Chinese and US labs now both ship frontier-class models, but they publish them under very different patterns of licensing, pricing and access, so comparing them one model at a time hides the bigger picture. This page sets the latest flagship from seven Chinese providers beside the latest flagship from six US providers, using the same fields for each, and then compares the two groups on open weights, licences, price structures, deployment regions and access status. Providers are grouped by the country of their headquarters and listed alphabetically within each group; order is never a ranking. Every figure comes from the provider's own documentation, 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
- Thirteen flagships, two groups. Chinese providers covered: Alibaba (Qwen3.8-Max), Baidu (ERNIE 5.1), DeepSeek (DeepSeek-V4-Pro), MiniMax (MiniMax-M3), Moonshot AI (Kimi K3.1 Pro), Tencent (Hy3) and Z.ai (GLM-5.3). US providers covered: Anthropic (Claude Fable 5.2), Google (Gemini 4 Argon), Meta (Muse Wonder), Microsoft (MAI Thinking 1), OpenAI (GPT-6 Astra) and xAI (Grok 4.7).
- Open weights split along the group line at the flagship tier. Six of the seven Chinese flagships have downloadable weights on Hugging Face; none of the six US flagships does. US labs do publish smaller open-weight models, such as gpt-oss-120b, Gemma 4 and Muse Spark Open.
- Licences differ even where weights are open. Tencent uses Apache 2.0 for Hy3, while Alibaba, MiniMax, Moonshot AI and Z.ai use their own model licences for their largest open models, so the licence file has to be read before commercial use.
- Price structures differ as well as price levels. DeepSeek and Baidu use peak and off-peak rates, Alibaba prices by deployment scope, and Anthropic, OpenAI and xAI charge about 10% more for US-only or regional processing.
- Access is not uniform on either side. Google's Gemini 4 Carbon on Google Cloud is for verified enterprise customers, Anthropic's Mythos 5.2 is limited to verified organisations, Meta's Muse Wonder is in preview, and Baidu bills in Chinese yuan through its Qianfan platform.
How the models were selected
For each provider this page takes the model that the provider's own documentation describes as its most capable, flagship or frontier model, or, where no such wording is used, its newest generally priced general model, as of 15 November 2026. Providers are grouped by headquarters country. ByteDance (Seed) is not in the tables because its current model IDs and prices were confirmed only from secondary sources, and Amazon (Nova) is not in the tables because its current prices could not be read on the official pages; both are left out rather than estimated. Mistral AI (France) and other non-US, non-Chinese labs are outside this page's scope; the cross-lab view is in Frontier AI Models Compared, which this page complements rather than repeats.
| Group | Provider | Flagship covered | Basis for selection |
|---|---|---|---|
| China | Alibaba (Qwen) | Qwen3.8-Max | Hosted "official version" of the 2.4T-parameter Qwen3.8 model |
| China | Baidu | ERNIE 5.1 | Newest generally priced ERNIE model on Qianfan; ERNIE-5.5-Flash is a preview |
| China | DeepSeek | DeepSeek-V4-Pro (0813) | Larger of the two current API models |
| China | MiniMax | MiniMax-M3 | Current full-size model; M3.1-Flash is the smaller variant |
| China | Moonshot AI | Kimi K3.1 Pro | Newest K3-generation model |
| China | Tencent | Hy3 (Hunyuan) | Current released model; Hy4 is a preview |
| China | Z.ai (Zhipu) | GLM-5.3 | Described by Z.ai as its latest flagship model |
| US | Anthropic | Claude Fable 5.2 | Recommended for the most demanding reasoning and long-horizon agentic work |
| US | Gemini 4 Argon | Google AI Studio calls it its flagship reasoning model; Gemini 4 Carbon is the Google Cloud frontier model for verified enterprise customers | |
| US | Meta | Muse Wonder | "Meta's most capable model" |
| US | Microsoft | MAI Thinking 1 | Microsoft AI's reasoning model |
| US | OpenAI | GPT-6 Astra | "Our most capable model for the most demanding work" |
| US | xAI | Grok 4.7 | Newest Grok model on the xAI API |
Chinese and US flagships at a glance
As of 2026-11-15. "Not stated" means the provider did not publish the fact on the pages checked (2026-11-15).
