Compare · Chinese vs American models
How close are Chinese models, and what do they cost?
Published benchmark results next to each lab's list price per million tokens. Every number links to its source; anything we could not find a published figure for says n/a.
Side by side
Sorted by Artificial Analysis Intelligence Index within each group. Highest score and lowest output price in the table are highlighted.
Savings calculator
Price your own workload
Pick the American model you use and your monthly tokens. Every assumption is listed under the result.
How to read this
- AA Index is the Artificial Analysis Intelligence Index, a composite of many evaluations run by a third party on each model's highest-effort setting. It is the only measure that covers nearly every model here, so it leads the table.
- Terminal-Bench (agentic tool use in a terminal), SWE-bench Verified (fixing real GitHub issues) and GPQA Diamond (graduate-level science questions) come from Vals AI's independent runs. Vals did not run several of the newest American models, so those cells are n/a rather than guessed.
- Long ctx is Artificial Analysis long-context reasoning (AA-LCR).
- Lab-published scores are often higher than third-party runs because harnesses differ. We use third-party numbers for comparability; see the source notes.
- Prices are list prices with no caching, batch or off-peak discounts. DeepSeek charges about half outside peak hours. Google has announced Gemini price increases for December 2026 and January 2027.
Sources
The full per-number source list, with read dates and caveats, ships with this site in assets/data/SOURCES.md. Main sources: Artificial Analysis leaderboards, Vals AI benchmark pages, and each lab's official API pricing page (DeepSeek, Alibaba Cloud Model Studio, Moonshot Kimi, Z.ai, MiniMax, Anthropic, OpenAI, Google, xAI).