Which AI Models Does Each Major Provider Offer in 2026?

- Key Takeaways - Why a Current AI Model Catalog Requires More Than a Leaderboard - OpenAI, Anthropic, Google, and xAI Define the Hosted Frontier - Meta, DeepSeek, Alibaba, Moonshot AI, and MiniMax Expand Open and Hybrid Access - Mistral, Cohere, AI21, and IBM Target Controlled Enterprise Deployment - Amazon, Microsoft, and NVIDIA Connect Models to Cloud and Compute Platforms - ByteDance, Baidu, Z.ai, Tencent, and Perplexity Broaden the Global Provider Map - Image, Video, Speech, Retrieval, Safety, and Robotics Models Form Distinct Markets - How Major AI Providers and Models Should Be Compared - Summary - Appendix: Useful Books Available on Amazon - Appendix: Top Questions Answered in This Article - Appendix: Glossary of Key Terms Key Takeaways - Frontier AI now covers language, coding, images, video, speech, search, safety, and robotics. - Model families, API identifiers, aliases, previews, and dated snapshots are not interchangeable. - The best provider depends on capability, price, control, deployment, distribution, and governance. Why a Current AI Model Catalog Requires More Than a Leaderboard On July 18, 2026, OpenAI released GPT-5.6 for general availability in three tiers named Sol, Terra, and Luna. That release arrived in a month that also included Meta Muse Spark 1.1, Tencent Hy3, new media models, pricing changes, and several approaching model-retirement dates. A comprehensive list of major AI providers and models can become outdated within days. A new model may replace a flagship, a familiar API alias may begin pointing to a different underlying system, and a provider may retain an older marketing name inside a consumer application. Preview access may also be available through one product but absent from another. The phrase “AI model” now describes systems that perform substantially different work. A large language model generates and transforms text. A multimodal model accepts combinations of text, images, audio, video, or documents. A reasoning model uses added inference computation before producing an answer. Coding models are optimized for software engineering. Image and video models synthesize media. Embedding models convert content into numerical representations for search, recommendations, classification, and retrieval. Rerankers reorder search results. Safety models classify prompts and outputs. Robotics and world models represent physical environments, movement, and possible actions. This breadth explains why a single benchmark cannot identify the best model for every organization. A model that performs well on advanced mathematics may be too slow or expensive for customer support. A compact open-weight model may suit a factory, hospital, government department, or satellite operator that needs local control. A hosted frontier model may be a stronger choice for a research group that values reasoning depth more than deployment independence. Speech latency matters for a voice agent. Image consistency matters for a media studio. Document-layout recognition matters for an insurer or law firm. Tool reliability matters for an autonomous coding system. The broader AI value chain includes chips, data centers, cloud services, model laboratories, application software, integrators, security products, and distribution channels. Model terminology creates another source of confusion. A family is a related group of models, such as GPT-5.6, Claude 5, Gemini 3, Qwen3, or Granite 4.1. A tier is a capability, speed, or price level within that family. An alias is a convenient API name that can be redirected to a newer version without requiring a customer to change code. A snapshot fixes a model to a dated release so its behavior is less likely to change unexpectedly. A preview provides public access before the provider promises long-term production stability. An experimental model carries greater uncertainty. A legacy model remains accessible but has a recommended successor. A deprecated model is scheduled for removal or is no longer recommended for new applications. A retired model is no longer available through the affected service. Open-weight and open-source also require separation. An open-weight release makes trained model parameters downloadable, but its license, source-code availability, training-data disclosure, acceptable-use rules, and commercial rights can differ. A model may publish its weights without disclosing the full training process. It may also permit commercial deployment under conditions that do not match a conventional open-source software license. Product names are equally unreliable as model identifiers. Microsoft Copilot can use Microsoft models and partner models. Amazon Bedrock hosts Amazon Nova beside models from outside suppliers. Perplexity can combine its Sonar models with external systems through its Agent API. Google deploys Gemini models through consumer applications, developer services, cloud platforms, workspace products, and research tools. An end user may see a single product label that hides the exact model, routing rule, reasoning level, or dated snapshot used for a request. Dynamic routing allows a product to send simple tasks to an economical model and difficult tasks to a more capable one. Search, moderation, image generation, and speech may be handled by separate models behind the same interface. Provider competition is also shaped by distribution. OpenAI, Anthropic, Google, and xAI emphasize hosted frontier access. Meta combines proprietary Muse models with downloadable Llama releases. Amazon, Microsoft, NVIDIA, and IBM connect models to cloud, enterprise, consulting, and hardware channels. Chinese developers such as Alibaba, DeepSeek, Moonshot AI, ByteDance, Baidu, Tencent, Z.ai, and MiniMax distribute models through domestic products, global developer portals, open repositories, and cloud platforms. The distribution of AI market share reflects those channels as much as benchmark performance. A provider can gain influence through an operating system, cloud platform, office suite, social network, developer environment, or search product even when another company leads a specific technical evaluation. A useful catalog must answer several questions for every entry. It must identify the recommended family, distinguish general models from specialist models, describe hosted access and downloadable availability, state preview or retirement status, and avoid treating a future announcement as a released product. Those rules also explain why older names remain relevant after a new flagship arrives. Deployed applications, contractual commitments, fine-tuned systems, regional availability, migration schedules, and reproducibility requirements can keep an earlier generation in active use for months or years. Enterprise spending is moving beyond experimentation toward model access, integration, security, data preparation, monitoring, and workflow redesign. The 2026 enterprise AI spending analysis describes a market in which the model is one cost inside a much larger implementation program. Greater spending does not remove the need for careful model identification. It makes version control, procurement language, audit records, regression tests, and exit planning more important. A provider can alter aliases, prices, context limits, safety behavior, or retirement dates after an application enters production. OpenAI, Anthropic, Google, and xAI Define the Hosted Frontier The largest proprietary providers offer portfolios rather than one general chatbot. Their catalogs separate high-compute reasoning from fast responses, coding from general work, and media generation from language processing. They also use routing systems that may select models according to the request, subscription, tool, or required response time. The operational question is not simply which company appears strongest. It is which exact model, version, context window, data policy, region, tool set, and service commitment applies to a particular workload. OpenAI The OpenAI model catalog places GPT-5.6 at the top of its July 2026 lineup. GPT-5.6 Sol is the highest-capability tier. GPT-5.6 Terra balances capability and operating cost. GPT-5.6 Luna is optimized for cost-sensitive workloads. All three support reasoning settings and multimodal input. OpenAI lists a context window of approximately 1.05 million tokens and a maximum output of 128,000 tokens for the family. Their model identifiers are gpt-5.6-sol , gpt-5.6-terra , and gpt-5.6-luna . The tiered structure lets developers remain within one generation when selecting a balance of capability, latency, and price. An organization can route demanding research or software tasks to Sol, routine professional work to Terra, and high-volume processing to Luna. Earlier GPT generations remain relevant to existing deployments. OpenAI’s current model guidance includes GPT-5.5, GPT-5.4, GPT-5.3 Codex, GPT-5.2, GPT-5.1, GPT-5, and GPT-4.1. Some earlier models remain available, and others have entered deprecation. GPT-5.6 is the recommended family for most new API work. Coding has its own model path. GPT-5.3 Codex is the current named Codex generation in OpenAI’s model guidance. Earlier GPT-5 Codex models may still appear in existing workflows. Codex models are designed for agentic software engineering, including repository navigation, file editing, command execution, debugging, and test-driven changes. The Codex label does not mean that other GPT models cannot write code. It identifies models and workflows optimized for longer software-development tasks. A general frontier model may explain an algorithm well, but a coding model can be better prepared to inspect a repository, modify several files, run tests, and revise its work after receiving tool output. OpenAI’s earlier o-series remains operationally relevant. Models such as o3 and o3-pro were designed for extended reasoning. Later GPT generations have