Every major AI lab now ships a flagship model, and the specs, prices, licences and access rules differ so much that a side-by-side view built from the labs' own documentation is the only reliable way to compare them. This page sets out the latest flagship model from twelve providers, including Chinese labs, with the same fields for each: API ID, release date, context window, maximum output, input types, parameters, licence, price and access status. It is written for developers, architects, buyers and anyone tracking the frontier of large language models. Providers are listed alphabetically; order is never a ranking. The page gives facts only, with no verdicts, and where a provider does not publish a fact the table says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Model details change quickly, so confirm figures against each provider's current documentation, linked in the sources list.
Key takeaways
- Twelve flagships, twelve sets of rules. The models covered are Qwen3.8-Max (Alibaba), Claude Fable 5.2 (Anthropic), DeepSeek-V4-Pro (DeepSeek), Gemini 4 Carbon (Google), Muse Wonder (Meta), MAI Thinking 1 (Microsoft), MiniMax-M3 (MiniMax), Mistral Large 4 (Mistral AI), Kimi K3.1 Pro (Moonshot AI), GPT-6 Astra (OpenAI), Grok 4.7 (xAI) and GLM-5.3 (Z.ai).
- A 1M-token context is now common. Ten of the twelve providers state a context window of about one million tokens for their flagship; xAI states 500,000 for Grok 4.7 and Microsoft states 256K for MAI Thinking 1.
- List prices span more than a tenfold range. Standard input prices on the pages checked run from US$0.30 per million tokens (MiniMax-M3, up to 512K input) to US$10 (GPT-6 Astra), and output prices from US$1.20 to US$50.
- Six flagships publish weights. Alibaba, DeepSeek, MiniMax, Mistral AI, Moonshot AI and Z.ai publish weights on Hugging Face, each under its own licence; Anthropic, Google, OpenAI and xAI list their flagships as proprietary, and Meta and Microsoft do not state a weights release on the pages checked.
- Access is not always open. Gemini 4 Carbon on Google Cloud is for verified enterprise customers in select regions, Muse Wonder is in preview on the Meta Model API, MAI Thinking 1's playground is for personal, non-commercial use, and Anthropic's Mythos 5.2 (same capabilities as Fable 5.2) is limited to verified organisations.
How the flagship models were selected
A "frontier" or "flagship" model is the most capable general model a lab offers. To keep the choice objective, this page takes, for each provider, the model that the provider's own documentation describes as its most capable, its flagship or its frontier model, as of 15 November 2026. Where a provider uses different words on different pages, the choice and the wording are stated below.
| Provider | Flagship covered | How the provider describes it |
|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max | The Hugging Face card calls Qwen3.8-Max "the official version" of the 2.4T-parameter Qwen3.8 model, with more features than the open weights |
| Anthropic | Claude Fable 5.2 | Recommended for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort falls short |
| DeepSeek | DeepSeek-V4-Pro (0813) | The official release of DeepSeek-V4-Pro, superseding the preview version |
| Gemini 4 Carbon | "Google's frontier model" on Google Cloud; Google AI Studio calls Gemini 4 Argon its flagship reasoning model | |
| Meta | Muse Wonder | "Meta's most capable model" |
| Microsoft | MAI Thinking 1 | An expert reasoning model for complex problems |
| MiniMax | MiniMax-M3 | A native multimodal model with a 1M context |
| Mistral AI | Mistral Large 4 | An open-weight, general-purpose multimodal model |
| Moonshot AI | Kimi K3.1 Pro | "Our most capable 3T-class model to date" |
| OpenAI | GPT-6 Astra | "Our most capable model for the most demanding work" |
| xAI | Grok 4.7 | A frontier model for coding, agentic tasks and knowledge work |
| Z.ai (Zhipu) | GLM-5.3 | "Z.ai's latest flagship model" |
Providers not covered here, such as Amazon, Baidu, ByteDance, Cohere and Tencent, have their own model hubs on this site. Each provider's full line-up, including smaller and older models, is on its hub page, for example OpenAI Models, Anthropic Models and Google Gemini Models.
