Amp swaps Haiku 4.5 for Gemini 3 Flash in its codebase search subagent
Amp’s codebase search subagent now uses Gemini 3 Flash in place of Haiku 4.5, producing roughly 3× faster performance while maintaining the same result quality. The change centers on how the model coordinates tool calls and drives search iterations, resulting in noticeably lower latency for equivalent F1 scores.
What changed under the hood
The new subagent behavior is characterized by greater parallelism and shorter search loops:
- Parallel tool calls: Haiku 4.5 averaged about 2.5 parallel calls per iteration, whereas Gemini 3 Flash issues approximately 8 parallel calls.
- Iterations to completion: Searches that needed around 9 turns with Haiku now tend to finish in about 3 turns with Gemini 3 Flash.
- Exploration strategy: Gemini 3 Flash issues more diverse queries and can conclude early when sufficient evidence is gathered.
These shifts enable the subagent to explore the codebase more broadly in each turn and to terminate once confidence thresholds are met, rather than continuing longer multi-turn exchanges.
Performance profile
A performance vs. latency chart accompanying the announcement positions Gemini 3 Flash in the optimal zone: high F1 score combined with low latency. The visual emphasizes the tradeoff improvement — similar quality outcomes delivered with substantially reduced response times.
Context and effect
The swap affects Amp’s codebase search subagent specifically and is presented as a backend model change to improve throughput without degrading search accuracy. The core measurable benefits reported are increased parallelism per iteration and a reduced number of iterations to reach the same quality, producing the overall ~3× speedup.
Original source: https://ampcode.com/news/gemini-3-flash-search