absorbed much of that role through selectable reasoning levels. A procurement document should not assume that an “o” prefix automatically indicates greater reasoning ability than a newer GPT model. The company’s specialist catalog is extensive. GPT Image 2 is the recommended image-generation and editing model. GPT Image 1.5 and GPT Image 1 represent earlier generations, with older DALL-E models placed on deprecation paths. Realtime services include GPT-Realtime-2.1, GPT-Realtime-2.1 mini, GPT-Realtime-2, GPT-Realtime-Translate, GPT-Realtime-Whisper, GPT-Realtime-1.5, GPT-Realtime, and GPT-Realtime mini. These models support combinations of live speech, transcription, translation, tool use, text, and audio output. Audio endpoints include gpt-audio-1.5 and gpt-audio. Transcription options include GPT-4o Transcribe, GPT-4o mini Transcribe, and GPT-4o Transcribe Diarize. TTS-1 and TTS-1 HD generate speech, and Whisper remains available for general speech recognition. OpenAI also offers open-weight models named gpt-oss-120b and gpt-oss-20b. These provide an alternative to fully hosted GPT services, though their capabilities, deployment requirements, licensing, and safeguards differ from GPT-5.6. Search and retrieval systems can use text-embedding-3-large, text-embedding-3-small, or the older text-embedding-ada-002. Omni-moderation supports content classification. The combined catalog covers frontier text, code, images, live audio, transcription, speech, embeddings, moderation, and local deployment. OpenAI’s strength lies in portfolio integration. The same customer can combine a frontier model, coding agent, image generator, realtime voice interface, embedding model, and moderation layer through related services. Its catalog risk is naming density. Buyers should record the exact model identifier and retirement policy rather than rely on the broad labels “ChatGPT” or “GPT.” Anthropic Anthropic’s current Claude catalog contains four main public tiers. Claude Fable 5 is the company’s highest-capability widely released model. Claude Opus 4.8 is recommended for complex agentic coding and enterprise work. Claude Sonnet 5 balances speed and intelligence. Claude Haiku 4.5 is the fastest current Claude tier. Their public API identifiers are claude-fable-5 , claude-opus-4-8 , claude-sonnet-5 , and claude-haiku-4-5-20251001 . The Haiku alias is claude-haiku-4-5 . Fable 5, Opus 4.8, and Sonnet 5 support context windows of up to 1 million tokens in Anthropic’s current documentation. Haiku 4.5 supports 200,000 tokens. All current public Claude models accept text and image input and produce text output. Anthropic’s hierarchy is not a simple version ladder. Fable represents the highest widely available capability. Opus remains a high-end model line with a substantial enterprise and coding role. Sonnet targets production systems that need strong performance at lower cost. Haiku emphasizes speed and economical processing. Claude Mythos 5 sits outside general public access. It shares many specifications with Fable 5 but is offered in limited availability for approved defensive cybersecurity customers through Project Glasswing. Claude Mythos Preview remains another restricted model. Neither should be presented as a self-service option available to every developer. Anthropic publishes model-deprecation schedules. Claude Opus 4.7 was scheduled for removal on July 24, 2026. The short interval between the July 20 accuracy date and the planned removal demonstrates why production systems need pinned identifiers, regression tests, migration plans, and fallback models. Claude’s public portfolio is narrower than Google’s media catalog or OpenAI’s collection of dedicated audio endpoints. Anthropic concentrates on language, image understanding, reasoning, coding, tool use, long-context processing, and agentic execution. Its enterprise position also connects to questions about model access and national control. The debate over Anthropic restrictions and sovereign AI shows why governments and regulated organizations examine their dependence on outside suppliers. The technical catalog cannot resolve every sovereignty concern. It does determine whether customers can self-host, inspect weights, select a region, control upgrades, or continue operating after a policy or availability change. Google DeepMind The Gemini API model catalog is one of the broadest among major providers. It covers general reasoning, high-volume processing, live conversation, speech translation, image generation, video, music, embeddings, research agents, computer control, and robotics. Gemini 3.5 Flash is Google’s current high-capability Flash model. Gemini 3.1 Flash-Lite serves high-volume workloads that need lower cost. Gemini 3.1 Pro remains a preview model for advanced reasoning, complex problem solving, coding, and agentic work. Gemini 3 Flash also remains in preview. Earlier Gemini 2.5 models continue to matter. Gemini 2.5 Pro supports demanding reasoning and coding. Gemini 2.5 Flash offers a balance of performance and speed. Gemini 2.5 Flash-Lite is the least expensive member of that generation. Google also provides specialized live and speech models. Gemini 3.5 Live Translate supports low-latency speech-to-speech translation. Gemini 3.1 Flash Live is designed for realtime dialogue and voice applications. Gemini 3.1 Flash TTS generates speech. Gemini 2.5 Flash Live, Gemini 2.5 Flash TTS, and Gemini 2.5 Pro TTS remain available for related workloads. Gemini Omni Flash adds conversational video generation and editing. It accepts text and images and allows generated video to be revised through natural-language instructions. This service demonstrates how the boundary between a multimodal language model and a dedicated media generator is becoming less distinct. Google uses the Nano Banana brand for image generation. Nano Banana 2 is the production-oriented high-efficiency model. Nano Banana 2 Lite targets lower latency and cost. Nano Banana Pro provides a higher-capability tier. The original Nano Banana remains associated with Gemini 2.5 Flash image generation and editing. The Veo model family handles video generation. Veo 3.1 and its associated product tiers support text-to-video and image-to-video workflows. The Lyria family handles music generation and interactive music creation. Google’s catalog also includes Deep Research services, computer-use capabilities, embedding models, and Gemini Robotics-ER. Robotics-ER extends Gemini’s visual and language reasoning toward embodied systems that interact with physical environments. Status labels matter throughout the Google portfolio. Preview models can change faster than stable releases. Experimental systems carry less assurance. Earlier models may receive shutdown dates even when they remain embedded in public discussions or old code examples. Google’s advantage is breadth. A customer can assemble language, speech, image, video, music, search, embeddings, computer control, and robotics services through related platforms. The tradeoff is catalog complexity. Developers must distinguish consumer Gemini product names from API models and cloud deployment identifiers. xAI xAI’s developer documentation recommends Grok 4.5 for general tasks, coding, agentic work, and knowledge-intensive applications. The model supports a context window of up to 500,000 tokens and configurable reasoning levels. Grok 4.5 became available through the xAI API on July 8, 2026. Its documented training-data cutoff is February 1, 2026. Access to realtime information depends on search tools rather than the model’s stored training knowledge. The company offers separate APIs for images, video, and voice. Grok Imagine supports image generation and editing. The same media program provides video generation. Grok Voice supports conversational audio. Speech-to-text and text-to-speech services are also available. xAI often presents specialist capabilities as service families rather than publishing a long list of separately branded public model identifiers. This simplifies the visible catalog but can provide less model-level detail than platforms that expose every audio, image, and video version. Earlier Grok generations may remain present in consumer products, third-party integrations, or dated API snapshots. For new general API workloads, Grok 4.5 is the central current identifier. xAI’s position is tied to large-scale computing infrastructure, integration with X, search access, coding tools, and a growing set of media services. Its catalog is narrower than Google’s and less open than Meta’s Llama program, but Grok 4.5 belongs in the leading proprietary group because it combines reasoning, code, tool use, long context, and multimodal input. These four providers use different portfolio strategies. OpenAI divides its offering among GPT, Codex, image, audio, embedding, and moderation lines. Anthropic concentrates on a compact Claude hierarchy. Google publishes many specialist families. xAI leads with a general flagship supported by voice and media services. Those differences can matter more than a small benchmark gap. They affect integration effort, vendor concentration, product coverage, migration cost, and the number of outside components required to build a complete application. Meta, DeepSeek, Alibaba, Moonshot AI, and MiniMax Expand Open and Hybrid Access Competition outside the largest hosted Western providers influences model pricing, open deployment, multilingual performance, agent design, coding, and media generation. These providers also blur geographic and commercial categories. Chinese developers serve international users, American companies publish open weights used worldwide, and cloud platforms distribute models across regional boundaries. Meta Meta operates two distinct model programs. Muse is a proprietary frontier family delivered through hosted services. Llama remains the company’s principal open-weight family. Muse Spark 1.1 is