A 4B router model beats fixed Opus 5.5 on Terminal-Bench, 77.78 vs 69.44
YangYou1991 · x · 2026-11-07
InfRec's Launch Week finale: a 4B router model beats fixed Opus 5.5 on Terminal-Bench, 77.78 vs 69.44, and routes across 16 omni-modal benchmarks — suggesting Opus 5.5 isn't the ceiling of intelligence.
YangYou highlights the striking number: even picking the best model per task type in hindsight captured only 8% of the gain from routing per input. "Which model is best for coding?" is too broad a question.
More from coding & agent
- Solo founder claims $20K months running an entire team's work through Claude — PrajwalTomar_ · 2026-11-07
- Thoughtworks Technology Radar tabs mattpocock/skills as Trial for agentic coding — mattpocockuk · 2026-11-07
- OnCall-Memory: open-source incident response agent that learns from past production incidents — NewAlarm8213 · 2026-11-07
- Should MCP server operators trust agent identities issued by a third-party service? — zach-benson · 2026-11-07
- Build a 'strategist' agent: route every agent's accumulated context into key business decisions — EXM7777 · 2026-11-07
- Terrarium open-sources local VM sandbox giving AI agents explicit, auditable permissions — v_0009 · 2026-11-07