GPT-5.3-Codex

OpenAI's agentic coding model for Codex or similar environments. These notes cover model-level behavior; instruction files, permissions, and context discovery belong to the harness using it.

• Provide the complete task specification up front: outcome, constraints, relevant context, and completion conditions.
• Use medium reasoning effort for interactive coding when it provides enough quality; reserve high or xhigh for difficult, long-running work and validate the trade-off on representative tasks.
• Make persistence and the intended end state explicit for multi-step work so the model can distinguish a finished result from an intermediate artifact.
• State the required validation and tool outcomes rather than assuming that a plausible patch is sufficient.

• Copying Codex-specific system prompts or harness instructions into a model-neutral setup.
• Using the highest reasoning effort for every task without checking latency, token use, and quality.
• Treating the model's agentic coding orientation as a guarantee that the current harness exposes the required tools or permissions.

OpenAI documents GPT-5.3-Codex for agentic coding in Codex or similar environments. The model can be paired with other harnesses, but the available tools, context assembly, and approval behavior remain harness-specific.

The model page documents a 400,000-token context window, 128,000-token maximum output, and image input without audio or video support. Behavior and the useful reasoning-effort setting still depend on the workload and the surrounding harness.