Meta’s current multimodal reasoning and agentic model. Released on July 9, 2026, it improves tool use, computer use, coding, and multimodal understanding over the original Muse Spark. Muse Spark 1.1 powers Meta AI and is also available to external developers through the Meta Model API. The developer service supports tool calling and agentic workflows. The model can process text, images, video, documents, and audio for understanding tasks. Meta also introduced Muse Image and Muse Video. Muse Image supports image generation, reference composition, editing, and integration with Muse Spark. Muse Video applies the related media program to video generation with native audio. Availability and product integration can differ across Meta services. The proprietary Muse line represents a change from the assumption that every advanced Meta model would follow the Llama distribution pattern. Meta can now compete in hosted frontier services and maintain an open-weight program at the same time. The Llama 4 family includes Llama 4 Maverick and Llama 4 Scout. Both are open-weight, natively multimodal mixture-of-experts models. Maverick targets higher capability, and Scout emphasizes efficiency and long-context use. Llama 4 Behemoth was introduced as a much larger teacher model under development. It was not released as a generally downloadable model alongside Maverick and Scout. Lists that include Behemoth should state that distinction. Earlier Llama 3.3, Llama 3.2, and Llama 3.1 models remain widely used. They have been incorporated into local software, cloud services, research projects, mobile tools, and enterprise systems. An older open-weight model can remain commercially important long after the provider announces a newer generation. Meta also distributes safety components. Llama Guard 4 classifies content risks. Prompt Guard 2 focuses on prompt attacks and malicious instructions. LlamaFirewall provides a framework for protecting model-based applications. These components show that an open model program requires more than a downloadable language checkpoint. Deployment teams also need input screening, output policies, tool restrictions, logging, identity controls, and application-level defenses. Meta’s hybrid strategy gives customers a choice between hosted frontier access and downloadable models. It also creates a branding challenge. Muse Spark 1.1 and Llama 4 are separate programs with different access methods, product goals, and deployment assumptions. DeepSeek DeepSeek V4 includes DeepSeek-V4-Pro and DeepSeek-V4-Flash. V4-Pro is the higher-capability model, with 1.6 trillion total parameters and approximately 49 billion active parameters. V4-Flash uses 284 billion total parameters and approximately 13 billion active parameters. Both support context windows of up to 1 million tokens and thinking or non-thinking operation. DeepSeek-V4-Pro targets demanding reasoning, coding, and agentic tasks. V4-Flash targets faster and less expensive inference. The direct API identifiers are deepseek-v4-pro and deepseek-v4-flash . The older aliases deepseek-chat and deepseek-reasoner map to non-thinking and thinking configurations of V4-Flash. DeepSeek scheduled those aliases for deprecation on July 24, 2026. This is a clear example of alias risk. A familiar identifier can begin pointing to a newer underlying model and then be removed. Developers should record the direct model identifier when consistent behavior and lifecycle planning matter. DeepSeek’s earlier open releases remain influential. DeepSeek-V3.2, DeepSeek-R1-0528, DeepSeek-R1, DeepSeek-V3-0324, and DeepSeek-V3 appear in local deployments and third-party services. The R1 distilled family includes smaller models based on Qwen and Llama architectures. These distilled editions range from 1.5 billion to 70 billion parameters. They attempt to transfer some reasoning behavior into packages that are easier to host on modest infrastructure. DeepSeek’s pricing and open releases continue to pressure rivals. The economic effect began with its earlier systems, which changed assumptions about model costs. The provider’s importance does not rest on price alone. Hosted and downloadable releases give developers more control over reasoning modes, local deployment, and outside inference services. Organizations still need to examine license terms, security review, regional data handling, export controls, model provenance, and differences between official endpoints and third-party hosts. Alibaba and Qwen Alibaba Cloud Model Studio identifies Qwen3.7-Max as its leading hosted model as of July 20, 2026. Qwen3.7-Plus provides a balanced alternative, and Flash tiers serve high-volume or cost-sensitive workloads. The hosted catalog also contains Qwen3.6 models, Qwen3.5 Omni services, Qwen3-Coder-Plus, and Qwen3-Coder-Next. Omni models process combinations of text, images, audio, and video. Coder models target software engineering and agentic development. The exact set of available models can differ between Alibaba Cloud’s international and mainland China regions. A model may appear in Qwen Chat before reaching every cloud endpoint, or remain available through one region after another service has begun migrating customers. Alibaba’s open Qwen3 family contains mixture-of-experts and dense models. Large mixture-of-experts releases include Qwen3-235B-A22B and Qwen3-30B-A3B. The notation indicates total parameters and the approximate number activated during inference. Dense Qwen3 releases include 32B, 14B, 8B, 4B, 1.7B, and 0.6B editions. These sizes support deployments ranging from data centers to personal computers and edge devices. Qwen3-Coder-480B-A35B-Instruct is a major open coding release. The hosted coding line includes Qwen3-Coder-Plus and Qwen3-Coder-Next. The open and hosted offerings are related but should not be treated as identical services. Qwen3-Embedding models are available at 0.6B, 4B, and 8B sizes. Qwen3-Reranker models use the same scale options. Qwen-MT supports translation, and Qwen3Guard provides safety classification. The Qwen program also includes visual-language, speech, image, and video research. Product identifiers can differ among Model Studio, ModelScope, Qwen Chat, open repositories, and regional cloud services. Alibaba’s strength is breadth across managed and downloadable deployment. A developer can choose a hosted Max or Plus model, deploy an open Qwen checkpoint, add a coding model, build retrieval with embeddings and reranking, and apply a separate safety classifier. The resulting catalog is powerful but complex. Organizations should record the provider region, model identifier, access method, weight license, and retirement policy rather than use “Qwen3” as a complete technical description. Moonshot AI Kimi K3 is Moonshot AI’s current flagship for general reasoning, coding, long-horizon work, multimodal understanding, and coordinated agent tasks. It powers the current Kimi consumer experience and related professional tools. Kimi K2.7 Code remains associated with Moonshot’s coding service and Kimi Code tools. Kimi K2.6 is an open model focused on coding, agent execution, multimodal work, and coordinated subagents. Earlier Kimi K2 and Moonshot V1 services may still appear in older integrations or regional documentation. Moonshot’s release cadence makes status tracking important. Consumer products, coding subscriptions, open repositories, and platform APIs may refer to different model generations. A model presented in an application may not have the same availability as a downloadable release. Kimi’s competitive identity centers on long context, coding, agent execution, office productivity, and aggressive pricing. The provider has also emphasized coordinated agents that can divide a task into parallel workstreams. The value of that approach depends on more than the base model. Agent orchestration can increase token use, tool calls, execution time, and opportunities for error. Customers should evaluate the complete workflow, including planning, delegation, result checking, and failure recovery. MiniMax MiniMax-M3 is MiniMax’s current flagship language and agent model. Released on June 1, 2026, it supports multimodal input, coding, agentic reasoning, tool use, and a context window of up to 1 million tokens. MiniMax-M2.7 and MiniMax-M2.7-highspeed remain available. The high-speed edition provides the same model generation through a faster inference service. M2.5, M2.1, and M2 remain listed for existing workloads, though M3 is the current recommended flagship. The Hailuo family handles video generation. MiniMax Hailuo 2.3 supports text-to-video and image-to-video generation. Hailuo 2.3Fast emphasizes faster and more economical image-to-video processing. Hailuo 02 remains a legacy video model. Speech-2.8-hd and speech-2.8-turbo are the current speech-generation models. They support 40 languages and several controllable emotions. Older speech-2.6 and speech-02 models remain listed as legacy options. MiniMax’s current music catalog includes Music-3.0, Music-2.6, and Music-Cover. Music-3.0 was released on July 16, 2026. Music-Cover generates cover versions using reference audio and supports style transfer and lyric-related workflows. The company also provides image-generation services, coding agents, and consumer products. Naming can differ between the MiniMax API, Hailuo, Talkie, and regional offerings. MiniMax belongs among major providers because it combines a frontier language model with developed video, speech, music, and image services. Its cloud distribution is smaller than Amazon’s or Microsoft’s, but its media portfolio and release pace make it relevant to creative applications and agent systems. This group demonstrates the decline of a simple open-versus-closed division. Meta operates proprietary Muse beside open Llama. Alibaba combines managed Qwen services with downloadable Qwen models. DeepSeek pairs low-cost hosted access with open releases. Moonshot