Superseded, correction (verified 2026-10-01)
The 2026-09-30 banner said
gpt-5.3-codexwas shut down on 2026-07-23: that was wrong. OpenAI’s deprecations page listsgpt-5-codex,gpt-5.1-codex,-max,-miniandgpt-5.2-codexfor 2026-07-23 (replacementsgpt-5.6-sol/gpt-5.6-terra), butgpt-5.3-codexappears there only as an example of a model class with 3-month notice. Its API model page shows it not deprecated (Responses API only; 400K context, 272K max input; knowledge cutoff 2025-08-31). In the Codex product, however, OpenAI’s models page says “GPT-5.2 and GPT-5.3-codex: already deprecated”, GPT-5.3-Codex-Spark retired 2026-09-14, GPT-5.4 and 5.4-mini retired 2026-08-31, GPT-5.5 retires from ChatGPT/Codex 2026-10-14, and Codex now recommends GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna (gpt-5-and-gpt-6-family). Net: an older, still-callable API model, no longer a Codex default. Dedicated “-codex” models are no longer OpenAI’s coding line; Codex (app, CLI, IDE, cloud) runs on the general GPT-6 models.
GPT-5.3 Codex
by OpenAI
OpenAI’s agentic coding model of early 2026 (announced 2026-02-05 per Wikipedia). No longer OpenAI’s most capable coding model.
See https://openai.com/blog/gpt-53-codex
Features
Per OpenAI’s API model page (fetched 2026-10-02): described as “the most capable agentic coding model to date” (launch wording), text and image input, low/medium/high/xhigh reasoning effort, Responses API only, function calling, web search, hosted shell and skills. Per Wikipedia (citing ZDNET): 25% faster than GPT-5.2-Codex, and described as OpenAI’s “first model that was instrumental in creating itself”. OpenAI’s launch post was not retrievable (403), so benchmark results and cybersecurity-classification claims from launch coverage are not repeated here. The API changelog lists later cyber-specific models (GPT-5.6 Cyber).
Availability
- Launch (2026-02-05, per Wikipedia): initially for ChatGPT Pro subscribers in Codex. Lighter variant GPT-5.3-Codex-Spark followed 2026-02-12 (research preview, text-only; Wikipedia). API release per changelog 2026-02-24 (Responses API).
- Now: still served on the API (not deprecated per its model page); deprecated for ChatGPT sign-in in Codex.
Pricing
API pricing is token-based with discounted cached input (model page, 2026-10-01); supports low, medium, high and xhigh reasoning effort.
Sources (accessed 2026-10-01 and 2026-10-02)
- https://developers.openai.com/api/docs/deprecations
- https://developers.openai.com/api/docs/models/gpt-5.3-codex
- https://developers.openai.com/api/docs/changelog
- https://learn.chatgpt.com/docs/models (Codex models; redirect from developers.openai.com/codex/models)
- https://github.com/openai/codex/discussions/17038
- https://en.wikipedia.org/wiki/GPT-5.3-Codex (search result; date)