| Group | Model | API ID | Context window | Max output | Input types | Open weights |
|---|---|---|---|---|---|---|
| China | Qwen3.8-Max | qwen3.8-max | Open weights 262,144 native, up to about 1,010,000 | Not stated by Alibaba | Text, image (hosted) | Yes (Qwen3.8-2.4T-A95B) |
| China | ERNIE 5.1 | ERNIE-5.1 | Not stated by Baidu (price tiers run to 128K input) | Not stated by Baidu | Not stated by Baidu | Not stated by Baidu |
| China | DeepSeek-V4-Pro | deepseek-v4-pro | 1M | 384K | Text | Yes |
| China | MiniMax-M3 | MiniMax-M3 | 1M | Not stated by MiniMax | Text, image, video | Yes |
| China | Kimi K3.1 Pro | kimi-k3.1-pro | 1,048,576 | Not stated by Moonshot AI | Text, image | Yes |
| China | Hy3 | Not stated by Tencent | 256K | Not stated by Tencent | Text | Yes |
| China | GLM-5.3 | glm-5.3 | 1M | 128K | Text | Yes |
| US | Claude Fable 5.2 | claude-fable-5-2 | 1M | 128K | Text, image | No |
| US | Gemini 4 Argon | gemini-4-argon | 1,048,576 (shared by input and output) | 65,536 per streaming segment; up to 1,048,576 with long decodes | Text, image, audio, video, PDF | No |
| US | Muse Wonder | muse-wonder-1.0 | 1,048,576 | Not stated by Meta | Text, image, video, audio, PDF | Not stated by Meta |
| US | MAI Thinking 1 | Not stated by Microsoft | 256K | Not stated by Microsoft | Text | No |
| US | GPT-6 Astra | gpt-6-astra | 1,050,000 | 128,000 | Text, image | No |
| US | Grok 4.7 | grok-4.7 | 500,000 | Not stated by xAI | Text, image | No |
Release dates stated by the providers: Claude Fable 5.2, 9 November 2026; Gemini 4 Argon, general availability 13 October 2026; Muse Wonder, announced 12 October 2026; MAI Thinking 1, public preview on Microsoft Foundry 12 August 2026; GPT-6 Astra, 3 September 2026; Grok 4.7, 21 September 2026; GLM-5.3, 18 August 2026; DeepSeek-V4-Pro, release date not stated on the model card checked; the API currently serves DeepSeek-V4-Pro-0813. Alibaba's model card cites August 2026 for Qwen3.8. Baidu, MiniMax, Moonshot AI and Tencent do not state a release date on the pages checked. The full release history for every provider is in the LLM Release Tracker.
How do Chinese and US flagship prices compare?
Standard per-million-token list prices on the provider's own API, as of 2026-11-15. Prices are in US dollars unless stated; DeepSeek and Alibaba show "$" without naming a currency code. Discounts, caching and regional rules are summarised in the notes column and explained in the next section.
| Group | Model | Input | Cached input | Output | Pricing notes from the provider |
|---|---|---|---|---|---|
| China | Qwen3.8-Max | $2 (Singapore, US); $1.65 (Global, Chinese mainland) | Not stated on the page checked | $6; $4.951 | Priced by deployment scope; tier up to 1M input tokens |
| China | ERNIE 5.1 | CNY 4 (up to 32K); CNY 6 (32K to 128K) | Not stated by Baidu | CNY 18; CNY 22 | Baidu quotes CNY per 1,000 tokens (0.004 and 0.018), converted here to per million; no batch option listed |
| China | DeepSeek-V4-Pro | $1.32 peak; $0.66 off-peak | $0.044; $0.022 | $3.96; $1.98 | Off-peak is half the peak rate |
| China | MiniMax-M3 | $0.30 (up to 512K); $0.60 (over 512K) | $0.06; $0.12 | $1.20; $2.40 | Priority tier at 1.5x; prices reflect a permanent 50% discount |
| China | Kimi K3.1 Pro | Not stated on the page checked | Not stated on the page checked | Not stated on the page checked | Batch billed at 60% of standard; automatic caching, 5-minute default TTL |
| China | Hy3 | Not stated by Tencent | Not stated by Tencent | Not stated by Tencent | Weights free to download under Apache 2.0 |
| China | GLM-5.3 | $1.40 | $0.26 | $4.40 | Also offered through the GLM Coding Plan |
| US | Claude Fable 5.2 | $8 | $0.20 | $40 | Batch 50% off; US-only inference 1.1x |
| US | Gemini 4 Argon | $2.00 to 30 Nov 2026; $4.00 from 1 Dec 2026 | $0.10; $0.20 | $10.00; $20.00 | Output includes thinking tokens; no free tier for this model |
| US | Muse Wonder | $4.00 | $0.20 | $20.00 | No long-context premium; Contributor mode not offered for this model |
| US | MAI Thinking 1 | Not stated by Microsoft | Not stated by Microsoft | Not stated by Microsoft | Enterprise use through Microsoft Foundry |
| US | GPT-6 Astra | $10.00 | $1.00 | $50.00 | Flex and Batch at half price; 10% uplift on regional processing endpoints for models released on or after 5 March 2026 |
| US | Grok 4.7 | $2.00 (under 200K); $4.00 (200K or more) | $0.50; $1.00 | $6.00; $12.00 | US regional endpoint 1.1x; long-context rates apply to the whole request |
Across the models with published prices, standard input prices for the Chinese flagships run from US$0.30 (MiniMax-M3) to US$2 (Qwen3.8-Max), and for the US flagships from US$2.00 (Gemini 4 Argon until 30 November 2026, Grok 4.7) to US$10.00 (GPT-6 Astra). These are list prices only: reasoning tokens, cache writes and long-context tiers change the real cost, as explained in LLM Pricing Explained. The LLM Pricing Tracker covers more models per provider.