combines hosted Kimi services and open generations. MiniMax emphasizes managed language and media services. Buyers must examine the exact model and license rather than infer deployment rights from a provider’s public reputation. Mistral, Cohere, AI21, and IBM Target Controlled Enterprise Deployment Enterprise selection often centers on control, predictable operating cost, multilingual support, retrieval quality, private deployment, document processing, and integration with existing systems. Mistral AI, Cohere, AI21 Labs, and IBM compete strongly on these requirements. Their catalogs include general models, but they devote substantial attention to embeddings, reranking, moderation, speech, formal reasoning, optical character recognition, and smaller systems that can operate inside customer-controlled infrastructure. Mistral AI Mistral AI’s current catalog includes Mistral Large 3, Mistral Medium 3.5, Mistral Small 4, and the Ministral 3 family. Mistral Large 3 is an open-weight general-purpose multimodal model. Mistral Medium 3.5 is a frontier-class multimodal model optimized for agents and coding. Mistral Small 4 combines instruction following, reasoning, and coding in a smaller package. Ministral 3 includes 3B, 8B, and 14B editions. These open models target local and edge deployment. The 8B and 14B editions support text and vision, and the family provides context windows of up to 256,000 tokens. Mistral’s specialist catalog is extensive. Devstral 2 is designed for software-engineering agents. Codestral serves code completion and programming workflows. Codestral Embed creates vector representations of source code for search and retrieval. Mistral OCR 4 is the current optical character recognition and document-understanding service. It adds structural block labels and paragraph-level bounding boxes, allowing applications to retain layout information rather than extract plain text alone. Leanstral 1.5 focuses on formal proof engineering in Lean 4. It is intended for mathematical formalization, proof generation, and automated theorem-proving workflows. This is a narrow use case that a general chat model may not handle with the same consistency. The Voxtral family supports speech and audio. Voxtral TTS generates speech with multilingual support and zero-shot voice cloning. Voxtral Mini Transcribe 2 handles transcription. Voxtral Mini Transcribe Realtime supports streaming audio. Other Voxtral models remain available for audio-language understanding. Mistral also offers moderation and embedding services. Earlier Magistral, Small, Medium, OCR, Voxtral, and Devstral generations have retirement schedules, with current documentation identifying their recommended replacements. Mistral’s position combines European corporate identity, open-model distribution, and managed services. That mix appeals to customers seeking alternatives to the largest American platforms and organizations concerned with regional control or procurement independence. It also supports hybrid deployments in which a larger hosted model performs demanding tasks and a smaller local model handles routine or sensitive work. Cohere Cohere’s model catalog focuses on enterprise language systems, retrieval-augmented generation, multilingual applications, translation, and private deployment. Command A+ is the newest member of the Command A family. It is Cohere’s first mixture-of-experts Command model, with 218 billion total parameters and approximately 25 billion active parameters. It accepts text and images and combines reasoning, agents, tool use, vision, and translation. Command A remains an important production model. It has 111 billion parameters and a context window of 256,000 tokens. It is optimized for tool use, retrieval-augmented generation, agents, and multilingual enterprise workloads. Command A Reasoning uses added reasoning computation for complex tasks. Command A Translate supports translation across 23 languages. Command R and Command R+ remain recognized retrieval-oriented models, though Cohere recommends newer Command A models for many applications. The Aya family addresses multilingual access. Aya Expanse supports multilingual text generation, and Aya Vision adds visual understanding. These models are relevant in markets where English-centered systems provide weaker performance or less appropriate responses. Cohere’s retrieval catalog is one of its strongest differentiators. Embed v4.0 accepts text, images, and mixed text-image input. Embed English v3.0 and Embed Multilingual v3.0 remain available, along with lighter editions. Rerank v4.0 Pro is optimized for complex retrieval tasks and higher quality. Rerank v4.0 Fast emphasizes latency and throughput. Rerank v3.5 and older v3 models remain present in existing systems. Cohere Transcribe extends the portfolio into multilingual automatic speech recognition. The company’s enterprise focus is visible across the catalog through private deployment, search quality, multilingual support, connectors, and control over organizational data. Retrieval components deserve separate evaluation from the language model. A general model can write an eloquent answer from weak material. Embedding and reranking models decide which internal documents reach that model. In many enterprise applications, retrieval quality directly affects accuracy, auditability, and user confidence. AI21 Labs AI21 Labs’ Jamba family combines transformer components with state-space modeling techniques intended to improve long-context efficiency. Jamba2 Mini uses approximately 12 billion active parameters and targets enterprise workflows that need a balance of capability, steerability, and deployment efficiency. Jamba2 3B is a compact model designed for on-device applications, private deployment, and agentic systems. Earlier Jamba Large and Jamba Mini generations may remain in customer systems or prior documentation, but Jamba2 Mini and Jamba2 3B are the primary current models presented in AI21’s foundation-model documentation. The architecture is designed to process longer contexts efficiently compared with a conventional transformer of similar scale. Jamba2 Mini serves broader enterprise tasks, and Jamba2 3B provides greater control for local or constrained use. AI21’s catalog is smaller than those of Google or Alibaba. That can simplify procurement. The company concentrates on a related architectural family rather than maintaining separate consumer, image, speech, and video programs. The tradeoff is that customers may need outside components for a complete multimodal platform. AI21 can still fit organizations seeking long context, private deployment, model steerability, and an alternative to the dominant transformer-only design. IBM IBM Granite 4.1 is a family of dense open models available in 3B, 8B, and 30B sizes. Each size is distributed in base and instruction-tuned forms, with optional FP8 editions for more efficient deployment. Granite 4.1 3B is intended for edge systems and resource-constrained environments. Granite 4.1 8B provides a balanced enterprise model. Granite 4.1 30B supports more complex reasoning and specialist tasks. IBM releases the family under the Apache 2.0 license. The current documentation emphasizes tool calling, instruction following, coding, mathematical reasoning, and model-transparency disclosures. Granite 4.0 remains relevant through dense and hybrid models. The hybrid editions combine Mamba-2 and transformer components, with mixture-of-experts designs used in selected sizes. Smaller Granite models support local inference, adaptation, and task-specific deployment. Granite Vision 4.1 handles chart, table, document, and key-value extraction. Granite Docling is a compact document-understanding model that converts PDFs, slides, scans, and page layouts into structured machine-readable output. Granite Speech 4.1 includes compact models for multilingual speech recognition and translation. Granite Guardian 4.1 evaluates prompts, outputs, jailbreak attempts, policy conditions, and possible hallucinations in tool and retrieval workflows. Granite Embedding includes multilingual embedding models and a reranker for semantic search and retrieval-augmented generation. IBM’s model program also extends beyond generative language through Flowstate, Tiny Time Mixer, and Time Series Pulse for forecasting, classification, and anomaly detection. IBM’s differentiator is fit with governed enterprise environments. Granite models are designed for inspection, adaptation, and customer-controlled deployment. They can integrate with IBM software and consulting services, but the open releases are not limited to IBM infrastructure. Compared with proprietary frontier systems, Granite may trade some broad benchmark performance for smaller computing requirements, transparent licensing, and deployment control. The four providers in this section share an enterprise orientation but approach it differently. Mistral offers a broad open and hosted portfolio. Cohere centers retrieval, multilingual work, and private deployment. AI21 concentrates on the Jamba architecture. IBM distributes compact models across language, vision, documents, speech, retrieval, safety, and time series. Their value becomes clearest when selection begins with the workload rather than the most familiar consumer brand. Amazon, Microsoft, and NVIDIA Connect Models to Cloud and Compute Platforms Amazon, Microsoft, and NVIDIA sell models and the infrastructure used to operate models. Amazon Web Services and Microsoft Azure host partner systems beside their own families. NVIDIA supplies accelerators, networking, inference software, model containers, and reference models. Their influence can shape procurement even when a customer selects a model developed by another organization. Amazon Amazon Nova 2 includes Nova 2 Lite, Nova 2 Sonic, and Nova Multimodal Embeddings. Nova 2 Lite accepts text, images, video, and documents and produces text. It