Published leaderboards disagree because they measure different things on different prompts. This page separates the numbers we can verify today — live price and capability coverage — from the ones we have not earned yet.
Every measurement records the exact model ID and provider. Family-level averages hide the thing you are actually choosing.
Accuracy without cost is marketing. We report pass rate and billed cost together, so a cheap model that needs two attempts loses.
If we have not run it three times with a published prompt set, the cell stays empty. No unnamed benchmarks, no vendor slides.
Cheapest enabled model in each family, straight from the catalogue. Updated on every request.
| Family | Models | Cheapest input / 1M | Cheapest per call | Capability coverage |
|---|---|---|---|---|
| DeepSeek | 7 | $0.1493 | — | Reasoning |
| Qwen | 39 | $0.0746 | $0.0269 | ReasoningCodeVisionVideoImage |
| Z-ai | 7 | $0.1194 | — | ReasoningCode |
| Moonshot AI | 4 | $0.597 | — | Reasoning |
| Hunyuan | 2 | $0.1493 | — | Reasoning |
| ByteDance | 19 | $0.0448 | — | ReasoningVideo |
| MiniMax | 6 | $0.3134 | — | ReasoningCode |
| Kling | 2 | — | $0.1939 | Video |
Task sets are ten problems per category from real repositories, run three times per model at a pinned temperature with a fixed step budget. Until a row is filled, we link to the protocol instead of a score.
| Task set | Models in run | Runs per task | Pass rate | Cost / completed task |
|---|---|---|---|---|
| Repo edit (10 tasks) | — | 3 | Not yet measured | — |
| Test repair (10 tasks) | — | 3 | Not yet measured | — |
| Long-document QA (10 tasks) | — | 3 | Not yet measured | — |
| Video prompt adherence (10 tasks) | — | 3 | Not yet measured | — |
We could copy a leaderboard. Instead the protocol is published and the harness is being built against these task sets. When a row is filled it will carry a date, a model ID and a re-run command.
Read this week's radar →Send the model ID, prompt set, run count and cost basis to research@chinesemodels.ai. Reproducible results get published with attribution, including ones that contradict us.
Usage rankings (live traffic) →