OpenAI released GPT-5.4 this week. Here's what's actually new and why it matters for builders.

It can now operate your computer

GPT-5.4 is the first OpenAI model that can directly control a browser and desktop, clicking, typing, and navigating through software the same way a human would. Agents built on this model can log into apps, fill out forms, navigate dashboards, and complete multi-step workflows across software that was never designed for AI. One company running property tax portals reported a 95% success rate on the first attempt, and 100% within three attempts, compared to 73-79% with previous computer-use models. That's the kind of result that makes automation actually usable in production.

It produces noticeably better documents and spreadsheets

The clearest upgrade for most users is output quality. Spreadsheet modeling, presentations, and financial analysis have all improved substantially over GPT-5.2. Human raters preferred its presentations citing better aesthetics and more effective use of image generation. If you're using AI to produce client-facing documents, pitch decks, or financial models, this is the most relevant change.

Tool search cuts costs for agent workflows

Previously, if you built an agent with lots of tools connected, every API call loaded all the tool definitions upfront, burning tokens before the model did anything useful. The new tool search feature loads definitions on demand instead, cutting token usage roughly in half for tool-heavy workflows. For anyone building with MCP servers or running complex automations through the API, that's a real cost reduction.

You can redirect it mid-thought

In ChatGPT, you can now adjust the model's output mid-response, so you can course-correct while it's working rather than waiting for a full response and starting over. Useful for long research tasks or anything where you realise halfway through that you need a different angle.

Fewer hallucinations

Factual accuracy has improved meaningfully over GPT-5.2, both at the level of individual claims and full responses. Not zero errors, but a step forward for anything where accuracy matters.

Pricing and access

The standard model is $2.50 per million input tokens and $12 per million output tokens via the API. Available now in ChatGPT (as GPT-5.4 Thinking for Plus, Team, and Pro users), the API, and Codex. The Pro variant is significantly more expensive and aimed at scientific research, so most builders won't need it.