supports a context window of up to 1 million tokens and a maximum output of 64,000 tokens. Amazon positions it as a cost-efficient model for automation, document processing, customer support, reasoning, and multimodal analysis. Nova 2 Sonic supports low-latency speech interaction. It is intended for voice agents, conversational applications, and speech-to-speech workflows. Nova Multimodal Embeddings converts text, documents, images, video, and audio into a shared semantic space. This allows an application to search one form of media using another, such as finding a video from a text description or retrieving an image related to an audio segment. The earlier Nova generation remains available through Amazon Bedrock. Nova Premier is its highest-capability understanding model. Nova Pro serves general multimodal work. Nova Lite targets lower-cost multimodal processing. Nova Micro is a compact text-only model. Nova Sonic handles conversational audio. Nova Canvas generates images, and Nova Reel generates video. Existing customers may continue using those models even as Nova 2 receives current development attention. Amazon’s strategy is tied closely to Bedrock. Customers can combine Amazon Nova with models from Anthropic, Meta, Mistral, Cohere, and other suppliers through one cloud service. Bedrock adds knowledge bases, agents, guardrails, evaluation, monitoring, permissions, and enterprise controls. A customer may benefit from a single cloud environment, but model behavior, license terms, and pricing still differ by provider. Amazon’s advantage is distribution into existing cloud accounts. A customer already using AWS identity, storage, networking, databases, and security products can adopt model services without establishing a separate infrastructure platform. The corresponding risk is platform dependence. Applications built deeply around Bedrock agents, proprietary knowledge-base features, and AWS-specific controls may require substantial effort to move elsewhere. Microsoft Microsoft’s MAI model family now covers reasoning, coding, images, transcription, and speech. MAI-Thinking-1 is Microsoft AI’s reasoning model. It is designed for mathematics, software engineering, analysis, and general problem solving. MAI-Code-1-Flash is a lightweight agentic coding model integrated with GitHub Copilot and Visual Studio Code. MAI-Image-2.5 supports text-to-image generation and image editing. MAI-Image-2.5-Flash offers a faster and less expensive tier. MAI-Transcribe-1.5 converts speech into text, including noisy and specialist audio. MAI-Voice-2 generates low-latency expressive speech. Earlier models include MAI-Image-2, MAI-Image-2-Efficient, MAI-Image-1, MAI-Transcribe-1, MAI-Voice-1, and MAI-1-preview. These releases document the development path toward the current family but are not equal recommendations for new systems. Microsoft also develops the Phi family. Phi-4, Phi-4-reasoning, Phi-4-mini-instruct, Phi-4-mini-reasoning, and Phi-4-multimodal-instruct support compact deployment, local applications, research, and narrower tasks. Phi models address a different need from MAI-Thinking-1. They trade frontier scale for lower hardware demand and easier deployment. A smaller model can be the better choice for classification, extraction, local assistance, or applications with strict cost limits. Microsoft’s product distribution extends beyond its own models. Azure provides access to OpenAI and outside open models. Microsoft 365 Copilot, GitHub Copilot, security products, and business applications can use different models or routing policies. The product name “Copilot” does not identify one fixed model. A feature may use MAI, OpenAI, Phi, or a combination selected by Microsoft. Enterprise buyers should ask which model family, data terms, and service boundary apply to the specific feature they plan to adopt. Microsoft’s in-house expansion reduces dependence on one model partner and creates systems optimized for its own software. Customers benefit from integration with identity, permissions, office applications, developer tools, and cloud infrastructure. They should still evaluate portability if a workflow becomes tightly coupled to Microsoft-specific services. NVIDIA NVIDIA Nemotron 3 is an open-model family designed for agents, reasoning, coding, multimodal work, retrieval, speech, safety, and document processing. Nemotron 3 Ultra has 550 billion total parameters and approximately 55 billion active parameters. NVIDIA positions it for demanding multi-agent workflows, long-running research, planning, code generation, synthesis, verification, and recovery. Nemotron 3 Super has 120 billion total parameters and approximately 12 billion active parameters. It targets software development, cybersecurity triage, and multi-agent work that needs high capability with lower computing demand than Ultra. Nemotron 3 Nano has 30 billion total parameters and approximately 3 billion active parameters. It is designed for efficient coding, reasoning, mathematics, and long-context processing. Nemotron 3 Nano Omni uses the same 30B-A3B scale and adds text, image, video, and audio understanding. It can act as a multimodal subagent within a larger system. The family also includes Nemotron Retriever for extraction, embeddings, reranking, and document retrieval. Nemotron Parse reconstructs text, tables, reading order, and layout from complex documents. Nemotron Speech supports voice and audio applications. NVIDIA distributes models through downloadable weights, NVIDIA Inference Microservices, build services, and enterprise software. A customer can use optimized containers, NVIDIA hardware, or managed model endpoints. The Cosmos program extends NVIDIA’s work into physical AI. Cosmos models support robotics, autonomous systems, synthetic data, world representation, and simulation. These tasks require models to reason about space, time, movement, physical constraints, and possible actions. NVIDIA benefits whenever demand rises for model training and inference, regardless of which provider leads a language benchmark. Nemotron gives the company its own model assets, but its larger influence comes from graphics processors, networking, software libraries, inference optimization, and reference architectures. The infrastructure connection creates convenience and concentration risk. A model optimized for one accelerator or cloud can offer stronger performance there. It may also become harder to move. Buyers should measure portability at the model, runtime, hardware, orchestration, data, and monitoring layers. Downloadable weights do not automatically guarantee simple migration. ByteDance, Baidu, Z.ai, Tencent, and Perplexity Broaden the Global Provider Map A provider list limited to American companies misses a substantial share of model development. ByteDance, Baidu, Z.ai, and Tencent operate major research and commercial programs in China, often combining general language models with image, video, speech, robotics, and enterprise services. Perplexity occupies a different category as a search-centered model provider and orchestration platform. Together, these companies show how competition is moving toward agents, multimedia, search grounding, and national technology strategies. ByteDance Seed ByteDance Seed2.1 was released on June 23, 2026. It is an agent-capable model family designed for professional work, coding, tool use, multimodal understanding, and longer task execution. Seed2.1 Pro is the higher-capability edition described in ByteDance’s evaluations. The family is available through Doubao and Volcano Engine services, with access and identifiers differing by region and product. Earlier Seed2.0 editions include Pro, Lite, Mini, and Code. They may remain present in existing workflows even after Seed2.1 becomes the recommended generation. Seedream 5.0 Pro is ByteDance’s higher-tier image-generation model. It emphasizes image-text alignment, structural coherence, typography, design work, and complex visual layouts. Seedream 5.0 Lite provides a less expensive image tier with reasoning and search-assisted generation. Seedance 2.0 handles video creation. It accepts combinations of text, images, audio, and video and can generate synchronized audiovisual output. It also supports video extension, editing, references, and camera-direction controls. Seeduplex is a full-duplex speech-language model. Full-duplex interaction allows the system to listen and respond with less rigid turn-taking than a conventional voice assistant. Seed Audio 1.0 supports audio generation. Seed3D addresses three-dimensional content, and Seed GR-RL applies reinforcement learning to robotics. Distribution occurs through ByteDance research releases, Volcano Engine, Doubao, Feishu, Coze, and related products. The model name visible to a developer may differ from the consumer-facing brand. Baidu Baidu ERNIE 5.1 is the company’s current general flagship. Released on May 9, 2026, it was derived from the ERNIE 5.0 training foundation through an elastic model architecture. Baidu states that ERNIE 5.1 uses approximately one-third of ERNIE 5.0’s total parameters and about half its active parameters. The design seeks to retain high capability with lower training and inference cost. ERNIE 5.0 remains an important model. It uses a unified multimodal architecture trained across text, images, video, and audio rather than attaching separate media encoders only after text-model training. ERNIE-Image is an open 8B text-to-image model based on a diffusion-transformer design. PaddleOCR-VL-1.5 is a compact vision-language model for document parsing, including complex layouts, tables, formulas, and scanned pages. Earlier ERNIE 4.5 and 4.x models remain present in open repositories, cloud services, and customer systems. Baidu connects the ERNIE family to search, cloud services, maps, consumer applications, enterprise software, and autonomous-system research. The unified multimodal approach can