Frontier AI models at a glance
As of 2026-11-15. Source: each provider's model page, model card or release notes, listed in the sources section. "Not stated" means the provider did not publish the fact on the pages checked (2026-11-15).
| Provider | Model | API ID | Release date | Context window | Max output | Input types |
|---|---|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max | qwen3.8-max | August 2026 (model card citation) | 1M by default (hosted); open weights 262,144 native, up to 1,010,000 | Not stated by Alibaba | Text, image (hosted); text only (open weights) |
| Anthropic | Claude Fable 5.2 | claude-fable-5-2 | 9 November 2026 | 1M | 128K | Text, image |
| DeepSeek | DeepSeek-V4-Pro (0813) | deepseek-v4-pro | Not stated by DeepSeek | 1M | 384K | Text (vision not supported) |
| Gemini 4 Carbon | gemini-4-carbon | 10 November 2026 (GA on Google Cloud) | 1,048,576 (shared by input and output) | Up to 1,048,576 with long decodes; 65,536 per streaming segment | Text, image, audio, video | |
| Meta | Muse Wonder | muse-wonder-1.0 | 12 October 2026 (announcement) | 1,048,576 | Not stated by Meta | Text, image, video, audio, PDF |
| Microsoft | MAI Thinking 1 | Not stated by Microsoft | 12 August 2026 (public preview on Microsoft Foundry, introduction post) | 256K | Not stated by Microsoft | Text |
| MiniMax | MiniMax-M3 | MiniMax-M3 | Not stated by MiniMax (paper published 11 June 2026) | 1M | Not stated by MiniMax | Text, image, video |
| Mistral AI | Mistral Large 4 | mistral-large-4 (also mistral-large-4-0, mistral-large-latest) | 5 November 2026 (version 26.10, GA) | 1M | Not stated by Mistral AI | Text, image |
| Moonshot AI | Kimi K3.1 Pro | kimi-k3.1-pro | Not stated by Moonshot AI | 1,048,576 | Not stated by Moonshot AI | Text, image |
| OpenAI | GPT-6 Astra | gpt-6-astra | 3 September 2026 | 1,050,000 | 128,000 | Text, image |
| xAI | Grok 4.7 | grok-4.7 | 21 September 2026 | 500,000 | Not stated by xAI | Text, image |
| Z.ai (Zhipu) | GLM-5.3 | glm-5.3 | 18 August 2026 | 1M | 128K | Text |
Every model in the table outputs text. Several also offer reasoning or "thinking" modes: Z.ai states that GLM-5.3 reasoning is always enabled, with effort levels low, high and max; the Qwen3.8 open weights always begin with reasoning, while the hosted Qwen3.8-Max adds a non-thinking option; DeepSeek offers thinking (the default) and non-thinking modes. For background on these modes see Reasoning Models Explained.
When was each frontier model released?
Ordered by date, oldest first, using only dates the providers publish. Models with no published date are listed after the timeline.
- August 2026: Qwen3.8 open weights (the Hugging Face citation gives August 2026; Qwen3.8-Max is the hosted version).
- 12 August 2026: MAI Thinking 1 enters public preview on Microsoft Foundry, per Microsoft AI's introduction post.
- 18 August 2026: GLM-5.3 released, per Z.ai's release notes.
- 3 September 2026: GPT-6 Astra released in the OpenAI API, described in the changelog as "our most capable model, built for the hardest end-to-end work".
- 21 September 2026: Grok 4.7 made available on the xAI API as grok-4.7.
- 12 October 2026: Meta announces Muse Wonder and Muse Wonder Full.
- 5 November 2026: Mistral Large 4 version 26.10 reaches GA, per its model card.
- 9 November 2026: Anthropic launches Claude Fable 5.2 and, for verified organisations, Claude Mythos 5.2.
- 10 November 2026: Gemini 4 Carbon reaches GA on Google Cloud.
Not dated on the pages checked: DeepSeek-V4-Pro (0813), Kimi K3.1 Pro and MiniMax-M3 (its sparse attention paper is dated 11 June 2026). The LLM Release Tracker keeps a running, dated list of releases across all providers.
How much do frontier models cost?
As of 2026-11-15. US dollars per million tokens, standard tier, as listed on each provider's pricing or model page. Cached input is the price for cache hits or cached reads. Prices change often; check the linked pricing page before budgeting.
