Amp's Librarian gets ~3x faster and 43% cheaper on GPT-5.5

Amp has updated its Librarian — the component responsible for searching codebases — by switching its underlying model from Sonnet-4.6 (medium reasoning) to GPT-5.5 with no reasoning, adding WebSocket mode, and revising its system prompt to encourage more parallel exploration. The result is a mean latency drop from 237s to 81s (a 2.9x speedup) and a cost reduction from $1.21 to $0.69 per search (43% cheaper), with quality remaining essentially flat at an F1 score of 0.48 versus 0.47.

According to internal evals, the model switch to GPT-5.5 is responsible for the larger share of the latency improvement (~2.2x), while OpenAI's WebSocket mode contributes the remaining ~1.3x.

According to internal evals, the model switch to GPT-5.5 is responsible for the larger share of the latency improvement (~2.2x), while OpenAI's WebSocket mode contributes the remaining ~1.3x. Behaviorally, the Librarian now fires approximately 8 tool calls in parallel per turn, up from ~3 with Sonnet, and wraps up a typical search in ~5 turns rather than ~15. A concrete example cited in the post shows a Kubernetes HorizontalPodAutoscaler query that previously took 2 minutes and cost $1.08 on Sonnet-4.6 now completing in 40 seconds at $0.47 on GPT-5.5.

Key facts

1. 01The Librarian now runs on GPT-5.5 (no reasoning) with WebSocket mode, replacing Sonnet-4.6 (medium reasoning).
2. 02Mean latency dropped from 237s to 81s — a 2.9x speedup.
3. 03The model switch to GPT-5.5 accounts for ~2.2x of the latency gain; WebSocket mode accounts for ~1.3x.
4. 04Average cost fell from $1.21 to $0.69 per search (43% cheaper).
5. 05Quality held steady: F1 score of 0.47 (Sonnet-4.6) vs. 0.48 (GPT-5.5).
6. 06Parallel tool calls per turn increased from ~3 to ~8, and turns per search dropped from ~15 to ~5.
7. 07A sample query on Kubernetes HorizontalPodAutoscaler went from 2 minutes / $1.08 to 40 seconds / $0.47.

Topics