Amp’s deep mode has a new engine: GPT-5.3-Codex. The update lands as a straightforward model swap on paper, but it changes the feel of deep mode in a way Amp frames as more fundamental—less about getting the agent to execute, more about being precise in directing it.
GPT-5.3-Codex becomes the default brain behind deep mode
With GPT-5.3-Codex now powering deep mode, Amp positions the model as “truly impressive” in practice and describes a shift in what limits outcomes. The core claim is that the coding agent is no longer the bottleneck; instead, the limiting factor is how well tasks are specified.
That framing fits an emerging pattern in AI-assisted coding tools: as models get more capable at multi-step work, the leverage increasingly moves to intent expression—clear constraints, tight definitions of “done,” and knowing what level of exploration is appropriate.
Amp also ties this change to a broader sentiment captured in its own post, “The coding agent is dead”, suggesting that the emphasis is moving away from the agent abstraction and toward the human-to-model interface: instructions, iteration, and control.
What’s different versus GPT-5.2-Codex
Amp describes GPT-5.3-Codex as better than GPT-5.2-Codex “in every dimension.” More specifically, it highlights a behavioral improvement that matters in day-to-day use: deep mode still “goes deep,” but it has “learned when to stop.”
In other words, the model is portrayed as retaining the ability to thoroughly investigate a codebase or problem, while reducing unnecessary digging when it’s not productive—an often-overlooked failure mode in “deep research” style coding workflows.
How to enable GPT-5.3-Codex deep mode in Amp
Amp says deep mode can be selected in two places:
- In the Amp CLI: open the command palette with
Ctrl-O, then runmode: use deep - In the Amp editor extension: select deep mode in the prompt field