OpenAI just shipped GPT-5.4 and this one feels built for real work

A single tool-heavy task in OpenAI's GPT-5.4 launch materials drops from 123,139 tokens to 65,320 with tool search turned on. That is the kind of change developers actually feel in the bill, in the latency, and in whether an agent makes it to the end of a job without falling apart. If that claim survives contact with production, GPT-5.4 is not just another model refresh. It is a cost and reliability play for people who actually leave agents running.

That is why this release caught my attention on March 5, 2026. OpenAI is not pitching a vague intelligence jump. It is pitching a work model. Keep the coding gains from GPT-5.3-Codex, add native computer use, make tool-heavy workflows cheaper, and push the whole thing across ChatGPT, the API, and Codex. GPT-5.4 Thinking replaces GPT-5.2 Thinking for paid ChatGPT users, GPT-5.4 Pro is the higher-performance tier, and Codex gets experimental support for a 1M token context window.

What actually changed

I think the most important part of this release is that OpenAI finally made the product story more direct. GPT-5.4 is supposed to be the mainline workhorse, not just another branch in the model picker that only benchmark nerds can explain. The company literally calls it its "most capable and efficient frontier model for professional work," which is a much clearer promise than the usual model-picker soup.

• GPT-5.4 launches in ChatGPT as GPT-5.4 Thinking, in the API as gpt-5.4, and in Codex on the same day.
• GPT-5.4 Pro launches for harder tasks in ChatGPT and the API as gpt-5.4-pro.
• OpenAI says GPT-5.4 is its first general-purpose model with native computer-use capabilities.
• Codex gets experimental 1M context support, while requests above the standard 272K context window are billed at 2x usage.
• The new original image detail setting supports full-fidelity perception up to 10.24M total pixels or a 6000-pixel maximum dimension.

OpenAI is also saying the model is more factual than GPT-5.2, with flagged factual claims 33% less likely to be false and full responses 18% less likely to contain any errors. Those are the sort of improvements that matter more in spreadsheet work, coding sessions, and research tasks than they do in screenshot-friendly prompt demos.

OpenAI describes GPT-5.4 as its "most capable and efficient frontier model for professional work."

The benchmark story is less boring than usual

I do not care much about tiny benchmark deltas when they never show up in real work. This launch has a few numbers that are harder to shrug off. On SWE-Bench Pro, GPT-5.4 scores 57.7%, only a modest lift over GPT-5.3-Codex at 56.8%. On OSWorld-Verified, though, it jumps to 75.0%, ahead of GPT-5.2 at 47.3% and even above the cited human score of 72.4%. BrowseComp also moves to 82.7% from 65.8% on GPT-5.2.

That is why I think the real story is not raw coding alone. It is long-horizon execution. OpenAI is clearly betting that developers and teams care more about models that can keep context, use tools correctly, and finish multi-step work without burning tokens or dropping the thread halfway through.

What native computer use means in practice

This is the section I wish more launch posts explained clearly. The native computer-use claim matters because GPT-5.4 is being tuned for jobs that involve tools, software environments, and real interfaces, not just answer generation. The cleanest example in OpenAI's release is MCP tool search. Instead of stuffing every tool definition into the prompt up front, GPT-5.4 can search for the right tool when it needs it, keep the cache healthier, and avoid paying for the same giant block of tool text on every step.

That also makes the 1M Codex context option more interesting than it first looks. I do not read it as a marketing flex. I read it as OpenAI trying to keep a model stable across a large repo, a large tool catalog, and a long chain of edits without forcing the user to constantly start over.

Why the tool search claim matters more than the price bump

OpenAI raised API pricing versus GPT-5.2. GPT-5.4 moves to $2.50 per million input tokens and $15 per million output tokens. GPT-5.4 Pro lands at $30 input and $180 output. On paper, that is not cheap. But OpenAI is making a specific counterargument. If the model wastes fewer tokens while searching tools and carrying out long jobs, the total task cost can still come down.

The proof point it chose is a good one. On Scale's MCP Atlas benchmark, OpenAI says GPT-5.4 tool search reduced token usage by 47% while preserving accuracy. That is not abstract. In the chart above, the example task drops from 123,139 total tokens to 65,320. If your team runs the same kind of tool-heavy workflow all day, that difference adds up fast.

Where I would still be careful

I would not treat the launch table as a free pass to swap every workflow over tomorrow morning. The SWE-Bench Pro gain over GPT-5.3-Codex is real, but small. That tells me the easy story here is not "coding got solved." The harder and more believable story is that OpenAI improved the messy parts around coding, tool usage, memory, and computer control. That matters, but it needs real evals on your own stack before it deserves blind trust.

I would also watch how the 1M Codex context option behaves under pressure. Huge context windows are only impressive when retrieval, tool choice, and edits stay coherent. If GPT-5.4 can hold together across a large repo plus a large MCP catalog, that is valuable. If it just turns into a more expensive way to lose focus later in the run, nobody will care how pretty the launch chart looked.

The community reaction is already split

I checked the launch-day Reddit threads as well, and the reaction is exactly what you would expect. One side is impressed by the OSWorld and BrowseComp gains, especially for Codex and agent workflows. The other side is already asking whether this is just benchmaxing with a nicer name and whether day-to-day prompting really feels different from 5.3. That skepticism is healthy. GPT launches usually look strongest on day one and messier a week later when people hit edge cases.

Still, I do not think this is an empty packaging update. The 1M Codex context option, native computer use, better image fidelity controls, and cheaper tool-heavy execution all point in the same direction. OpenAI wants GPT-5.4 to be the model you leave running while real work happens.

The bet OpenAI is making

My read is that OpenAI is trying to move the conversation away from chatbot vibes and back toward execution. Less back-and-forth, more tool use, more software control, longer context, and less token waste. That is a better story than arguing over one more abstract reasoning chart.

The open question is whether GPT-5.4 keeps this shape outside the benchmark lab. If it really reduces retries, preserves context, and stays accurate across long jobs, it will earn the price increase. If it does not, developers will notice fast, because this release is aimed at workflows where wasted tokens and broken loops are impossible to hide. If you are building an agent stack this month, this is one of the few launches worth testing with real evals instead of vibes.

Primary sources

• OpenAI launch post for GPT-5.4
• OpenAI rollout and pricing details for GPT-5.4 Thinking and GPT-5.4 Pro
• OpenAI GPT-5.4 system card
• OpenAI developer docs for image detail levels
• Reddit launch discussion
• Reddit benchmark reaction thread