OpenAI just rolled out a new brain for its coding assistant, Codex. They call it GPT-5-Codex. This isn’t just another update. This version of GPT-5 learns to think for itself. It decides how much effort a coding task needs. The model is already live for a bunch of ChatGPT users. This includes Plus, Pro, Business, Edu, and Enterprise accounts. Early reports from TechCrunch say it’s nailing coding tests. It also delivers better code reviews, even with engineers watching.
This new GPT-5-Codex model can spend anywhere from a few seconds to seven hours on a single coding task. OpenAI made it available in many places. You can find it in terminals, integrated development environments (IDEs), GitHub integrations, and the ChatGPT interface itself. For now, only subscribers to ChatGPT Plus, Pro, Business, Edu, and Enterprise can use it. OpenAI plans to offer it to API clients later, but there’s no set date for that yet.
What’s New Compared to GPT-5
GPT-5-Codex really steps up its game against the standard GPT-5 model. TechCrunch points out that it performs much better on benchmarks specifically for agent-focused coding. It shows big improvements in SWE-bench Verified, for example. This benchmark measures how well AI agents can code.
OpenAI also ran tests where the model had to reorganize code from big, existing software projects. GPT-5-Codex did better than GPT-5 in these trials too. The company trained this new model especially for code reviews. Experienced software engineers looked at the model’s suggestions. They found fewer wrong comments and more high-impact ideas. These upgrades help Codex stand out in a busy market of AI coding tools.
🚀⚡ GPT-5-Codex: OpenAI raises the stakes
Adapts: ‘thinking’ seconds→7 hours
Already in Plus/Pro/Business/Edu/Enterprise
Better on SWE-bench; more precise reviews
API and costs still no date pic.twitter.com/SEdHlLiq3w— Diario฿itcoin (@Blaze Trends)
How the “Thinking” Works
Alexander Embiricos, who leads the Codex product at OpenAI, shed some light on this new “thinking” process. He compared it to how GPT-5 uses a “router” to send complex requests to different models. But GPT-5-Codex works differently. It doesn’t use a router. Instead, the model itself figures out in real-time how long it needs for a task.
This distinction is important. A router decides upfront how much power and time to give a task. GPT-5-Codex, however, can start with an estimate. If it realizes minutes later that the job needs more effort, it can reallocate resources. Embiricos mentioned seeing cases where the model took over seven hours to finish a tricky task.
Rollout and Access
As mentioned, the new model is already built into Codex products. You can get to it through terminals, IDEs, GitHub integrations, and the regular ChatGPT experience. OpenAI made sure that ChatGPT Plus, Pro, Business, Edu, and Enterprise users all have access. The company didn’t say if there are different features or limits across these plans.
OpenAI also promised that API customers will get GPT-5-Codex at a later time. They didn’t share specific dates or prices for this API offering. For developers and software teams, GPT-5-Codex is a big deal. It’s a tool that adjusts its effort based on how complex a task truly is. It can also provide more useful code reviews, according to the experts quoted by TechCrunch.
Competition in AI Coding Tools
This launch happens in a very competitive market. TechCrunch points out other tools like Claude Code, Anysphere’s Cursor, and Microsoft’s GitHub Copilot. The article notes that Cursor hit over $500 million in annual recurring revenue (ARR) by early 2025. This shows how fast this part of the market is growing.
They also mentioned Windsurf, a code editor. An attempt to buy Windsurf led to its team splitting up between Google and Cognition. These examples show why OpenAI is working hard to improve Codex. They want to stay strong against well-funded rivals and growing company adoption.
What This Means for Teams and Businesses
For engineers, the promise of more accurate and impactful code reviews could speed up their work. It might also cut down the time spent on manual reviews. Still, teams must check these automated suggestions. The TechCrunch article highlights that human engineers still rated GPT-5-Codex’s comments to ensure their quality.
For businesses, a model that spends hours on a single task raises questions. They’ll need to think about computing costs and how they set runtime policies for production. OpenAI hasn’t detailed pricing for variable compute time. They also haven’t set clear policies on maximum limits. Organizations will need more information before planning large-scale integrations.