Pricing structures: peak hours, scopes and regional uplifts
The two groups use different levers in their price lists, and the lever often matters more than the headline price.
| Pricing lever | Providers using it (from the pages checked) | What it means in practice |
|---|---|---|
| Time-of-day pricing | Baidu (ERNIE-5.5-Flash-Preview: peak 08:00 to 22:00, off-peak overnight); DeepSeek (peak 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays) | Batch or scheduled jobs moved off-peak cost half as much on DeepSeek |
| Deployment scope pricing | Alibaba (Singapore, US Virginia, Global, Chinese mainland) | The same model has different prices depending on where it is deployed |
| Regional or US-only processing uplift | Anthropic (1.1x for US-only inference); OpenAI (10% on regional processing endpoints); xAI (1.1x on the US regional endpoint) | Data residency choices add about 10% to token costs |
| Long-context tiers | Alibaba (by input tier on some models); Baidu (32K and 128K tiers); MiniMax (512K threshold); xAI (200K threshold) | Crossing the threshold can re-price the whole request |
| Priority or service tiers | MiniMax (Priority 1.5x) | Pay more for higher service priority |
| Batch or flex discounts | Anthropic (50%); Moonshot AI (batch at 60% of standard); OpenAI (Flex and Batch at half price) | Asynchronous workloads cost less |
| Scheduled price changes | Google (Gemini 4 Argon doubles from 1 December 2026) | Budget for the dated change, not today's rate |
| Training-use discount | Meta (Contributor mode for Muse Spark, not Muse Wonder) | Lower prices in exchange for allowing Meta to train on prompts and responses |
Is any Chinese or US frontier model open source?
"Open weights" means the trained parameters can be downloaded; it does not mean the training data or code are published, and each licence sets its own conditions. Background is in Open-Weight vs Closed Models Explained. The table lists the largest current open-weight model each provider publishes, as of 2026-11-15.
| Group | Provider | Largest current open-weight model | Parameters (total / active) | Licence |
|---|---|---|---|---|
| China | Alibaba | Qwen3.8-2.4T-A95B | 2.4T / 95B | Alibaba's own Qwen3.8-Max licence; smaller Qwen3.8-27B under Apache 2.0 |
| China | Baidu | ERNIE-5.5-Flash-Base (base checkpoint) | 280B / 14B | ERNIE 5.5 Flash Base Model License v1.0 |
| China | DeepSeek | DeepSeek-V4-Pro (Hugging Face card; API serves the 0813 version) | 1.6T / 49B | MIT (repository and weights) |
| China | MiniMax | MiniMax-M3 | About 428B / about 23B | MiniMax Community License |
| China | Moonshot AI | Kimi K3.1 Pro | 2.8T / 104B | Kimi K3.1 Pro License |
| China | Tencent | Hy3 (Hy4-preview also published) | 295B / 21B | Apache 2.0 |
| China | Z.ai | GLM-5.3 | 753B (Hugging Face size); active not stated | Z.ai's own GLM-5.3 licence; GLM-5.3-Flash under MIT |
| US | Anthropic | No open-weight model on the pages checked | Not applicable | Proprietary models |
| US | Gemma 4 (31B dense; 26B A4B mixture of experts) | 31B; 26B / 4B | Apache 2.0 | |
| US | Meta | Muse Spark Open (gated download) | 816B / 42B | Muse Model License, with limited commercial terms |
| US | Microsoft | Phi-4 family on Hugging Face | Up to 15B | See each model card |
| US | OpenAI | gpt-oss-120b | 117B / 5.1B | Apache 2.0 |
| US | xAI | No current open-weight model on the pages checked | Not applicable | Proprietary models |
The pattern on the pages checked: the Chinese providers release their largest or near-largest models as weights, often under custom licences, while the US providers that publish weights do so for models below their flagship tier, with the exception of Meta's 816B Muse Spark Open. Custom licences can add conditions a model card summary does not show, so read the licence file itself. Self-hosting brings its own security work, covered in the Self-Hosted LLM Security Guide.
