OpenAI has launched a new, improved version of the GPT model. GPT-5.4 can independently use tools, analyze images better, and consumes fewer tokens.
In a blog post, OpenAI announces GPT-5.4. The new version of the GPT model is immediately available via ChatGPT, Codex, or as an API. GPT-5.4 improves agentic and image processing capabilities, among other things, and should work more efficiently than GPT-5.2.
OpenAI has significantly ramped up the pace of GPT releases. GPT-5.2 was only announced in December. The GPT-5.3 Instant model, rolled out three days ago, turned out to be nothing more than a stepping stone to GPT-5.4. It appears in three versions: a base model, GPT-5.4 Thinking, and GPT-5.4 Pro.
Working independently
OpenAI states that GPT-5.4 is suitable for knowledge work and points to various benchmarks to support that claim. In practice, the model improves ChatGPT’s agentic capabilities to work independently. The model can use a computer without assistance and determine which tools or documents are needed for each task to execute it. This means you no longer have to provide ChatGPT with a list of which tools it should use.
GPT-5.4 is also much more flexible than its predecessors. The model can ‘change course’ in the middle of an answer to better align the output with what you need, without having to enter follow-up prompts. The Thinking model comes with better search skills and maintains context better for complex issues. Factual errors in responses have decreased by eighteen percent.
Better eyes
GPT-5.3 is not only a better thinker, but the model also sees more thanks to improved computer vision. Computer vision is the ability of an algorithm to extract and analyze meaningful data from images. This allows you to feed the model images at a very high resolution, up to 10 million pixels, without having to compress them. This way, details are not lost, and the model can also extract more precise information from the image.
According to OpenAI, this also benefits other skills. On the OSWorldVerified benchmark, a simulation that tests the ability to interact with applications, GPT-5.4 achieves a score of 75 percent, which is higher than the score the average human achieves on this test (72.4 percent). The model scores 87.3 percent on spreadsheet analysis: eight percent higher than GPT-5.2. The new version is also better at handling presentations and browsing the internet.
Less greedy with tokens
Improved capabilities often come with a higher cost, as large models swallow tokens faster. As a result, users sometimes hit their subscription limit in no time. This should not be the case with GPT-5.4: according to OpenAI, the model consumes up to seventy percent fewer tokens for similar prompts than GPT-5.2.
This is offset by a small price increase per token. The input price rises from 1.75 dollars per million tokens to 2.5 dollars, and the output price to 15 dollars per million tokens. For the Pro version, it is thirty dollars (input) and 180 dollars (output) per million tokens.
Diversionary tactic
The timing of the GPT-5.4 launch could not be better for OpenAI. The company is facing a lot of criticism for snatching the Pentagon contract away from Anthropic. OpenAI is accused of blatant opportunism and messy communication, and has even seen users switch to its rival. With a new AI model, OpenAI hopes the conversation will return to its technology.