GPT-5.4
GPT-5.4 is still a sound API choice for long documents and computer-use agents, but it is no longer OpenAI's recommended model. OpenAI released it on 5 March 2026 across ChatGPT, the API and Codex. It merged GPT-5.3-Codex's coding skills with general reasoning, added built-in computer use (the model can operate a desktop by clicking and typing) and offered a context window of up to 1.05 million tokens.
On OpenAI's figures it scored 75.0% on OSWorld-Verified, above the 72.4% average for humans, and 83.0% on GDPval. Smaller GPT-5.4 mini and nano followed on 17 March 2026. GPT-5.5 replaced it as the top ChatGPT model on 23 April 2026. The API price is $2.50 per million input tokens and $15 per million output tokens.
GPT-5.4 was the release where OpenAI's models started acting like office workers, and it remains a capable API model. It brought three big changes at once:
- Computer use built in. 75.0% on OSWorld-Verified, up from 47.3% for GPT-5.2 (OpenAI figures).
- A 1.05M-token context you can switch on for very long documents or codebases.
- Better professional work. 83.0% on GDPval and, OpenAI says, 33% fewer factual errors than GPT-5.2.
Reviewers agreed on the gains. ZDNET praised the drop in hallucinations but noted it sometimes ignored parts of a prompt. Vice liked the research quality and the lower token use.
Who should use it: API teams already tuned to GPT-5.4, especially for computer-use agents or very long inputs, where its behaviour is well understood.
Who should not: new projects. GPT-6 Sol costs less ($2 / $10 vs $2.50 / $15) and is OpenAI's current mid-tier model. Watch the long-context surcharge too: past 272K input tokens, GPT-5.4's price jumps to 2x input and 1.5x output for the whole session.
Score breakdown
Best for
- Existing computer-use agents built on GPT-5.4
- Very long document or codebase analysis via the 1.05M context
- Spreadsheet, slide and document automation where GPT-5.4 prompts are tuned
What we like
- Built-in computer use scored 75.0% on OSWorld-Verified, above the 72.4% human average (OpenAI)
- Opt-in 1.05M-token context for whole codebases or document sets
- Five reasoning-effort levels, from none to xhigh, to trade speed for depth
- 83.0% GDPval and 33% fewer factual errors than GPT-5.2 (OpenAI claims)
- mini and nano versions for cheaper, faster jobs
Watch out for
- Price doubles for input (and rises 1.5x for output) once prompts pass 272K tokens
- More expensive than the newer GPT-6 Sol
- Reviewers reported occasional prompt-following slips
- GPT-5.4 Pro is costly at $30 / $180 per million tokens
Specs at a glance
| Developer | OpenAI |
|---|---|
| Release date | 5 March 2026 (Thinking and Pro); 17 March 2026 (mini and nano) |
| Variants | GPT-5.4 Thinking, GPT-5.4 Pro, GPT-5.4 mini (free ChatGPT users), GPT-5.4 nano (API only) |
| API snapshot | gpt-5.4-2026-03-05 |
| Context window | 1,050,000 tokens (opt-in); 272,000 by default |
| Max output | 128,000 tokens |
| Knowledge cutoff | 31 August 2025 |
| Input / output | Text and image in; text out |
| Reasoning effort | none (default), low, medium, high, xhigh |
| API pricing | $2.50 input, $0.25 cached input, $15 output per 1M tokens |
| Long-prompt surcharge | Over 272K input tokens: 2x input and 1.5x output price for the session |
| GPT-5.4 Pro API pricing | $30 input, $180 output per 1M tokens |
| Batch / Flex / Priority | 50% / 50% / 200% of standard price |
| Predecessor / successor | GPT-5.2 and GPT-5.3-Codex / GPT-5.5 |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| GDPval (wins or ties vs experts) | 83.0% | OpenAI | GPT-5.2: 70.9% |
| OSWorld-Verified | 75.0% | OpenAI | GPT-5.2: 47.3%; human average 72.4% |
| SWE-bench Pro | 57.7% | OpenAI | GPT-5.3-Codex: 56.8% |
| BrowseComp | 82.7% | OpenAI | GPT-5.2: 65.8% |
| Toolathlon | 54.6% | OpenAI | GPT-5.2: 46.3% |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| API input | $2.50 per 1M tokens | 2x above 272K input tokens |
| API cached input | $0.25 per 1M tokens | |
| API output | $15 per 1M tokens | 1.5x above 272K input tokens |
| GPT-5.4 Pro (API) | $30 in / $180 out per 1M tokens | |
| Batch and Flex | 50% of standard | |
| Priority processing | 200% of standard |
GPT-5.4 vs GPT-5.3-Codex vs GPT-5.2
| Benchmark | GPT-5.2 | GPT-5.3-Codex | GPT-5.4 |
|---|---|---|---|
| GDPval | 70.9% | 70.9% | 83.0% |
| SWE-bench Pro | 55.6% | 56.8% | 57.7% |
| OSWorld-Verified | 47.3% | 74.0% | 75.0% |
| Toolathlon | 46.3% | 51.9% | 54.6% |
| BrowseComp | 65.8% | 77.3% | 82.7% |
All figures come from OpenAI's GPT-5.4 launch post. Note that OpenAI's earlier GPT-5.3-Codex post listed 64.7% on OSWorld-Verified, so test settings differ between posts.