Where are Chinese and US flagships available?
| Group | Model | Provider API and console | Other routes stated by the provider |
|---|---|---|---|
| China | Qwen3.8-Max | Alibaba Cloud Model Studio, with Singapore, US (Virginia), Global and Chinese mainland deployment scopes | Open weights on Hugging Face |
| China | ERNIE 5.1 | Baidu AI Cloud Qianfan, billed in CNY | Not stated on the page checked |
| China | DeepSeek-V4-Pro | DeepSeek API with OpenAI-format and Anthropic-format base URLs | Open weights on Hugging Face |
| China | MiniMax-M3 | MiniMax platform, Standard and Priority tiers | Open weights on Hugging Face |
| China | Kimi K3.1 Pro | Kimi platform API, with batch inference | Open weights on Hugging Face |
| China | Hy3 | Tencent TokenHub | Open weights on Hugging Face |
| China | GLM-5.3 | Z.ai API and GLM Coding Plan | Open weights on Hugging Face |
| US | Claude Fable 5.2 | Claude API | Amazon Bedrock and Claude Platform on AWS; Claude models on Google Cloud |
| US | Gemini 4 Argon | Gemini API (paid tier) in Google AI Studio | Google Cloud |
| US | Muse Wonder | Meta Model API (preview) | Muse Wonder Full for Muse Code subscribers |
| US | MAI Thinking 1 | MAI Playground for personal, non-commercial use | Microsoft Foundry for enterprise use |
| US | GPT-6 Astra | OpenAI API | Not stated on the model page |
| US | Grok 4.7 | xAI API, global endpoint and US regional endpoint | Not stated on the model page |
Two points matter for buyers outside China. First, an open-weight Chinese model can be run on infrastructure you choose, so the location of processing is set by your deployment rather than the provider. Second, a hosted API call goes to the provider's own service and is governed by that provider's terms and privacy policy, wherever the provider is based. DeepSeek's API also accepts OpenAI-format and Anthropic-format requests at separate base URLs, which makes switching easy but means base URLs and keys need careful control; see LLM Base URL and Endpoint Override Security.
Restricted-access and preview models
Access status is stated plainly for every provider in both groups:
- Alibaba: Qwen4-Max-Preview is listed as a preview on Model Studio; Qwen3.8-Max is on public API.
- Anthropic: Claude Mythos 5.2 is available only to organisations verified through Anthropic's verification programmes; Fable 5.2 is generally available.
- Baidu: ERNIE-5.5-Flash is a preview; ERNIE 5.1 is on public pay-as-you-go pricing.
- DeepSeek: both current API models are public; no verification step is stated.
- Google: Gemini 4 Carbon on Google Cloud is for verified enterprise customers in select regions; cyber-permissive variants of Gemini 4 models are allowlist only.
- Meta: Muse Wonder is in preview on the Meta Model API.
- Microsoft: MAI Thinking 1 is in public preview on Microsoft Foundry.
- MiniMax, Moonshot AI and Z.ai: the flagships are on public API, with no verification step stated.
- OpenAI: GPT-6 Astra is on public API; GPT-6 Cyber is restricted to approved defenders.
- Tencent: Hy4 is published as a preview; Hy3 is released.
- xAI: Grok 4.7 is on public API, with no verification step stated.
Benchmarks
Both Chinese and US labs publish benchmark tables for their own flagships, but each chooses its own tests, effort settings and comparison models, so provider-reported scores cannot be lined up fairly across the two groups. This page lists no scores for any provider. If you compare figures, use the same benchmark version and settings and a dated independent source, and keep provider-reported results separate. How to Read LLM Benchmarks sets out what to check.