support tasks that combine spoken instructions, video, written documents, and generated visual content. It also requires separate evaluation for each modality because strong text performance does not establish reliable video, audio, or image behavior. Z.ai GLM-5.2 is Z.ai’s current flagship. Released on June 16, 2026, it is designed for long-horizon tasks, software engineering, tool use, reasoning, and agents. GLM-5.2 supports a context window of up to 1 million tokens. Z.ai positions it for project-scale engineering in which one task can move from requirements through implementation and delivery. Earlier models include GLM-5.1, GLM-5, GLM-5-Turbo, GLM-4.7, GLM-4.6, GLM-4.5, and lower-cost Flash or Air editions. Visual models include GLM-5V-Turbo, GLM-4.6V, and GLM-4.5V. Z.ai also provides GLM-Image and CogView for image generation, video services based on CogVideo technology, GLM-OCR for document extraction, embedding models, and reranking models. The rapid progression from GLM-5 to GLM-5.1 and GLM-5.2 illustrates why model lists need an exact date. A catalog accurate in April 2026 could be outdated by June. Z.ai offers models through its international developer platform, Chinese services, coding subscriptions, and downloadable releases. Availability, pricing, and aliases can differ among those channels. Tencent Tencent Hy3 was released on July 6, 2026, following an earlier preview. Hy3 is a mixture-of-experts model with 295 billion total parameters and approximately 21 billion active parameters. It supports a context window of up to 256,000 tokens and combines fast and slower reasoning modes. Tencent positions Hy3 for reasoning, coding, agents, office productivity, financial modeling, front-end design, and game production. It is integrated into Tencent services and is available through Tencent Cloud. Hy-MT2-30B-A3B is an open translation model supporting 33 languages and five regional forms. HunyuanImage supports image generation. Hunyuan3D generates three-dimensional assets. HY-World addresses world generation and simulation, and HY-Embodied supports embodied-intelligence research. Tencent uses both Hy and Hunyuan branding. Newer general and translation models use the shorter Hy name, and media, 3D, world, and earlier systems retain Hunyuan branding. The naming change can make outside model lists appear inconsistent even when they describe related research and product programs. Tencent’s distribution strength comes from cloud services, communications, gaming, media, enterprise software, and consumer applications. Its models can enter products through integrated services rather than direct developer selection. Perplexity Perplexity Sonar is designed around web-grounded answers rather than standalone offline generation. Sonar provides fast and economical search-based responses. Sonar Pro performs broader retrieval and synthesis. Sonar Reasoning Pro combines search with extended reasoning. Sonar Deep Research conducts larger multi-source research tasks and produces long analytical outputs. The Sonar API remains supported but is in maintenance mode. Perplexity recommends the Agent API for new integrations. The Agent API can use Perplexity tools and models from external providers through one orchestration service. Perplexity also offers a Search API for ranked web results and embedding services for retrieval. The company’s value comes from the combination of search, ranking, source retrieval, orchestration, and language generation. This makes Perplexity an edge case in provider comparisons. It develops Sonar models, but much of the service’s performance depends on the retrieval system around them. Comparing Sonar with an offline language model without accounting for search access would be misleading. The model and the information system should be evaluated together. These providers have strategic significance beyond product features. Governments and companies increasingly treat model access, compute capacity, cloud control, training data, and technical talent as national assets. The debate over sovereign AI concerns whether dependence on foreign model services can affect public administration, industrial policy, security, bargaining power, and economic resilience. A global model catalog should describe capability without ignoring the legal and regional conditions that determine who can use it. Image, Video, Speech, Retrieval, Safety, and Robotics Models Form Distinct Markets General language models receive most public attention, but specialist model categories often determine whether a commercial system works. An image generator needs visual composition, editing controls, reference consistency, and readable text. A speech agent needs low latency, interruption handling, and natural prosody. A retrieval system needs embeddings that place related items close together. A safety classifier needs dependable policy labels. A robotics model needs spatial and physical reasoning. Treating all of these as minor features of a chatbot hides the technical and commercial structure of the market. Image generation has moved from one-shot text prompts toward repeated editing and agentic control. OpenAI offers GPT Image 2. Google offers Nano Banana 2, Nano Banana 2 Lite, and Nano Banana Pro. Meta offers Muse Image. Microsoft offers MAI-Image-2.5 and its Flash edition. Amazon offers Nova Canvas from its earlier Nova generation. ByteDance offers Seedream 5.0 Pro and Lite. Baidu offers ERNIE-Image. Tencent offers HunyuanImage. Z.ai offers GLM-Image and CogView. MiniMax retains an image-generation service. These systems differ in typography, identity consistency, prompt adherence, editing precision, output resolution, safety filters, and commercial terms. An impressive demonstration does not prove that a model will retain a person’s appearance across 20 images or produce accurate text in a dense infographic. Video generation is developing as a separate contest. Google’s Veo family, xAI’s Imagine service, Amazon Nova Reel, ByteDance Seedance 2.0, MiniMax Hailuo 2.3, Meta Muse Video, and Tencent’s media and world models use different approaches. Some generate short clips from text or images. Others support editing, camera control, reference composition, synchronized sound, character consistency, or continuation of an existing sequence. Release status needs careful wording. A model introduced in a research announcement may not be broadly available through a public API. A consumer application may provide limited access before an enterprise service launches. Regional availability may also differ. Music models form another category. Google’s Lyria family supports generated music and interactive creation. MiniMax offers Music-3.0 and related services. ByteDance’s Seed Audio program includes broader audio-generation work. Copyright, performer identity, training-data governance, and commercial licensing can matter as much as sound quality. A technically strong sample does not establish that a model suits a publisher, broadcaster, game studio, or advertising campaign. Speech covers at least four tasks: recognition, generation, speech-to-speech interaction, and audio-language understanding. OpenAI’s Realtime and audio families cover live interaction, transcription, diarization, translation, and generated speech. Google offers Live, Translate, and text-to-speech Gemini variants. Amazon Nova 2 Sonic supports conversational audio. Microsoft offers MAI-Transcribe-1.5 and MAI-Voice-2. Mistral’s Voxtral family provides transcription, realtime recognition, speech generation, and audio-language models. Cohere Transcribe adds multilingual recognition. ByteDance Seeduplex supports full-duplex conversation. MiniMax offers speech-2.8-hd and speech-2.8-turbo. Latency can dominate speech-model selection. A system that waits several seconds before responding may be acceptable for transcription but fail as a conversational agent. Full-duplex systems attempt to handle interruption, overlap, and short acknowledgement sounds closer to human conversation. Voice products also require consent controls, identity protection, recording policies, accent testing, noise handling, and protections against impersonation. Embedding and reranking models support retrieval-augmented generation, commonly abbreviated as RAG. An embedding model turns a query, paragraph, image, audio clip, or video segment into a vector. A search system retrieves nearby vectors, and a reranker reorders the candidates before a language model writes an answer. OpenAI, Google, Cohere, Alibaba, Amazon, IBM, Mistral, NVIDIA, Perplexity, and Z.ai provide embedding or reranking components. Cohere’s Rerank v4 models, Amazon’s multimodal embeddings, Qwen’s paired embedding and reranking families, and IBM’s compact multilingual models show how retrieval has become its own competitive category. Good retrieval can reduce unsupported answers by supplying relevant organizational material, but it does not guarantee truth. Source documents may be wrong, outdated, incomplete, or poorly divided into searchable units. Access controls can fail if an index ignores document permissions. A reranker can prefer persuasive but less authoritative text. Production systems need document governance, source dates, permission-aware search, evaluation sets, and visible provenance. Safety models operate at several layers. OpenAI offers omni-moderation. Meta provides Llama Guard 4, Prompt Guard 2, and LlamaFirewall. Alibaba offers Qwen3Guard. IBM offers Granite Guardian 4.1. NVIDIA provides Nemotron safety technology. Mistral offers moderation services. No single classifier can secure an application. Tool permissions, rate limits, identity controls, data-loss prevention, logging, approval rules, and human review remain necessary. Agent security has become more important because models can operate browsers, write code, access files, send messages, and call business systems. A harmless-looking webpage can contain