GPT-5.4
- Provider
- OpenAI
- Status
- available
- Context
- 1,100,000 tok
- SWE-bench
- 57.7%
- Price
- $2.5 / $15 /MTok
GPT-5.4 is the OpenAI model released on 5 March 2026 that, for about seven weeks, was the company’s flagship — until GPT-5.5 superseded it on 23 April. It remains one of the most-used models in the GPT-5 line because of its value: at $2.50/$15 per million tokens it is half the price of GPT-5.5, and on the standardized public SWE-bench Pro leaderboard it still leads on accuracy-per-dollar. GPT-5.4 was the release where OpenAI folded coding, reasoning and computer use into a single unified model rather than shipping separate specialist models.
Its headline achievement was computer use: GPT-5.4 scored 75% on OSWorld, the first model to cross the 72.4% human-expert baseline, and it led knowledge-work benchmarks (GDPval ~83%) at launch. As of mid-June 2026 it is superseded as flagship but actively recommended for cost- and latency-sensitive work, with mini and nano variants covering cheaper and edge use cases.
Quick specs
| Provider | OpenAI |
| Tier | Strong (former flagship) |
| Released | 5 March 2026 (mini/nano 17 March) |
| Status | Available (superseded by GPT-5.5) |
| API model ID | gpt-5.4 |
| Context window | ~1,100,000 tokens |
| Max output | 128,000 tokens |
| Input price | $2.50 / MTok |
| Output price | $15.00 / MTok |
| Reasoning control | reasoning_effort: none → xhigh |
| OSWorld (computer use) | 75% (first model past the 72.4% human baseline) |
| SWE-bench Pro | 57.7% (OpenAI); ~59% on the standardized leaderboard |
| Best for | Value, computer use, knowledge work, low-latency agentic tasks |
| Limitations | Superseded by GPT-5.5; vendor-run headline benchmarks |
What GPT-5.4 is
GPT-5.4 sits in the middle of the GPT-5 line — the successor to GPT-5.3 and the predecessor to GPT-5.5. Its defining design choice was unification: rather than offering separate models for chat, reasoning, coding and computer use, GPT-5.4 rolled them into one model with configurable reasoning_effort, so the same model handles a quick reply or a long agentic task depending on the effort you set.
Two things made it stand out at launch. First, computer use: its 75% on OSWorld was the first score to beat the human-expert baseline of 72.4%. Second, knowledge work: it led GDPval (~83%) among models available in March 2026. Both strengths carried forward into GPT-5.5, but GPT-5.4 remains the value tier of the family.
Model variants
GPT-5.4 shipped as a small family, with mini and nano following on 17 March 2026.
| Variant | API ID | What it is |
|---|---|---|
| GPT-5.4 | gpt-5.4 | The standard model; configurable reasoning; $2.50/$15 |
| GPT-5.4 Pro | gpt-5.4-pro | Parallel test-time compute for higher accuracy; $30/$180 |
| GPT-5.4 mini | gpt-5.4-mini | |
| GPT-5.4 nano | gpt-5.4-nano | Smallest variant, for edge and embedded use |
The mini variant is the value standout: it scores about 54% on SWE-bench Pro — within a few points of the standard model’s 57.7% — at roughly a sixth of the cost, which makes it a popular default for high-volume coding and agentic workloads.
Benchmark performance
OpenAI-reported figures from the 5 March 2026 launch; treat vendor numbers as a ceiling.
| Benchmark | GPT-5.4 | Notes |
|---|---|---|
| OSWorld (computer use) | 75% | First model past the 72.4% human-expert baseline |
| GDPval (knowledge work) | ~83% | Knowledge-work leader at release |
| SWE-bench Pro | 57.7% | OpenAI-reported; ~59% on the standardized public leaderboard (leads) |
GPT-5.4’s clearest strengths are computer use and knowledge work. On software engineering it is strong rather than class-leading — Claude leads the hardest coding benchmarks — but on the standardized public SWE-bench Pro scaffold GPT-5.4 leads on accuracy, and at its price it is the field’s value leader. See best AI for coding and best AI models for where it sits.
