What is GPT 5.4?
GPT 5.4 is OpenAI's March 5, 2026 model and the first general-purpose OpenAI model with computer use built in. It folds reasoning, coding, and agent work into one model, replaces GPT-5.2 Thinking, and runs on Unifically as openai/gpt-5.4. It accepts text and image input, outputs text, and takes up to 1M tokens of context with a 128k-token max output. The standout skill is operating software: it reads screens, clicks, types, and carries a plan across long tool chains, scoring 75.0% on OSWorld-Verified, above the 72.4% human baseline.
Key features of GPT 5.4
Native computer use
It scores 75.0% on OSWorld-Verified, past the 72.4% human baseline and far ahead of GPT-5.2 at 47.3%. The model reads screenshots and issues mouse and keyboard actions itself, with no separate computer-use model.
Up to 1M tokens of context
Large codebases, long agent sessions, and multi-document research fit in one run, with a 128k-token max output for long answers and big diffs.
Tool search for big tool catalogs
The model pulls in tool definitions on demand instead of loading every definition upfront. Agents with hundreds of tools spend far fewer tokens per turn. On BrowseComp it scores 82.7%, up from 65.8% for GPT-5.2.
Stronger coding
It scores 57.7% on SWE-bench Pro, ahead of GPT-5.3-Codex at 56.8%, and 75.1% on Terminal-Bench 2.0, up from 62.2% for GPT-5.2.
Abstract reasoning gains
It reaches 93.7% on ARC-AGI-1 and 73.3% on ARC-AGI-2, up from 52.9% for GPT-5.2.
Fewer factual errors, fewer tokens
Individual claims are 33% less likely to be false than GPT-5.2, and full responses are 18% less likely to contain any error. On GDPval knowledge work it scores 83.0% versus 70.9% for GPT-5.2.
Best for
Computer-use agents
Browser and desktop workflows that read screens, click, and type, with clear action limits.
Agentic coding
Multi-step fixes across a large repo with terminal and tool access.
Deep web research
Long research runs that search, cross-check, and compile many sources.
Long-context work
Contracts, filings, and codebases held together in one context.
Knowledge work
Reports, spreadsheets, and financial models, with human review on the output.
Agents with many tools
Tool search keeps token spend flat as the tool catalog grows.
Use cases
Build an operations agent that runs a whole workflow inside internal web tools: open the app, fill the form, check the result, and report back. Use it for coding agents that trace a bug across a repo, run tests in a terminal, and open a fix. It also fits research assistants that browse, verify claims across sources, and write up findings with fewer false statements. For document-heavy teams, GPT 5.4 holds contracts, filings, or a full codebase in one context and answers questions that span all of it.
Limitations
GPT-5.3-Codex still leads on Terminal-Bench 2.0 at 77.3% versus 75.1%. For pure terminal-driven coding, the specialist model remains slightly ahead.
SWE-bench Pro sits at 57.7%. That leads the line, but it also means hard repo tasks fail almost half the time. Review agent output before merging.
OSWorld is a controlled benchmark. Real apps change layouts, throw popups, and time out, so give computer-use agents narrow permissions and clear stop rules.
Blind preference boards place it mid-pack: Arena's Text board ranks gpt-5.4-high at #12 with a 1478±4 score, behind Claude Opus 4.6 at #4 with 1499±4.
GPT 5.4 vs GPT-5.2
GPT 5.4 is a wide step up from the GPT-5.2 Thinking model it replaces. OSWorld-Verified computer use jumps from 47.3% to 75.0%. BrowseComp web research rises from 65.8% to 82.7%. ARC-AGI-2 abstract reasoning climbs from 52.9% to 73.3%, and GDPval knowledge work moves from 70.9% to 83.0%. On top of the scores, its claims are 33% less likely to be false and it spends fewer reasoning tokens per problem, so long agent runs cost less and drift less.
When to use GPT 5.4
Use GPT 5.4 when the task is agentic: operating a browser or desktop, running long tool chains, or fixing code across a big repo. Pick it over GPT-5.2 in every case. Drop to GPT 5.4 Mini or Nano when the job is high-volume drafts, routing, or quick replies where cost per call matters more than depth.
API examples
Call GPT 5.4 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/openai/gpt-5.4.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "openai/gpt-5.4",
"messages": [
{ "role": "user", "content": "Plan the steps to reproduce and fix this flaky integration test." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
openai/gpt-5.4, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
Text and image input with text output. It reads screenshots, charts, and documents in the same request as code or prose.
Up to 1M tokens in the API, with a max output of 128k tokens. That fits large codebases, long agent sessions, or many documents in one run.
Yes. Computer use is built into the model itself. It scores 75.0% on OSWorld-Verified, above the 72.4% human baseline, by reading screens and issuing mouse and keyboard actions.
Yes. It scores 57.7% on SWE-bench Pro, ahead of GPT-5.3-Codex at 56.8%, and 75.1% on Terminal-Bench 2.0, a big jump from GPT-5.2 at 62.2%.
A way to handle large tool catalogs. The model loads tool definitions on demand instead of holding every definition in context, which cuts token overhead for agents with many tools.
Yes. Its individual claims are 33% less likely to be false, and full responses are 18% less likely to contain any error. It also solves problems with fewer reasoning tokens.
Set clear action limits for agents that operate real accounts or desktops. Hard repo tasks still fail often at 57.7% SWE-bench Pro, so review agent output before merging.