| Provider | Model | Input | Cached input | Output | Pricing notes from the provider |
|---|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max | 2.00 | Not stated by Alibaba (caching discount supported) | 6.00 | Singapore and US (Virginia) rows; the Global deployment row is 1.65 input and 4.951 output; single tier up to 1M tokens |
| Anthropic | Claude Fable 5.2 | 8.00 | 0.20 | 40.00 | Batch 4.00 input and 20.00 output; 5-minute cache writes 10.00, 1-hour 16.00; fast mode 16.00 and 80.00; US-only inference adds a 1.1x multiplier |
| DeepSeek | DeepSeek-V4-Pro (0813) | 1.32 (peak) | 0.044 (peak) | 3.96 (peak) | Off-peak is half the peak rate; peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday |
| Gemini 4 Carbon | Not stated by Google on the pages checked | Not stated by Google on the pages checked | Not stated by Google on the pages checked | Google Cloud supports Standard and Priority PayGo, Batch and Provisioned Throughput for this model | |
| Meta | Muse Wonder | 4.00 | 0.20 | 20.00 | Standard mode only; not offered in Contributor mode |
| Microsoft | MAI Thinking 1 | Not stated by Microsoft | Not stated by Microsoft | Not stated by Microsoft | Playground pricing not stated |
| MiniMax | MiniMax-M3 | 0.30 (up to 512K input) | 0.06 | 1.20 | Above 512K input: 0.60, 0.12 and 2.40; prices include a stated permanent 50% discount; Priority tier is 1.5x Standard |
| Mistral AI | Mistral Large 4 | 0.70 | 0.07 | 2.10 | From the model card |
| Moonshot AI | Kimi K3.1 Pro | Not stated by Moonshot AI on the pages checked | Not stated by Moonshot AI on the pages checked | Not stated by Moonshot AI on the pages checked | Pricing is on the Kimi platform pricing page |
| OpenAI | GPT-6 Astra | 10.00 | 1.00 | 50.00 | Cache writes 12.50; prompts over 272K input tokens cost 2x input and cache rates and 1.5x output for the whole request |
| xAI | Grok 4.7 | 2.00 | 0.50 | 6.00 | Higher rates apply to requests over 200K tokens |
| Z.ai (Zhipu) | GLM-5.3 | 1.40 | 0.26 | 4.40 | Cached input storage listed as free for a limited time |
Headline per-token prices are only part of the cost. Reasoning tokens are usually billed as output, long-context surcharges change the rate for the whole request on some models, and batch, off-peak or priority tiers move the price in either direction. Model Pricing Explained covers how tokens and caching are billed, and the LLM Pricing Tracker follows price changes across providers.
What is the context window of each frontier model?
The context window is the total number of tokens a model can consider in one request. The figures in the first table are the providers' own. Three points matter when comparing them:
- Shared or separate limits: Google states that Gemini 4 Carbon's 1,048,576-token limit is shared by input and output, while OpenAI lists a 1,050,000-token context and a separate 128,000-token output limit for GPT-6 Astra.
- Hosted versus open weights: the Qwen3.8 open weights are 262,144 tokens natively, extensible to 1,010,000; the hosted Qwen3.8-Max uses 1M by default.
- Price tiers by length: MiniMax prices M3 in two tiers split at 512K input tokens, OpenAI changes GPT-6 Astra's rate above 272K, and xAI changes Grok 4.7's rate above 200K.
See Context Windows Explained and LLM Context Windows Compared Across Providers for a wider view.
Is any frontier model open source?
Six of the twelve flagships publish model weights on Hugging Face. "Open weights" means the trained parameters can be downloaded; the licence decides what you may do with them, and several are custom licences rather than standard open source licences. As of 2026-11-15.
| Provider | Model | Open weights | Licence | Parameters | Knowledge cutoff |
|---|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max | Yes (Qwen3.8-2.4T-A95B) | Custom, listed as "qwen3.8-max" | 2.4T total, 95B active (mixture of experts) | Not stated by Alibaba |
| Anthropic | Claude Fable 5.2 | No | Proprietary | Not stated by Anthropic | June 2026 (reliable knowledge cutoff) |
| DeepSeek | DeepSeek-V4-Pro (0813) | Yes | MIT | 1.7T total; active count not stated | Not stated by DeepSeek |
| Gemini 4 Carbon | No | Proprietary | Not stated by Google | March 2026 | |
| Meta | Muse Wonder | Not stated by Meta | Not stated by Meta | Not stated by Meta | Not stated by Meta |
| Microsoft | MAI Thinking 1 | Not stated by Microsoft | Not stated by Microsoft | About 1T total, about 35B active (sparse mixture of experts) | Not stated by Microsoft |
| MiniMax | MiniMax-M3 | Yes | MiniMax community licence | About 428B total, about 23B active | Not stated by MiniMax |
| Mistral AI | Mistral Large 4 | Yes (Mistral-Large-4-1T-A52B) | Mistral Ecosystem Open Weight License | 1.05T total, 52B active, plus a 1.6B vision encoder | Not stated by Mistral AI |
| Moonshot AI | Kimi K3.1 Pro | Yes | Kimi K3.1 Pro License (custom) | 2.8T total, 104B active (mixture of experts) | Not stated by Moonshot AI |
| OpenAI | GPT-6 Astra | No | Proprietary | Not stated by OpenAI | 30 April 2026 |
| xAI | Grok 4.7 | No | Proprietary | Not stated by xAI | Not stated by xAI |
| Z.ai (Zhipu) | GLM-5.3 | Yes | Custom, listed as "glm-5.3" | 753B | Not stated by Z.ai |
Running a model of this size yourself needs serious hardware: the smallest open flagship here, MiniMax-M3, has about 428B parameters. Open-Weight vs Closed Models Explained covers the trade-offs, and the Self-Hosted LLM Security Guide covers the controls you take on when you host weights.