Data handling and security
For any hosted model, data handling is set by the provider's API terms, privacy policy and the deployment route, not by the model or the provider's country alone. Points documented on the pages checked:
- Processing location: Alibaba publishes separate deployment scopes, including Chinese mainland (Beijing) and US (Virginia); Anthropic offers US-only inference; OpenAI offers regional processing endpoints; xAI offers a US regional endpoint.
- Training use: Google lists free-tier Gemini API content as used to improve products (Gemini 4 Argon has no free tier); Meta's Contributor mode trades lower prices for training use. Read each provider's terms for the plan you buy, and record the answer in your AI inventory.
- Self-hosting: open weights from either group let you keep prompts inside your own environment, but you take on patching, access control, logging and model provenance checks.
- Credentials: every hosted API is reached with an API key or cloud credential, which is a non-human identity. Store keys in a secrets manager, scope them per application and set spend limits; see the LLMjacking Guide.
Retention periods and training defaults by provider are covered in AI Data Retention and Deletion. Organisations with legal or policy restrictions on particular providers or jurisdictions should check those rules against both the hosted API and any self-hosted weights they plan to use.
How to choose between Chinese and US models: factual criteria
These are criteria, not recommendations. Score each candidate against your own workload, policies and evaluations.
| Criterion | Question to answer | Where to look |
|---|---|---|
| Policy and jurisdiction | Do internal policies, contracts or regulators restrict particular providers, countries or processing locations? | Your legal and procurement teams; provider terms |
| Deployment model | Will you call a hosted API or run weights on your own infrastructure? | Open weights and availability tables |
| Licence | Does the licence permit your commercial use, fine-tuning and redistribution? | The licence file in the model repository |
| Processing region | Can the provider process data where you need it, and at what uplift? | Pricing structures table; provider data residency pages |
| Cost at your volume | What is the blended cost with caching, off-peak windows, batch and long-context tiers? | Pricing tables and provider pricing pages |
| Currency and billing | Is the price in US dollars or Chinese yuan, and which billing entity contracts with you? | Provider pricing and terms pages |
| Access status | Is the model generally available, preview or verified only? | Restricted-access section |
| Change cadence | How often are models replaced, and how much notice is given? | LLM Deprecation and Retirement Tracker |
Small teams usually pilot on a hosted API with published prices; larger organisations with sovereignty or sector rules weigh processing location, licences and self-hosting more heavily.
Practitioner checklist
- Record each candidate model's provider, headquarters country, API ID, version and date checked in your AI inventory.
- Confirm with legal and procurement whether any policy restricts providers or processing locations before running a pilot.
- Read the full licence file for any open-weight model and record it in your software bill of materials.
- Model costs with the provider's real levers: off-peak windows, deployment scope, regional uplifts, batch discounts and dated price changes.
- Check the billing currency and contracting entity for each provider.
- Run your own evaluation set across at least one Chinese and one US candidate rather than relying on provider-reported benchmarks.
- Store every provider API key in a secrets manager, scope it per application and set spend alerts.
- Review provider pricing, release and deprecation pages monthly, since flagships in both groups change every few months.
Standards and references
- Alibaba Cloud: Model Studio model pricing (accessed 2026-11-15)
- Alibaba Qwen: Qwen3.8-2.4T-A95B model card (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Models overview (accessed 2026-11-15)
- Baidu AI Cloud: Qianfan model pricing (accessed 2026-11-15)
- DeepSeek: Models and pricing (accessed 2026-11-15)
- DeepSeek-V4-Pro model card, Hugging Face (accessed 2026-11-15)
- Google AI Studio: Gemini API pricing (accessed 2026-11-15)
- Google: Gemma 4 model overview (accessed 2026-11-15)
- Meta: Model API pricing and rate limits (accessed 2026-11-15)
- Meta: Muse Spark Open (accessed 2026-11-15)
- Microsoft AI: Models (accessed 2026-11-15)
- MiniMax: Pay-as-you-go pricing (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro model card (accessed 2026-11-15)
- Moonshot AI: Kimi platform chat pricing (accessed 2026-11-15)
- OpenAI: API pricing (accessed 2026-11-15)
- OpenAI: gpt-oss-120b model card (accessed 2026-11-15)
- Tencent: Hy3 model card (accessed 2026-11-15)
- xAI: API pricing (accessed 2026-11-15)
- Z.ai: Pricing (accessed 2026-11-15)
Related NHI Mgmt Group resources: Frontier AI Models Compared, DeepSeek Models, Alibaba Qwen Models, OpenAI Models and Anthropic Models.