OpenAI released GPT-5.4 today — and this one landed fast. Just two days after rolling out the GPT-5.3 Instant update, the company pushed out what it's calling its most capable and efficient model yet. It's available now across ChatGPT, the API, and OpenAI's Codex developer platform.

Here's the short version: GPT-5.4 merges the coding strengths of GPT-5.3 Codex with better reasoning, fewer errors, and dramatically improved efficiency across professional tasks like documents, spreadsheets, and presentations. It also introduces native computer use — meaning AI agents can now control a computer to complete multi-step workflows across apps. And for the first time, you can interrupt ChatGPT mid-thought and redirect it before it finishes.

Three Versions, Different Access Levels

GPT-5.4 comes in three flavors:

• GPT-5.4 Thinking — the default reasoning model, replacing GPT-5.2 Thinking. Available now for ChatGPT Plus ($20/month), Team, and Pro users.
• GPT-5.4 Pro — a maxed-out version for complex tasks. Reserved for Pro ($200/month) and Enterprise plans.
• GPT-5.4 (API) — available to developers through the API and Codex, with support for up to 1 million tokens of context.

Free ChatGPT users will also get access, but only when their queries are automatically routed to the model — not on demand.

GPT-5.2 Thinking will stick around in the model picker under "Legacy Models" for paid users until June 5, 2026, then it's gone.

What's Actually Better

You can now interrupt it mid-thought. This is the feature most regular users will notice first. GPT-5.4 Thinking now shows you an upfront plan of its reasoning before it starts generating the full response. You can click "Update" in the sidebar to add context, correct direction, or change priorities — and the model adjusts on the fly without starting over. If you've ever watched ChatGPT burn through a long response that's heading in the wrong direction, this is the fix. Available now on chatgpt.com and Android, with iOS coming soon.

Fewer errors. OpenAI says GPT-5.4 is 33% less likely to make factual mistakes on individual claims compared to GPT-5.2, and overall responses are 18% less likely to contain any errors. That's a meaningful improvement for anyone using ChatGPT for research, writing, or analysis.

Way more efficient. This is the big one for developers and power users. GPT-5.4 uses up to 47% fewer tokens to solve the same problems — that's the internal "thinking cost" the model burns through when reasoning. Fewer tokens means faster responses and lower API costs. OpenAI argues this efficiency offsets the higher per-token pricing (more on that below).

1 million token context window (now out of beta). In the API and Codex, GPT-5.4 supports up to 1M tokens of context — that label is no longer in beta. That's roughly 750,000 words — enough to feed it an entire codebase, a long legal contract, or months of meeting transcripts in a single prompt. Worth noting: OpenAI charges double the per-token rate once input exceeds 272,000 tokens.

Native computer use. GPT-5.4 is OpenAI's first general-purpose model with built-in computer control capabilities. In Codex and the API, agents can now operate a desktop — clicking, typing, navigating between apps — to complete complex workflows. On the OSWorld benchmark (which tests desktop navigation), GPT-5.4 hit 75% success rate, up from 47.3% with GPT-5.2. For reference, human performance on the same test is 72.4%. On BrowseComp (web browsing tasks), GPT-5.4 Pro hit 89.3% — a new state of the art.

Stronger coding. GPT-5.4 incorporates the coding capabilities from GPT-5.3 Codex while matching or beating it on SWE-Bench Pro (57.7% vs 55.6%). A new /fast mode in Codex delivers 1.5x faster token velocity — same model, same intelligence, just quicker. OpenAI also released an experimental Codex skill called "Playwright (Interactive)" that lets the model visually debug web apps in real time.

Professional task performance. On OpenAI's GDPval benchmark — which tests how well AI handles real knowledge work across 44 occupations — GPT-5.4 matched or outperformed human professionals 83% of the time, up from 70.9% with GPT-5.2. On spreadsheet modeling tasks typical of a junior investment banking analyst, it scored 87.3% — up from 68.4% with GPT-5.2. Human raters preferred its presentations 68% of the time over GPT-5.2's output, citing better design and visual variety.

Better at legal and financial work. Harvey, the AI legal platform, tested GPT-5.4 on its BigLaw Bench evaluation and reported a 91% score — calling it a new bar for document-heavy legal work. Zapier's CEO said GPT-5.4 "finished the job where previous models gave up" on their multi-step tool use benchmarks.

Improved vision. OpenAI introduced a new "original" image input detail level that supports up to 10.24 million pixels — a big upgrade for anyone working with dense documents, high-res images, or screenshots. This feeds directly into better document parsing and computer use accuracy.

New for Developers: Tool Search

If you build with the OpenAI API, this one matters. GPT-5.4 introduces Tool Search, a new system for managing tool calling. Previously, every tool definition had to be loaded into the system prompt — which got expensive as the number of tools grew. Now the model can look up tool definitions on demand, cutting token usage by 47% in testing while keeping accuracy the same.

API Pricing: Higher Per Token, But Fewer Tokens

This is worth calling out. GPT-5.4 costs more per token than GPT-5.2:

• GPT-5.4 Thinking: $2.50 input / $0.25 cached input / $15 output per million tokens (up from $1.75 / $0.175 / $14 for GPT-5.2)
• GPT-5.4 Pro: $30 input / $180 output per million tokens (up from $21/$168 for GPT-5.2 Pro)

Batch and Flex processing are available at half the standard rate. Priority processing costs double.

That's a notable price increase. OpenAI's argument is that the 47% token efficiency makes up for it — you're paying more per token, but using far fewer of them. Whether that math works out depends on your specific use case. For simple queries, you'll likely pay a bit more. For complex reasoning tasks that previously burned through tokens, you could actually save money.

For developers, the API model strings are gpt-5.4 and gpt-5.4-pro.

ChatGPT for Spreadsheets

Alongside GPT-5.4, OpenAI launched ChatGPT for Excel and Google Sheets (in beta). This embeds ChatGPT directly into your spreadsheets to build, analyze, and update financial models using the formulas and cell structures you already have. They're also launching integrations with FactSet, MSCI, Third Bridge, and Moody's — aimed squarely at finance teams.

On an internal investment banking benchmark, model performance jumped from 43.7% with GPT-5 to 88% with GPT-5.4 Thinking.

The Bigger Picture

This release comes at a turbulent time for OpenAI. The company is managing fallout from its Pentagon deal, a viral #QuitGPT movement, and reports of slowing user growth. GPT-5.4 feels like OpenAI's answer to all of that: ship faster, ship better, and give paying users a reason to stay.

The competition is real. Anthropic's Claude Opus 4.6 and Google's Gemini 3.1 lineup both launched in recent weeks, and both are performing well on independent leaderboards. OpenAI's monthly release cadence — from GPT-5.3 Codex to GPT-5.3 Instant to now GPT-5.4 in a matter of weeks — shows how seriously they're taking the pressure.

For business users: GPT-5.4 Thinking is probably the most practical ChatGPT upgrade in months. Less back-and-forth, fewer errors, and the new spreadsheet tools could genuinely save time. If you're on Plus or higher, it's already in your model picker.

For developers: the 1M context window, computer use, and Tool Search are the real headline. These are building blocks for serious AI agents, not chatbot tricks.