Pricing
| Input (per MTok) | Output (per MTok) | Cached input | |
|---|---|---|---|
| GPT-5.4 | $2.50 | $15.00 | $0.25 |
| GPT-5.4 mini | ~$0.40 | ~$1.60 | — |
| GPT-5.4 Pro | $30.00 | $180.00 | — |
| GPT-5.5 (successor) | $5.00 | $30.00 | $0.50 |
GPT-5.4 is half the per-token price of GPT-5.5, which is the main reason a vocal “5.4 holdout” contingent stuck with it after the upgrade — for many workloads the value is better even though GPT-5.5 is more capable and uses fewer tokens. The mini variant pushes the value further for high-volume use.
How to access GPT-5.4
GPT-5.4 is generally available with no waitlist:
- API —
gpt-5.4,gpt-5.4-mini,gpt-5.4-nano,gpt-5.4-provia the Responses and Chat Completions APIs. - Microsoft Foundry (Azure) and third-party routers such as OpenRouter.
- ChatGPT — available historically as the flagship; the default has since moved to GPT-5.5.
How GPT-5.4 compares
vs GPT-5.5
GPT-5.5 improves on GPT-5.4 across nearly every benchmark while using fewer output tokens — but at double the per-token price. For latency- or cost-sensitive workloads, GPT-5.4 (and its mini/nano variants) remains the pragmatic choice; for the best capability, GPT-5.5 is the upgrade.
vs Claude and Gemini
GPT-5.4 trails Claude Opus 4.8 on the hardest coding benchmarks, but its computer-use and knowledge-work scores are frontier-class, and its price undercuts the top Claude and Gemini models. For value and computer-use-heavy agentic work it is a strong pick; for the hardest software engineering, Claude leads.
Known limitations
Superseded as flagship. GPT-5.5 is the current frontier model; GPT-5.4 is now the value tier rather than the top of the line.
Coding isn’t class-leading. Claude leads the hardest coding benchmarks; GPT-5.4’s edge is value and computer use, not raw SWE-bench Verified.
Vendor-run headline benchmarks. The launch figures are OpenAI’s own; lead with standardized leaderboards where decisions ride on the number.
Version history
| Version | Released | Key points |
|---|---|---|
| GPT-5.4 | 5 Mar 2026 | Unified model; 75% OSWorld (first past human baseline); $2.50/$15; mini/nano 17 Mar |
| GPT-5.5 | 23 Apr 2026 | Successor flagship; $5/$30 |
| GPT-5.3 | Feb 2026 | Predecessor; GPT-5.3-Codex for agentic coding |
FAQ
What is GPT-5.4?
GPT-5.4 is an OpenAI model released on 5 March 2026 that unified coding, reasoning and computer use into one model. It was OpenAI’s flagship until GPT-5.5 arrived on 23 April 2026, and remains widely used as the value option in the GPT-5 line.
How much does GPT-5.4 cost?
$2.50 per million input tokens and $15 per million output — half the price of GPT-5.5. The mini variant is roughly six times cheaper (~$0.40/$1.60), and GPT-5.4 Pro is $30/$180.
Is GPT-5.4 still worth using over GPT-5.5?
For cost- and latency-sensitive workloads, often yes. GPT-5.5 is more capable, but GPT-5.4 is half the price and leads the standardized public SWE-bench Pro leaderboard on value. Many developers kept using it after the upgrade for exactly this reason.
What is GPT-5.4 best at?
Computer use (75% on OSWorld, the first model past the human-expert baseline) and knowledge work (GDPval ~83%), plus strong, well-priced agentic coding.
What’s the context window?
About 1.1 million tokens, with up to 128,000 output tokens.
Last verified 18 June 2026. GPT-5.4’s headline benchmark figures are OpenAI-reported from the 5 March 2026 launch; the standardized SWE-bench Pro placement comes from the public leaderboard. Pricing and standing change quickly — confirm against OpenAI’s official pages before relying on them.