ImprovementModels ·
GPT-5.3 Codex and Gemini 3.1 Pro Preview
Try in the builder
Create an agent or open an existing agent in the builder, then follow the steps below. Model and feature availability may depend on your plan.
What it does
Two models joined the allowlist: OpenAI's Codex variant and Google's Gemini 3.1 Pro Preview.
Before this
Two additions to the selector, both filling gaps rather than replacing anything. Nothing on the list was trained specifically on code, and Gemini's long-context Pro line was not yet available.
Why it matters
Codex for agents that read or write code; Gemini 3.1 Pro for long-context work over big document sets.
The specs
- GPT-5.3 Codex
- 1,000,000-token context, 128,000-token output, text and images, Actions and reasoning (Fast to Smartest). Every plan.
- Gemini 3.1 Pro Preview
- 1,048,576-token context, 65,536-token output, text and images, Actions and reasoning. Gold and Pro.
Codex is trained on code, which shows less in whether it can write a function and more in how it handles structured output — JSON, config, and anything with a schema an agent has to hit exactly.