OpenAI upgrades Codex for autonomous coding

OpenAI is taking the fight directly to Cursor and Claude Code with its latest weapon: a custom-tailored coding model called GPT-5-Codex.

Rolled out to all paying ChatGPT subscribers and integrated across the terminal, IDEs, and GitHub, the new model replaces brute-force answering with dynamic, persistent problem-solving. A week after its initial debut, OpenAI also made it available via the Responses API at the same price point as standard GPT-5.

Real-time thinking that scales from seconds to 7 hours

Most current coding tools use model routers to estimate how hard a prompt is upfront. Codex product lead Alexander Embiricos explained that GPT-5-Codex ditches the router entirely, instead deciding dynamically on the fly how much compute to burn.

If a task looks quick, it replies in a snap. But five minutes into a messy refactor, it might decide it needs another hour to tinker, run test suites, and fix regressions. In OpenAI's internal testing, the agent worked independently for over seven straight hours on massive tasks until tests turned green.

Internal traffic numbers show just how wide the dynamic range is. On the simplest 10% of turns, GPT-5-Codex used 93.7% fewer tokens than the base GPT-5. On the toughest 10%, it spent more than double the time thinking, editing, and running tests. It also beat standard GPT-5 on SWE-bench Verified and dedicated code refactoring benchmarks.

CLI makeovers and autonomous PR reviews

Alongside the model, OpenAI gave the entire Codex workflow an overhaul. The open-source Codex CLI now handles images and wireframes, displays task progress via structured to-do lists, and integrates external data via MCP servers and web search. Permission modes have been streamlined into three simple security tiers.

In the cloud, container caching slashed median task completion times by 90%. The agent can now automatically install dependencies, fire up a headless browser to inspect frontend styling, and attach visual progress screenshots back into your GitHub pull requests.

Perhaps the biggest workflow upgrade is autonomous code review. Instead of relying on static syntax analysis, Codex checks PR descriptions against diffs, parses deep repo dependencies, and actually runs your tests. OpenAI claims the model reviews the vast majority of its own internal pull requests, flagging hundreds of bugs daily before human engineers even open the tab.

Why it matters

The AI coding space has exploded into an all-out turf war. With Cursor surpassing $500 million in ARR earlier in 2025 and Anthropic pushing terminal workflows with Claude Code, OpenAI could no longer rely on developers simply plugging standard GPT models into third-party editors.

We think the real breakthrough here isn't just raw benchmark scores; it's the shift toward persistent agency. By letting the model independently iterate against unit tests for hours, OpenAI is morphing Codex from a passive auto-complete widget into a tireless junior developer. For solo builders and lean startups, delegating tedious UI scaffolding, dependency upgrades, and sanity reviews is an immediate productivity win.

However, giving an autonomous agent extended access to your terminal and cloud environment carries real operational risks. While OpenAI locks down sandboxes by default, opening up MCP access and automated command execution still requires immense discipline. Developers will love the horsepower, but engineering leads will want to keep a close watch on edge-case hallucinations and runaway agent loops.