Yesterday OpenAI announced the release of GPT 5.4 thinking and GPT 5.4 pro. As the name of this newsletter is Price Per Token I am going to take a moment to point out the exuberant cost of GPT 5.4 pro:
The cost does come with a more powerful model, to be far. Notable is the increased context length of 1M tokens for both pro and thinking, that is up from 400k tokens for the 5.2 / 5.3 generation.
In terms of practical usage, computer use improvement was the prominent feature OpenAI talks about in their release blog. They trained the model to be able to handle high resolution images and issue accurate mouse clicks as a core capability.
It is able to receive screenshots (taken automatically and sent to it via an API) and perform actions without using a tool like Playwright. This opens up possibilities for it to control your computer outside of the browser environment. In the OSWorld - Computer Use test it is by far their best general purpose model at controlling a computer and performing slightly better than Codex.
This test works by giving a model access to a Ubuntu Desktop VM and giving it tasks like looking for emails in an email client or performing excel tasks. With its match of Codex in computer use and coding (it scores similar in coding benchmarks) OpenAI is making it the default model in the Codex app, replacing the specialized Codex model. This will bring an experience close to Claude Code where you have a general use model in a code-first app that can also perform other tasks.
It also brings new possibilities for developing using the OpenAI API. 5.4 exposes “computer” as a tool allowing you to more easily build apps that control the end user’s computer with simple prompts. No need to build anything custom, that is now a native feature of the model.
There are other improvements including a new strategy for tool calling that saves tokens and decreased latency, but the most significant improvements are covered above.