OpenAI released GPT-5.4 on Thursday, its latest frontier model that combines coding improvements from GPT-5.3-Codex with two capabilities new to its mainline lineup, native computer use and a one-million-token context window. The model replaces GPT-5.2 across the API and Codex, OpenAI's coding tool.
GPT-5.4 marks the first time OpenAI has shipped computer use inside a production model rather than as a separate tool. The feature lets agents interact with desktop software by reading screenshots and issuing keyboard and mouse commands, completing tasks in what OpenAI describes as a "build-run-verify-fix loop." Anthropic shipped a similar capability last year with Claude, and Google has been testing agentic desktop control through Project Mariner.
One million tokens. That's the new context window, four times what GPT-5.2 offered, and big enough to fit a full codebase in one shot. OpenAI also baked in what it calls "native compaction support," which means the model taught itself to shed token weight during long agent runs without losing track of earlier steps. The practical pitch is straightforward: agents that can work longer before they start forgetting.
What Changed
- GPT-5.4 ships native computer use and a 1M token context window, replacing GPT-5.2 across API and Codex
- Input token pricing rises to $2.50 per million from $1.75, with five reasoning effort levels including new "xhigh"
- GPT-5.4 Thinking available to Plus, Team, Pro, and Enterprise users; Free and Go tiers excluded
- Microsoft's Claude adoption in Copilot 365 pressured OpenAI to sharpen its professional tools pitch
Pricing goes up, reasoning gets more granular
API customers will pay more for the upgrade. Input tokens cost $2.50 per million, up from $1.75 with GPT-5.2, according to Engadget. OpenAI pitches token efficiency gains to offset the increase, claiming GPT-5.4 completes multi-step agent tasks with fewer tool calls and less total token consumption.
Stay ahead of the curve
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
Reasoning effort now spans five levels: none, low, medium, high, and xhigh. The default remains "none," continuing the shift OpenAI started with GPT-5.2 away from always-on chain-of-thought reasoning. Temperature, top_p, and logprobs controls only work at the "none" setting. Developers who need deeper reasoning trade away those fine-grained sampling knobs.
The professional pivot sharpens
OpenAI positioned GPT-5.4 squarely at paying customers and developers. Inside ChatGPT, GPT-5.4 Thinking rolls out to Plus, Team, and Pro users starting Thursday, replacing GPT-5.2 Thinking. GPT-5.2 Thinking stays available under Legacy Models for three months before retirement on June 5. GPT-5.4 Pro is limited to Pro and Enterprise plans. Free and Go users get nothing. The company says individual factual claims are 33 percent less likely to be false compared to GPT-5.2, and responses with any errors dropped 18 percent. Accuracy numbers that matter most to the paying customers OpenAI is nervous about losing.
Microsoft added Anthropic's Claude models to Copilot 365 last September after finding Claude outperformed OpenAI's systems on spreadsheet and presentation tasks. That stung. OpenAI now generates roughly $25 billion in annualized revenue, The Information reported, but remains unprofitable with over $1.4 trillion in data center commitments on the books.
Tool search and deferred loading
A quieter addition, tool search, addresses a scaling problem for developers building agents with large tool libraries. Instead of loading every tool definition into context upfront, GPT-5.4 can defer tool schemas until runtime, loading only what it needs per request. OpenAI says this cuts token usage and improves both cache performance and tool selection accuracy.
Two modes. Hosted search, where candidate tools are known at request time, and client-executed search, where your application decides dynamically what to load. If you're running MCP servers or managing dozens of function definitions, the practical effect is lower latency and smaller context windows per call.
When GPT-5 first arrived in August 2025, early testers reported modest gains over GPT-4. Seven months and three point releases later, OpenAI has shifted the pitch from raw intelligence gains to infrastructure for autonomous work. Whether GPT-5.4's computer use and extended context actually deliver on that promise will depend on what developers build with it in the coming weeks.
Frequently Asked Questions
Is GPT-5.4 available to all ChatGPT users?
GPT-5.4 Thinking is available to Plus, Team, Pro, and Enterprise users. GPT-5.4 Pro is limited to Pro and Enterprise plans. Free and Go users don't get access.
How much does GPT-5.4 cost for API developers?
Input tokens cost $2.50 per million, up from $1.75 with GPT-5.2. OpenAI says the model uses fewer tokens per task to offset the price increase through better tool-call efficiency.
What is native computer use in GPT-5.4?
The model reads screenshots and issues keyboard and mouse commands to operate desktop software directly. It's the first OpenAI production model with this built in, similar to Anthropic's Claude computer use feature.
How does GPT-5.4's context window compare to competitors?
At one million tokens, GPT-5.4 matches Google Gemini's context capacity. It's four times GPT-5.2's 250,000-token limit, enough to process entire codebases in a single request.
What is tool search and why does it matter?
Tool search lets GPT-5.4 defer loading tool definitions until runtime, pulling only relevant schemas per request. This reduces token usage and latency for developers running large tool libraries or MCP servers.
