Amp: Gemini 3 Pro as default model (ampcode.com)

Amp has switched its smart-agent default to Gemini 3 Pro after months of experiments with Claude, GPT-5, Gemini 2.5, Grok and others. The team picked Gemini 3 because it matches or exceeds Claude’s rare combo of speed, instruction-following and eagerness/ability to use tools: internal feedback was overwhelmingly positive and Terminal‑Bench 2.0 scores jumped ~17 percentage points when moving from Sonnet 4.5 to Gemini 3. Technically, Gemini 3 shows stronger codebase pattern adherence, cleaner tool interactions, fewer inappropriate behaviors (no emoji overuse, less blind affirmation), and higher dexterity in invoking and composing tools — all qualities that materially improve agent workflows and developer experience. Amp users who want to stay on Sonnet 4.5 must downgrade to an older Amp build (VS Code/Cursor/Windsurf extension v0.0.1763398025 or the CLI package @sourcegraph/amp@0.0.1763397734-g58e8bc via npm/pnpm/yarn). The team also lists remaining failure modes: occasional indefinite “thinking” loops, thinking-like prose leaking into outputs (very long token streams), control-character or fake tool-call leakage, reluctance to execute bash, and unwanted git commits. Despite these quirks, the move highlights how rapidly agent-capable models are improving and underscores the importance of ongoing prompt engineering, tooling safeguards and benchmark-driven evaluation as new base models are adopted.