Addendum to GPT-5 system card: GPT-5-Codex
OpenAI
OpenAI dropped a coding-focused GPT-5 variant called GPT-5-Codex. It thinks fast on easy tasks and grinds longer on hard ones, on its own.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has quietly published an addendum to the GPT-5 system card, and buried in the paperwork is a new model worth paying attention to: GPT-5-Codex. It's not a fresh foundation model so much as a specialized cut of GPT-5, tuned specifically for the kind of agentic coding work that happens inside Codex, OpenAI's coding tool.
The headline feature is how it manages its own effort. Instead of applying a flat amount of reasoning to every request, GPT-5-Codex scales its thinking time based on what the task actually demands. Ask it something simple, a quick syntax question or a small function tweak, and it answers fast, the way you'd want from a tool you're using in a live coding session. But hand it something gnarly, a multi-file refactor or a bug that requires tracing logic across a codebase, and it will sit with the problem longer, working independently rather than bailing out early or padding its response with filler reasoning.
That dynamic allocation matters more than it might sound. Anyone who's used LLM coding assistants knows the frustration of a model overthinking a one-line fix or, worse, rushing through something that needed careful multi-step reasoning. OpenAI is essentially trying to make the model behave more like a competent engineer who knows when to move fast and when to slow down, rather than a system that treats every prompt with the same fixed budget of compute.
This is also a signal about where OpenAI thinks the real product battle is. Agentic coding tools have become one of the most competitive corners of the AI industry, with GitHub Copilot, Cursor, Cognition's Devin, and others all racing to prove their models can work autonomously on real engineering tasks, not just autocomplete lines. By shipping a GPT-5 variant specifically optimized for that use case, and documenting it formally in a system card addendum rather than a marketing post, OpenAI is treating agentic coding less like a feature and more like its own category requiring dedicated safety and capability review.
My take — AI-written commentary, not fact-checked reporting
I like that OpenAI is willing to fork GPT-5 into task-specific variants instead of pretending one model should be great at everything, that's the honest engineering move. But calling this a system card addendum rather than a proper launch tells you OpenAI is still hedging on how much autonomy it's comfortable admitting these coding agents actually have.
Read more about this at: OpenAI