The frontier AI models you can use without Chinese labs
A growing number of teams cannot use Chinese-made models at all: a government-device rule, a data-residency policy, or a procurement clause takes them off the table before capability is even discussed. This page ranks the frontier models that are left, made in the United States or Europe, two ways: by benchmark score and by value, the points you buy per dollar of tokens.
Which frontier AI model is best if you cannot use Chinese labs?
For teams barred from Chinese-made models, the strongest frontier LLM you can buy is GPT-5.6 Sol (80 of 100 on the coding composite). But the best value is GPT-5.6 Luna, at about 33.3 coding points per dollar of blended token price, roughly 8.8x the value of the priciest option, Claude Fable 5. All 7 models here are made in the United States or Europe, so none carry the procurement, data-residency, or export-control exposure that has pushed US agencies and enterprises away from Chinese labs.
Capability against price
Coding composite score against blended token price, for the non-China frontier. Better value sits lower and further right: a high score for a low price. The GPT-5.6 tiers and Grok 4.5 cluster in that corner; Claude Fable 5 and Claude Opus 4.8 sit high, where a top score comes at a premium token rate.
| Item | Coding composite score (0-100) | Blended price per Mtok (USD) |
|---|---|---|
| GPT-5.6 Luna | 75 | $2.30 |
| Grok 4.5 | 76 | $3 |
| Gemini 3.1 Pro | 69 | $4.50 |
| GPT-5.6 Terra | 77 | $5.60 |
| Claude Opus 4.8 | 74 | $10 |
| GPT-5.6 Sol | 80 | $11.30 |
| Claude Fable 5 | 77 | $20 |
The value ranking
Coding points per dollar of blended token price, best value first. The order is nearly the inverse of the raw-score order: the cheaper tiers score within range of the flagships at a fraction of the price, so they buy far more measured ability per dollar.
| Item | Value |
|---|---|
| GPT-5.6 Luna | 33.3 |
| Grok 4.5 | 25.3 |
| Gemini 3.1 Pro | 15.3 |
| GPT-5.6 Terra | 13.7 |
| Claude Opus 4.8 | 7.4 |
| GPT-5.6 Sol | 7.1 |
| Claude Fable 5 | 3.8 |
The full table
| Model | Origin | Coding | Blended $/Mtok | Value |
|---|---|---|---|---|
| GPT-5.6 LunaOpenAI | United States | 75 | $2.30 | 33.3 |
| Grok 4.5xAI | United States | 76 | $3 | 25.3 |
| Gemini 3.1 ProGoogle | United States | 69 | $4.50 | 15.3 |
| GPT-5.6 TerraOpenAI | United States | 77 | $5.60 | 13.7 |
| Claude Opus 4.8Anthropic | United States | 74 | $10 | 7.4 |
| GPT-5.6 SolOpenAI | United States | 80 | $11.30 | 7.1 |
| Claude Fable 5Anthropic | United States | 77 | $20 | 3.8 |
Value is the coding composite divided by the blended token price, ((3 × input) + output) ÷ 4, in dollars per million tokens. Higher is more coding ability per dollar.
Why exclude Chinese labs
For most teams this is a rule they inherit, not a benchmark call. On raw scores the leading Chinese open-weight models (GLM-5.2, Qwen3.7, Kimi, DeepSeek) are genuinely competitive, and several rank near the top on value in the full value leaderboard. What takes them off the table is policy. Several US Commerce Department bureaus and states including Virginia, Texas and New York have restricted DeepSeek on government devices, as reported by the South China Morning Post; a bipartisan bill introduced in 2025, the No Adversarial AI Act, would bar federal agencies from using AI models developed in China, Russia, Iran or North Korea. Beyond government, many enterprises apply data-residency or supply-chain rules that rule out models hosted by Chinese providers. This page answers the question those rules create: with the Chinese labs excluded, what is the best model left, and what does it cost.
Model availability is now a geopolitical variable in its own right. Claude Fable 5 spent three weeks suspended worldwide in June 2026 under a US export-control directive before being restored, the first known use of that authority against a specific commercial frontier model. For the wider picture of which countries produce which models, see AI by country.
How the ranking works
Two grounded inputs, one derived number. The benchmark scores are independent composite coding scores; for models the Price Per Token dataset has not composited yet (the GPT-5.6 tiers and Grok 4.5), the score is read directly from Artificial Analysis, the same independent benchmark family. The prices are this site's verified per-token API rates, the same numbers behind the model release tracker. From those two the page computes value: the coding score divided by a blended token price that weights input and output 3 to 1, the input-heavy mix an agentic coding session actually bills. Value is a cost-efficiency measure, not a verdict on quality: a model can lead on value and still be the wrong choice for work that needs the highest absolute score.
To turn these rates into the cost of a real job, use the cost-per-task calculator or put two models head to head. For the ranking across every model, Chinese labs included, see the value leaderboard.
Frequently asked questions
- What is the best non-Chinese AI model?
- Among frontier models made outside China that you can buy today, GPT-5.6 Sol posts the highest coding composite (80 of 100). On a value basis, points bought per dollar of blended token price, GPT-5.6 Luna leads at about 33.3 coding points per dollar. Which is "best" depends on whether the constraint is raw capability or capability per dollar.
- Why would you exclude Chinese AI models?
- For most teams it is a compliance or procurement rule, not a capability judgment. US federal bureaus and several states have restricted models such as DeepSeek on government devices, proposed federal legislation would bar Chinese-made models from federal agencies outright, and many enterprises apply data-residency or supply-chain rules that rule out models hosted by Chinese labs. On raw benchmarks the leading Chinese open-weight models are competitive; this page simply answers what is left once they are off the table.
- Which models count as frontier here?
- Models you can buy right now, made outside China, with an independent coding composite of 60 or higher. That floor is stated so the cut is reproducible. It yields 7 models from US and European labs. Lower-scoring open-weight models (for example Nvidia Nemotron, Mistral Devstral 2, and Llama 4 Maverick) are tracked on the full leaderboard but fall below the frontier line used here.
- Are there any European frontier models?
- France-based Mistral is the leading European frontier lab, and its open agentic coding model Devstral 2 is fully usable, but its coding composite currently trails the US frontier and falls below the floor used on this page. For teams that need EU-headquartered options specifically, Mistral is the primary choice today; on pure capability and value the ranking is dominated by US labs.
Sources
- Artificial Analysis. Independent LLM benchmarks and intelligence index (coding composite scores for the newest frontier models). https://artificialanalysis.ai/
- Price Per Token. LLM API Pricing and Benchmarks dataset (composite coding scores). https://pricepertoken.com/
- Cybernews (2025). US lawmakers push new bill to ban DeepSeek and other Chinese AI models across government agencies (the No Adversarial AI Act). Secondary reporting. cybernews.com
- South China Morning Post. China-tied AI tools like DeepSeek face US federal ban over threat to national security. Secondary reporting. scmp.com
- Capital & Compute. AI model value leaderboard and AI model registry (verified per-token prices). /ai-model-leaderboard/