Gemini 3 Pro is now the new main model in Amp, powering the smart agent mode. The team had previously used Anthropic’s Claude as the default after testing other models such as GPT-5 and Gemini 2.5. Gemini 3 Pro impressed the team with speed, instruction following, and tool use, and in an early internal run of Terminal-Bench 2.0 switching from Sonnet 4.5 to Gemini 3 raised the score by 17 percentage points. Staff reactions in internal channels were strongly positive, and the team notes the model still has imperfections.
The update ships alongside multiple product changes. Amp introduced a new rush mode that is 67 percent cheaper per token and 50 percent faster per token than smart; it is recommended for small, well defined tasks but not for complex refactors. Workspace settings are now supported via .amp/settings.json so project specific preferences and MCP server configurations can be stored per repository. Amp can now reference and fetch threads directly using a read_thread tool that retrieves a thread as Markdown and uses another model to extract relevant context. The CLI and editor extensions gained a command palette to run commands without editing prompts.
Several collaboration and tooling features were announced. The review panel in the VS Code extension provides an Artificial Intelligence summary of changes, guided tours through diffs, code navigation, and staging or committing with suggested commit messages to streamline human review of agent output. Compaction has been replaced by handoff, which extracts relevant context into new threads with a generated prompt and list of related files. Tooling improvements include toolboxes and commands amp tools make, amp tools show, and amp tools use to create, inspect, and run custom agent tools. The Librarian subagent lets Amp search public and private GitHub repositories after configuring a GitHub connection. Additional items in the release list include workspace entitlements, Amp Free changes that remove training opt in, and various performance and UX enhancements across the product.
