GPT-5-Codex now available via API (platform.openai.com)

OpenAI has released GPT-5-Codex on the API — a GPT-5 variant tuned for "agentic" coding workflows. It brings a massive 400,000-token context window and up to 128,000 output tokens, image inputs (for context only) and enhanced reasoning-token support, with a September 30, 2024 knowledge cutoff. The model supports chat, responses, realtime, assistants and batch endpoints, and features streaming, function calling and structured outputs to make it easier to orchestrate tools and return machine-readable results. Fine-tuning, distillation and predicted outputs are not supported; snapshots let you pin specific model versions. Why it matters: the extended context and long-output capacity make GPT-5-Codex well-suited for multi-file code generation, long-running agent sessions, complex refactors and codebase-level reasoning that earlier models struggled with. Built-in function calling, streaming and structured outputs simplify integration with toolchains and developer assistants, while multiple endpoints (including realtime and assistants) enable low-latency agent deployment. Trade-offs include higher output-token pricing ($10 per 1M output tokens; $1.25 per 1M input tokens, cached input $0.125) and no fine-tuning, so behavior customization relies on prompts, snapshots and tool orchestration. Rate limits scale by usage tier (up to ~15k RPM and 10M TPM on top tiers), making the model usable from prototypes to large deployments.