How suggestions are made
The price builder suggests a model for a kind of work only when it has a published score on a benchmark built for that work. Among those, it picks the cheapest to run while staying close to the original model's quality.
| Work | Suggested | Benchmarks that count |
|---|---|---|
| Software | Qwen3.6 35B | LiveCodeBench, Terminal-Bench, SWE-bench |
| Agents | Qwen3.6 35B | Terminal-Bench, MCP-Atlas, SWE-bench |
| Documents and research | Qwen3.6 35B, Qwen3 Embedding 8B | GPQA Diamond, HLE, IFBench, MTEB |
| Support and questions | Qwen3.6 35B, Qwen3 Embedding 8B | GPQA Diamond, IFBench, MTEB |
Reading a score
- A score from the model's maker and a score from an independent board for the same test are shown separately, never averaged. They often differ.
- A benchmark's version is part of its name. Terminal-Bench 2.0 and 2.1 are different tests.
- You can still pick any model yourself, suggested or not.