How the 1M-token context works
GPT-5.4 runs with a 272,000-token window by default. To use the full 1,050,000 tokens you must switch it on in your API settings (in Codex this is done with the model_context_window and model_auto_compact_token_limit settings). Keep an eye on cost:
- Up to 272K input tokens: $2.50 in / $15 out per million.
- Above 272K: 2x input ($5) and 1.5x output ($22.50) for the whole session.
For most jobs, splitting documents into smaller chunks is cheaper than one giant prompt.
GPT-5.4 in ChatGPT
In ChatGPT, GPT-5.4 Thinking and GPT-5.4 Pro were for paid plans only. From 17 March 2026 free users got GPT-5.4 mini. GPT-5.2 Thinking stayed in the Legacy menu until 5 June 2026. GPT-5.5 Thinking and Pro took over on 23 April 2026, and today ChatGPT runs the GPT-6 family on paid plans. The API model has no deprecation notice as of 23 September 2026.
- Leave reasoning effort at
nonefor simple lookups and raise it only for hard tasks. Higher effort uses more output tokens, which cost $15 per million. - Before turning on the 1M context, check whether your prompt really needs it. Crossing 272K tokens raises the price for the whole session.
- Run large, non-urgent jobs through Batch or Flex to pay 50% of standard rates.
- Use GPT-5.4 mini or nano for classification and extraction, and keep full GPT-5.4 for multi-step agent work.
- When testing computer-use agents, add a human check before any irreversible click such as sending or buying. A 75% benchmark score still means one task in four can fail.
Jargon explained
- Computer use
- The model looks at screenshots and controls a mouse and keyboard to complete tasks in normal apps.
- Reasoning effort
- A setting that tells the model how long to think before answering. More effort means better answers on hard problems but slower, pricier replies.
- BrowseComp
- A test of how well an AI can dig up hard-to-find facts by browsing the web.
- Batch processing
- Sending many API requests to run within a day instead of instantly, in exchange for a lower price.
- Context window
- How much text a model can read at once. 1.05 million tokens is roughly 750,000 words.
Alternatives to consider
GPT-6 Sol
Current OpenAI mid-tier model and cheaper at $2 / $10 per million tokens.
GPT-5.5
Direct successor from April 2026.
GPT-6 Astra
OpenAI's flagship for the hardest reasoning and agent tasks.
Gemini
Google's models also offer very long context windows.
Claude
Strong rival for coding and computer-use agents.
Frequently asked questions
When was GPT-5.4 released?
5 March 2026 for GPT-5.4 Thinking and Pro in ChatGPT, the API and Codex. GPT-5.4 mini and nano followed on 17 March 2026.
How much does GPT-5.4 cost?
$2.50 per million input tokens, $0.25 cached and $15 per million output tokens. Prompts above 272K tokens cost 2x input and 1.5x output. GPT-5.4 Pro costs $30 / $180.
Does GPT-5.4 really have a 1 million token context?
Yes, 1,050,000 tokens, but you must switch it on. The default is 272,000 tokens, and anything above that costs more.
Is GPT-5.4 better than humans at using a computer?
On one test, by a small margin. OpenAI reports 75.0% on OSWorld-Verified against a 72.4% human average. That is a lab benchmark, not proof it can run any desktop job unsupervised.
Is GPT-5.4 still available?
Yes in the API, with no deprecation notice as of 23 September 2026. In ChatGPT it has been superseded by GPT-5.5 and then the GPT-6 family.
GPT-5.4 or GPT-6 Sol?
For new work, GPT-6 Sol. It is newer and cheaper ($2 / $10 vs $2.50 / $15). Stay on GPT-5.4 only if your prompts or agents are tuned to it and you have not tested a switch.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Introducing GPT-5.4 (OpenAI)
- GPT-5.4 model page (OpenAI API docs)
- GPT-5.4 (Wikipedia)
- GPT-5.4 deep dive: pricing, context limits, and tool search explained (OpenAI Developer Community)
- GPT-5.4 on OpenRouter (OpenRouter)
- Introducing GPT-5.3-Codex (OpenAI)
- GPT-5.5 (Wikipedia)