Where are frontier models available?
As of 2026-11-15, from the pages checked. Availability on third-party clouds and model marketplaces changes often and is not listed unless the provider's page states it.
| Provider | Model | Provider API and apps | Other routes stated by the provider |
|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max | Alibaba Cloud Model Studio, with Singapore, US (Virginia) and Global deployment scopes | Open weights on Hugging Face, compatible with vLLM, SGLang and TokenSpeed |
| Anthropic | Claude Fable 5.2 | Claude API, including fast mode and Batch API | Claude Mythos 5.2 offers the same capabilities to verified organisations |
| DeepSeek | DeepSeek-V4-Pro (0813) | DeepSeek API, concurrency limit 500 | Open weights on Hugging Face |
| Gemini 4 Carbon | Google Cloud (Gemini Enterprise Agent Platform), Global plus US and EU multi-regions | A cybersecurity variant, Gemini 4 Carbon for cyber, is documented alongside it | |
| Meta | Muse Wonder | Meta Model API (preview) and Muse Code | Muse Wonder Full for Muse Code subscribers |
| Microsoft | MAI Thinking 1 | MAI Playground, for personal, non-commercial use | Microsoft Foundry for enterprise evaluation and deployment |
| MiniMax | MiniMax-M3 | MiniMax API, Standard and Priority tiers | Open weights on Hugging Face |
| Mistral AI | Mistral Large 4 | Mistral API | Open weights on Hugging Face |
| Moonshot AI | Kimi K3.1 Pro | Kimi platform API (platform.kimi.ai) | Open weights on Hugging Face |
| OpenAI | GPT-6 Astra | OpenAI API: Responses, Chat Completions, Realtime, Batch and other endpoints | Not stated on the model page |
| xAI | Grok 4.7 | xAI API; us-east-1 listed, with regional pricing on other clusters | Not stated on the model page |
| Z.ai (Zhipu) | GLM-5.3 | Z.ai API and GLM Coding Plan | Open weights on Hugging Face |
Restricted-access and preview models
Access status is stated plainly for every provider:
- Anthropic: Claude Mythos 5.2 offers the same capabilities as Fable 5.2 only to organisations verified through Anthropic's verification programmes, such as the Cyber Verification Program. Fable 5.2 itself is generally available.
- Google: Gemini 4 Carbon on Google Cloud is available "to verified enterprise customers in select regions", which Google links to platform reliability, capacity and security.
- Meta: the Meta Model API documentation states that Muse Wonder is in preview; Muse Wonder Full is limited to Muse Code subscribers.
- Microsoft: the MAI Playground is for personal, non-commercial use; enterprise use runs through Microsoft Foundry.
- Others: the pages checked for Alibaba, DeepSeek, MiniMax, Mistral AI, Moonshot AI, OpenAI, xAI and Z.ai describe public API access for these flagships, with no verification step stated.
Benchmarks
Every lab publishes benchmark results for its own flagship, but each chooses its own tests, settings, reasoning effort and comparison models, so provider-reported scores are not comparable across labs. Meta's Muse Wonder announcement, for example, claims state-of-the-art agentic performance without giving scores and says a full report will follow. This page therefore lists no scores. When you compare benchmark figures, use the same benchmark version, the same effort setting and a dated independent source, and keep provider-reported and independent results apart. How to Read LLM Benchmarks explains what to check.
Data handling and security
Data handling is set by each provider's API terms and the deployment route, not by the model. Facts from the pages checked:
- Data residency and region: Alibaba prices Qwen3.8-Max by deployment scope (Singapore, US, Global); Anthropic offers US-only inference for Fable 5.2 at a 1.1x multiplier; Google lists Global, US and EU options for Gemini 4 Carbon processing; xAI lists us-east-1 for Grok 4.7.