instructions designed to redirect an agent. A compromised document can attempt to extract secrets. An over-permissioned coding agent can alter production systems. Safety evaluation must examine the complete tool-enabled application rather than the language model in isolation. Physical AI extends model competition into robots, vehicles, industrial systems, and simulation. Google’s Gemini Robotics-ER, NVIDIA Cosmos, ByteDance Seed GR-RL, Tencent HY-Embodied, and related world models process perception, action, or physical environments. They may combine camera feeds, language instructions, state estimation, simulation, and control policies. Their errors can affect machinery and people, placing greater weight on testing, operational boundaries, redundancy, and human override. Specialization also affects business models. General chat access is often sold by token or subscription. Image and video services may be priced by output, duration, resolution, or computing unit. Speech can be priced by audio minute. Embeddings may be priced by input volume, and rerankers by request. Robotics systems can combine software licenses with sensors, hardware, simulation, and integration. The question of AI profitability depends partly on whether providers can earn enough from these specialist services to cover model development, computing, energy, and distribution costs. General models may absorb some specialist functions over time, but dedicated models will remain valuable where latency, cost, accuracy, hardware limits, or governance dominate. A compact transcription model can outperform an expensive general agent on a narrow audio pipeline. A reranker can improve search at a small fraction of the cost of asking a frontier model to inspect every document. A moderation model can process large request volumes before a more expensive model receives the prompt. The commercial market is likely to support broad multimodal models and focused components at the same time. How Major AI Providers and Models Should Be Compared A credible comparison begins with a defined workload. “Best AI model” is too broad to guide procurement because a customer-support assistant, coding agent, scientific-research tool, media generator, document-processing pipeline, and warehouse robot have different success criteria. Evaluation should begin with representative tasks, approved data, expected request volume, response-time requirements, legal conditions, and the consequences of an error. Capability tests should use the organization’s own material where permitted. Public benchmarks can identify candidates, but they may be optimized, contaminated by training data, or weakly related to production work. A legal group needs accurate document analysis and reliable quotations. A software team needs patches that compile and pass tests. A multilingual service needs performance in the exact languages and regional forms it serves. A media team needs consistent subjects, readable text, editing control, and suitable commercial rights. Model status comes next. A stable production model carries different expectations from a preview, experimental release, restricted system, legacy endpoint, or announced product. Contracts and technical documentation should record the exact identifier, version policy, context limit, output limit, region, and retirement-notice period. Aliases that automatically move to newer models can simplify upgrades but introduce behavior changes without a code change. Fixed snapshots improve reproducibility but can miss later improvements or safety updates. Cost requires more than the advertised input-token price. Reasoning models can use billed reasoning tokens. Long prompts raise input costs and latency. Tool calls, web search, image generation, speech minutes, video duration, embeddings, storage, and reranking add separate charges. A cheaper model that produces more errors can raise review and correction costs. A larger model may reduce workflow steps enough to cost less overall. The relevant metric is the cost of completing an acceptable task rather than the cost of one token. Context windows require similar care. A 1 million-token limit allows very large inputs, but effective recall can vary across the prompt. Sending an entire document repository on every request is rarely efficient. Retrieval, structured databases, summaries, and workflow memory can provide better results at lower cost. Long context is useful for source code, legal records, research collections, and agent histories, but it does not replace information architecture. Openness should be measured through several questions. Are the weights downloadable? Does the license permit commercial use and modification? Can the model run on available hardware? Are training methods, data sources, evaluations, and safety practices documented? Can the customer fine-tune it? Does deployment require provider software or a proprietary runtime? An open-weight model can still create dependence on a cloud service, accelerator, inference library, or data pipeline. Privacy and governance must be assessed at the service level. A provider may offer different retention terms for a consumer chatbot, business subscription, cloud endpoint, or self-hosted model. Regional processing, encryption, logging, administrator controls, identity integration, and training-data use can differ. Regulated organizations should obtain contractual terms rather than rely on a general product description. Tool use changes the risk profile. A model that can search internal systems, execute code, purchase services, or send communications needs permission boundaries. Each tool should expose the minimum required capability. High-impact actions should require confirmation or approval. Logs should show the request, model, tool call, result, and user identity. Security testing should include prompt injection, secret extraction, malicious files, compromised websites, and unexpected tool sequences. Multimodal systems need separate evaluation for every input and output type. Strong text reasoning does not prove accurate image interpretation. Good image generation does not prove reliable optical character recognition. Natural speech does not prove accurate transcription in noisy environments. A provider’s general model name can hide different internal components for audio, vision, search, and media generation. Testing must follow the actual service path used in production. Portability deserves a planned budget. Prompts written for one provider may depend on its tool schema, system-message rules, safety policies, structured-output format, or agent runtime. Embeddings from different models generally cannot share one index without rebuilding it. Fine-tunes can be provider-specific. Evaluation suites, abstraction layers, standard data formats, and export procedures reduce switching costs. A multi-model strategy can limit dependence but adds engineering and governance work. Routing may send simple requests to a small model and difficult requests to a frontier system. Sensitive work may remain on-premises, and public research may use a hosted service. Coding can use a specialist model, with another model reviewing the result. The router itself must be evaluated because poor classification can send data or tasks to the wrong destination. Human oversight should match consequence rather than model size. Low-risk formatting can run automatically. Financial, medical, legal, safety, personnel, and public-policy outputs require stronger review. Agents that change systems or communicate externally need approval boundaries. A model’s confidence language should not determine the review level because fluent systems can sound certain when wrong. Organizations also need continuous evaluation. Model behavior can change after an alias update, safety adjustment, routing change, or provider-side optimization. A regression suite should test accuracy, refusal behavior, tool use, latency, cost, structured output, and security. Results should be stored by model identifier and date. Retirement notices should trigger migration tests before the deadline. Competitive claims require restraint. Providers publish selected benchmarks under different prompts, tool settings, reasoning budgets, and sampling methods. Independent tests can add perspective but may still omit cost, latency, regional access, and production reliability. The debate over whether AI will become a commodity turns on more than model quality. Distribution, proprietary data, workflow integration, brand, security, hardware efficiency, and switching costs can preserve differentiation even when raw capabilities converge. Natural-language interfaces will also influence adoption. People increasingly instruct agents in ordinary language and delegate multi-step work. That creates a contest over the conventions used for permissions, confirmation, memory, accountability, and control. A provider that makes delegation understandable, observable, and reversible may win users without leading every benchmark. A practical selection process can use six stages: - Define the task, data, users, volume, latency, and acceptable error rate. - Select a candidate set using status, modality, region, and deployment requirements. - Test representative cases and score quality, cost, latency, and safety. - Review contracts, privacy, licensing, retention, and retirement terms. - Pilot the complete workflow with tools, retrieval, monitoring, and human approval. - Maintain regression tests, fallback models, and an exit plan. No provider wins every stage. OpenAI and Anthropic may lead a demanding reasoning task. Google may provide the most complete multimodal and media portfolio. Meta, DeepSeek, Alibaba, Mistral, IBM, AI21, Cohere, NVIDIA, or Tencent may provide stronger deployment control. Amazon or Microsoft may fit an existing cloud environment. Perplexity may suit search-grounded work. ByteDance and MiniMax may be attractive