Repository navigation
v1.56.2
Haiku only for a quick question the local model is sure about
Model per message now sends a message to Haiku only when all of these are true:
- the bge-small vote is for Haiku at 70% or more;
- the message looks like a known example;
- it is short (20 words or fewer);
- it asks for nothing that sounds like work: making or changing something, a status check, a "why", something not showing up, a file path or a link.
Anything else runs on Sonnet. The word rules alone never pick Haiku, and a sure vote now overrides any rule that disagrees.
Measured on 3,000 real prompts: heavy turns sent to Haiku (15+ tool calls or 12k+ output tokens) fell from 167 (11.4% of heavy turns in v1.56.1) to 2 (0.1%). Decision time is unchanged (about 30 ms at the median).
For contributors
npm run router:eval measures the router on your own prompts. It runs outside the test suite and coverage, and nothing leaves the computer.
- It scores the 120 seed examples against each other.
- It replays up to 3,000 of your recent typed Claude Code prompts and compares each pick with how much work answering actually took.
- It writes a 300-prompt labelling sheet. Fill it in and re-run to get true accuracy.
Full Changelog: v1.56.1...v1.56.2