One day after the GPT-5.3 Instant release, OpenAI follows up with the big one: GPT-5.4 is here — in two variants. GPT-5.4 Thinking for all paid ChatGPT subscribers and GPT-5.4 Pro for $200/month Pro and Enterprise users. And it ships with a capability that’s been Anthropic’s home turf: native computer use.
What GPT-5.4 Brings to the Table
The headline feature: GPT-5.4 is the first OpenAI model that can operate computers autonomously. It can open applications, navigate interfaces, and execute complex workflows across multiple programs. Anthropic has had Computer Use with Claude for a while — now OpenAI catches up.
Beyond that, there’s a 1 million token context window in the API version. That’s the largest context OpenAI has ever offered, putting them on par with Claude Opus 4.6.
Efficiency improvements are notable too: GPT-5.4 uses 47% fewer tokens on certain tasks compared to its predecessors. And accuracy is up — individual claims are 33% less likely to be false, and complete responses contain 18% fewer errors compared to GPT-5.2, according to OpenAI.
Codex Comes to Windows
Alongside the model launch, OpenAI announced the Codex App for Windows. The desktop application targets ChatGPT Business customers and lets you run multiple Codex agents in parallel — with isolated worktrees and interoperability between CLI and IDE. That’s a direct shot at Claude Code.
My Take
OpenAI is catching up. Computer use, 1M token context, better accuracy — these are all features Claude users already know. But the fact that OpenAI bundles them in one package and launches a real developer offensive with the Codex Windows app shows the competition is heating up.
The interesting question is how well computer use actually works in practice with GPT-5.4. Anthropic had months of head start and just acquired Vercept to deepen its computer use capabilities. OpenAI’s first generation will be measured against that standard.
Sources: