OpenAI ships GPT-5.3-Codex, retires legacy GPT-5 variants

OpenAI released GPT-5.3-Codex, billing it as its most capable agentic coding model to date, and simultaneously began retiring legacy GPT-5 Instant and Thinking variants — a forced-migration move that has created 'window is closing' pressure for developers still on older endpoints.

The model is being positioned squarely against Anthropic. Developers circulating early benchmarks said GPT-5.3-Codex 'absolutely demolished' Claude Opus 4.6 on Terminal-Bench 2.0, posting 77.3% versus 65.4%, while running roughly 25% faster with significant token-efficiency gains. OpenAI is also embedding Codex more broadly inside the ChatGPT app and shipping enterprise-focused Codex updates and business plugins.

The timing is pointed: GPT-5.3-Codex landed the same week Anthropic claimed 80% of its production code is Claude-authored and xAI announced its Cursor-trained Grok V9-Medium. Coding has become the most contested capability axis in frontier AI, because it is both a clear revenue driver (enterprise dev tools) and the substrate for agentic self-improvement.

The context to watch is regulatory: CEO Sam Altman met bipartisan lawmakers on June 4 to oppose government pre-approval requirements for model releases, a separate but related thread as ever-more-capable coding agents raise safety and liability questions. Skeptics also note benchmark leadership is fleeting — Terminal-Bench results can swing release to release, and real-world agentic reliability often diverges from leaderboard scores.