for media production. The correct decision is workload-specific and can change after a model release, price adjustment, policy update, or retirement notice. Summary The major model families available as of July 20, 2026, include OpenAI GPT-5.6, Anthropic Claude Fable 5 and Claude 5 models, Google Gemini 3 models, xAI Grok 4.5, Meta Muse Spark 1.1 and Llama 4, DeepSeek V4, Alibaba Qwen3.7, Moonshot Kimi K3, MiniMax-M3, Mistral’s current Large, Medium, Small, and Ministral families, Cohere Command A+, Amazon Nova 2, Microsoft MAI and Phi, NVIDIA Nemotron 3, IBM Granite 4.1, AI21 Jamba2, ByteDance Seed2.1, Baidu ERNIE 5.1, Z.ai GLM-5.2, Tencent Hy3, and Perplexity Sonar. That list still hides the main development. Providers are dividing their portfolios among reasoning, speed, coding, images, video, speech, embeddings, reranking, safety, agents, and physical AI. They are also mixing hosted access, open weights, cloud distribution, local deployment, restricted models, previews, and automatically updated aliases. The catalog has become a map of access and operating models as much as a list of technical systems. Competition is likely to remain fluid. New releases can replace a flagship within weeks, and retirement notices can force migration soon after an update. The most dependable approach is to maintain a dated inventory, record exact identifiers, test representative workloads, and separate announced capability from available service. The deeper commercial issue is whether models become interchangeable components or remain differentiated platforms. Open-weight releases and falling inference prices push toward interchangeability. Proprietary data, agents, search, developer tools, media services, cloud integration, safety controls, and user distribution push in the opposite direction. Organizations will adopt complete workflows rather than model names alone. The provider that delivers an acceptable combination of capability, price, control, reliability, integration, and governance will often win the deployment even when another company leads a selected benchmark. Appendix: Useful Books Available on Amazon - The Alignment Problem - The Age of AI - The Coming Wave - Co-Intelligence - Life 3.0 - AI Superpowers - Supremacy Appendix: Top Questions Answered in This Article Which Company Has the Most Capable AI Model in July 2026? No provider can be declared the universal leader because performance depends on the task, service mode, cost, latency, tools, and evaluation method. OpenAI GPT-5.6 Sol, Anthropic Claude Fable 5, Google Gemini 3.1 Pro, and xAI Grok 4.5 belong to the leading proprietary group. Open and hybrid providers can be stronger choices when local control, licensing, price, or specialized deployment matters. What Is the Difference Between a Model Family and a Model Identifier? A family groups related models under one generation or brand, such as GPT-5.6 or Claude 5. A model identifier is the exact technical name used in an application programming interface. Identifiers matter because they can specify a tier, snapshot, preview, or dated version that behaves differently from the broad family name. Why Do Providers Offer Several Sizes of the Same Model Family? Different sizes balance capability, speed, hardware demand, and cost. A large model may handle complex reasoning and agents, and a small model may process classification, extraction, or high-volume requests more efficiently. Organizations often combine sizes so expensive models receive only tasks that need them. What Does Open-Weight Mean? Open-weight means the trained numerical parameters can be downloaded under stated license terms. It does not guarantee that the training data, full source code, training process, or commercial rights are unrestricted. Each license and release package requires separate review before deployment or modification. Are Preview Models Suitable for Production Systems? A preview model can be suitable for pilots and controlled use, but it carries greater change and retirement risk than a stable release. Providers may alter behavior, limits, pricing, or identifiers with less notice. Production use requires regression tests, fallback plans, and a clear understanding of the service commitment. Why Are Embedding and Reranking Models Important? Embedding models make documents and queries searchable by meaning rather than exact words. Rerankers reorder retrieved results so the most relevant material reaches the generative model. Weak retrieval can produce inaccurate answers even when the language model itself is highly capable. Do Large Context Windows Remove the Need for Retrieval? No. A large context window allows more material in one request, but cost, latency, and uneven attention can still limit results. Retrieval selects relevant material, preserves access controls, and reduces repeated input. Many systems benefit from combining long context with structured search and document management. Can an Organization Switch AI Providers Easily? Switching can be difficult when prompts, tool schemas, embeddings, fine-tunes, safety rules, or agent runtimes depend on one provider. Portability improves when teams use standard data formats, maintain independent evaluation suites, isolate provider-specific code, and preserve export procedures. Embedding indexes commonly need rebuilding after a model change. Why Can the Same Product Use More Than One Model? Products increasingly route requests according to complexity, modality, latency, safety, or subscription level. A consumer may see one product name even though a small model handles simple requests and a frontier model handles difficult work. Search, speech, image, and moderation components may also come from separate models. How Often Should an AI Model Catalog Be Updated? A public catalog should be checked at least monthly, and production inventories should monitor provider notices continuously. Fast release periods may require weekly review because flagships, aliases, previews, and retirement dates can change within days. Every entry should include a verification date and exact provider documentation. Appendix: Glossary of Key Terms Application Programming Interface A software interface that lets an application send requests to a model and receive outputs. Model providers use application programming interfaces to expose text generation, images, audio, embeddings, tools, and related services under defined technical and commercial terms. Agent A model-based system that plans and performs multi-step work using tools, memory, files, browsers, code, or business applications. An agent differs from a basic chatbot because it can take actions and continue working toward a task rather than produce one isolated response. Alias A convenient model name that can point to a provider-selected underlying version. Aliases simplify upgrades because application code may remain unchanged, but they can reduce reproducibility when the provider changes the model behind the name. Context Window The maximum quantity of tokens a model can process in one request, including input, output, and sometimes internal reasoning. A larger window can hold more documents or conversation history, though it does not guarantee equal recall across all supplied material. Deprecated Model A model that a provider discourages for new use and may remove after a stated date. Deprecation usually starts a migration period during which customers should test a successor, update code, and verify changed behavior. Embedding Model A model that converts text, images, audio, or other content into numerical vectors. Systems use those vectors for semantic search, clustering, recommendations, duplicate detection, and retrieval of material related by meaning. Foundation Model A broadly trained model that can be adapted to many tasks through prompting, tools, retrieval, fine-tuning, or application design. Language, multimodal, image, speech, video, and physical-world systems can all function as foundation models. Mixture of Experts A model architecture containing groups of parameters called experts, with selected experts activated for each token or input. The design can provide high total capacity with lower inference cost than activating the full model for every calculation. Multimodal Model A model that processes or generates more than one type of content, such as text, images, audio, video, or documents. Multimodal capability should be tested separately for each input and output type because performance can differ substantially. Open-Weight Model A model whose trained parameters are available for download. Rights to use, modify, redistribute, or commercialize the model depend on its license, and the release may not include the original training data or complete training code. Preview Model A publicly accessible model offered before the provider gives it full production status. Preview releases may change quickly, carry shorter support commitments, have lower quotas, or disappear sooner than stable models. Reasoning Model A model configured or trained to spend added computation on multi-step problems before producing an answer. Reasoning can improve mathematics, coding, planning, and analysis, though it commonly increases latency and cost. Reranker A model that receives an initial list of search results and reorders them according to relevance. Reranking often improves retrieval-augmented generation by placing stronger source material near the top of the model’s context. Snapshot A fixed, often dated model version that preserves behavior more reliably than a moving alias. Snapshots help regression testing and regulated deployment, though they may miss later performance improvements, bug fixes, or safety changes. Token A unit used by language models to represent pieces of text, code, or other encoded content. Providers commonly measure context limits and pricing in tokens, and one token does not consistently equal one word.