GPT-5-Codex
OpenAI's version of GPT-5 optimized for agentic coding inside Codex. It was RL-trained on real-world coding tasks and served through Codex CLI, the IDE extension, cloud, and the Responses API. OpenAI shut
gpt-5-codexdown in the API on 2026-07-23 and namedgpt-5.6-solas its replacement. Codex now runs the GPT-6 family; OpenAI Models has the current lineup. Source: OpenAI deprecations, checked 2026-11-03
What it is
GPT-5-Codex is a GPT-5 variant fine-tuned for agentic coding in Codex or similar environments. Like its predecessor codex-1, it was trained with reinforcement learning on real-world coding tasks across many environments to produce code that mirrors human style and pull-request preferences, follow instructions precisely, and iteratively run tests until they pass. It was exposed through the OpenAI Responses API only, and the underlying snapshot was updated regularly until the 2026-07-23 shutdown. Source: OpenAI, https://openai.com/index/gpt-5-system-card-addendum-gpt-5-codex/, 2025 Source: OpenAI API docs, https://developers.openai.com/api/docs/models/gpt-5-codex, 2026-05-31; Source: OpenAI deprecations, checked 2026-11-03
How developers reach it
Codex (the product around the model) runs in several surfaces:
- Codex CLI - terminal-based agentic coding.
- IDE extension - in-editor agent, including VS Code-family editors.
- Cloud - Codex web, GitHub integration, and the ChatGPT mobile app for delegating tasks.
Safety is handled at both layers: model-level training against harmful tasks and prompt injection, plus product-level agent sandboxing and configurable network access so the agent's blast radius is bounded. Source: OpenAI, https://openai.com/index/gpt-5-system-card-addendum-gpt-5-codex/, 2025
Where it fits Kevin's stack
GPT-5-Codex is the OpenAI counterpart to Cursor's Composer (Cursor 2.0 and Composer) and the Claude Code harness (Everything Claude Code - Agent Harness Performance System) - a coding agent rather than a chat model. It is also a first-class alternative agent in security tooling: Deepsec - Agent-Powered Vulnerability Scanner runs against either Claude Opus or the GPT-5.x line at maximum reasoning. As an OpenAI model it slots directly into the OpenAI Agents SDK for tool-using agents. The Codex family has continued to advance (GPT-5.3-Codex was the top coding snapshot on 2026-05-31; OpenAI lists it shutting down on 2027-04-01), so treat "GPT-5-Codex" as the line, not a frozen checkpoint. Source: OpenAI API docs, https://developers.openai.com/api/docs/models/gpt-5-codex, 2026-05-31; Source: OpenAI deprecations, checked 2026-11-03
Timeline
- 2026-11-03 | Marked as shut down 2026-07-23 (replacement
gpt-5.6-sol); Codex runs the GPT-6 family. Source: OpenAI deprecations, checked 2026-11-03; OpenAI Models - 2025 | OpenAI shipped GPT-5-Codex with an addendum to the GPT-5 system card: RL-trained for agentic coding, available via Codex CLI, IDE extension, cloud, and Responses API, with agent sandboxing and configurable network access. Source: OpenAI, https://openai.com/index/gpt-5-system-card-addendum-gpt-5-codex/, 2025
- 2026-05-31 | Tool page created. The Codex coding-model line has iterated to GPT-5.3-Codex; GPT-5-Codex remains Responses-API-only with regularly updated snapshots. Source: OpenAI API docs, https://developers.openai.com/api/docs/models/gpt-5-codex, 2026-05-31; Source: OpenAI deprecations, checked 2026-11-03