- Training use: Meta offers a Contributor mode for Muse Spark at lower prices in exchange for training use; Muse Wonder is not offered in that mode. Check each provider's terms for your plan.
- Self-hosting: open weights let you keep prompts inside your own environment, but you then own patching, access control and logging for the inference stack.
- API keys: every hosted flagship is reached with an API key or cloud credential. Treat these as non-human identities: store them in a secrets manager, scope them per application and set spend limits. See the LLMjacking Guide and How to Set Spend Limits on LLM API Keys.
How to choose a frontier model: factual criteria
These are criteria, not recommendations. Score each candidate against your own workload and evaluations.
| Criterion | Question to answer | Where to find it |
|---|---|---|
| Access | Can your organisation use it today, or is verification, preview status or a subscription required? | Restricted-access section above; provider model page |
| Deployment | Do you need hosted API only, a specific cloud region, or downloadable weights? | Availability and open weights tables |
| Licence | Does the licence allow your commercial use, fine-tuning and redistribution? | The licence file linked from the model card |
| Context and output | Does your longest prompt plus expected output fit, and is the limit shared? | At-a-glance table |
| Input types | Do you need image, audio, video or PDF input? | At-a-glance table |
| Cost at your volume | What is the blended cost with caching, reasoning tokens, long-context tiers and batch? | Pricing table and provider pricing page |
| Data handling | Where is data processed, how long is it kept, and is it used for training? | Provider API terms and data controls |
| Change and retirement | How often does the provider replace models and how much notice does it give? | LLM Deprecation and Retirement Tracker |
For a smaller organisation, a hosted API with published pricing and no verification step is usually the quickest route to a pilot. A larger organisation with residency or sovereignty requirements will weigh regional processing options and open weights more heavily, and should plan for the operating cost of self-hosting.
Practitioner checklist
- Pin production code to an explicit API ID and record the model, version and date in your AI inventory.
- Run your own evaluation set on two or three candidates before switching; do not rely on provider-reported scores alone.
- Model the full cost, including reasoning tokens, cache writes, long-context surcharges and peak or priority tiers.
- Read the licence file before downloading open weights, and record the licence in your software bill of materials.
- Confirm access status (general availability, preview, verified only) before committing a delivery date.
- Check region and data residency options against your data protection obligations.
- Store API keys in a secrets manager, scope them per application and set spend alerts.
- Review provider release and deprecation pages monthly; flagships are being replaced every few months.
Standards and references
- Alibaba Qwen: Qwen3.8-2.4T-A95B model card (accessed 2026-11-15)
- Alibaba Cloud: Model Studio model pricing (accessed 2026-11-15)
- Anthropic: Models overview (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Claude Platform release notes (accessed 2026-11-15)
- DeepSeek: Models and pricing (accessed 2026-11-15)
- DeepSeek: DeepSeek-V4-Pro-0813 model card (accessed 2026-11-15)
- Google Cloud: Gemini 4 Carbon (accessed 2026-11-15)
- Google AI Studio: Models (accessed 2026-11-15)
- Meta: Model API models (accessed 2026-11-15)
- Meta: Model API pricing and rate limits (accessed 2026-11-15)
- Meta: Introducing Muse Wonder (accessed 2026-11-15)
- Microsoft AI: MAI Thinking 1 (accessed 2026-11-15)
- Microsoft AI: Introducing MAI Thinking 1 (accessed 2026-11-15)
- Z.ai: Release notes (accessed 2026-11-15)
- MiniMax: MiniMax-M3 model card (accessed 2026-11-15)
- MiniMax: Pay-as-you-go pricing (accessed 2026-11-15)
- Mistral AI: Mistral Large 4 model card (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro model card (accessed 2026-11-15)
- OpenAI: GPT-6 Astra model page (accessed 2026-11-15)
- OpenAI: API changelog (accessed 2026-11-15)
- xAI: Grok 4.7 model page (accessed 2026-11-15)
- xAI: Release notes (accessed 2026-11-15)
- Z.ai: GLM-5.3 guide (accessed 2026-11-15)
- Z.ai: Pricing (accessed 2026-11-15)
- Z.ai: GLM-5.3 model card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Small and Low-Cost LLMs Compared, Coding LLMs Compared, LLM Release Tracker, LLM Security Guide.