GPT-5.4 Pro
GPT-5.4 Pro is the version of GPT-5.4 that uses more compute to think harder on each answer, released by OpenAI on Mar 5, 2026 in ChatGPT and on the API. It costs US$ 30 per million input tokens and US$ 180 per million output, 12 times GPT-5.4, and was replaced on Apr 23 by GPT-5.5 Pro, which costs the same and scores higher on 7 of 8 comparable tests. It is still available on the API with no shutdown date; today it only makes sense where it is already integrated or where you need agent tools its successor does not list.
GPT-5.4 Pro in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- OpenAI
- API identifier
gpt-5.4-pro(snapshotgpt-5.4-pro-2026-03-05)- Release
- Mar 5, 2026, in ChatGPT (Pro and Enterprise plans) and on the API
- Input
- US$ 30 per million tokens
- Output
- US$ 180 per million tokens
- Cache
- no cached input discount
- Above 272 thousand input tokens
- input doubled and output 1.5x, for the whole session: US$ 60 and US$ 270
- Batch and Flex
- US$ 15 and US$ 90 in the pricing table, although the model page says the Batch API is not supported
- Fast mode
- not offered
- Context window
- 1,050,000 tokens
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Aug 31, 2025
- Reasoning
- effort
medium(default),highandxhigh - Where it runs
- OpenAI Responses API only; some requests take minutes
- Starting limit (Tier 1)
- 50 requests and 50,000 tokens per minute
- Cyber risk
- OpenAI treated GPT-5.4 as High in cybersecurity
- Status
- replaced by GPT-5.5 Pro on Apr 23, 2026; no deprecation notice on the API
Where GPT-5.4 Pro does well, where it falls behind and who it is for
Where it does well
- Web research: 89.3% on BrowseComp, which OpenAI called a new state of the art at launch, against 82.7% for GPT-5.4 and 77.9% for GPT-5.2 Pro.
- Hard knowledge with tools: 58.7% on Humanity's Last Exam with tools. In the April table it came out ahead of even GPT-5.5 Pro (57.2%) and GPT-5.5 (52.2%).
- Abstract reasoning: 83.3% on ARC-AGI-2, against 73.3% for GPT-5.4 and 54.2% for GPT-5.2 Pro.
- Research math: 38.0% on FrontierMath Tier 4, against 27.1% for GPT-5.4 and 31.3% for GPT-5.2 Pro.
- Agent tools on the API: it lists computer use, tool search and
apply_patch, which GPT-5.5 Pro does not have.
Where it falls behind
- A successor at the same price: GPT-5.5 Pro costs the same US$ 30 and US$ 180 and scores higher on BrowseComp, FrontierMath, GeneBench and investment banking modeling.
- Sometimes below regular GPT-5.4: 83.6% against 87.3% on investment banking modeling and 82.0% against 83.0% on GDPval, while costing 12 times more per token.
- Older knowledge: it runs to Aug 31, 2025, against Dec 1, 2025 for GPT-5.5 Pro and Apr 30, 2026 for GPT-6 Astra.
- Slow and expensive to repeat: some requests take several minutes, the Tier 1 limit is 50 requests per minute and there is no cache discount.
Who it makes sense for
- Older integrations on
gpt-5.4-prothat rely on computer use, tool search orapply_patchwith Pro: GPT-5.5 Pro does not list those tools. - Comparisons with analyses run between March and April 2026: the
gpt-5.4-pro-2026-03-05snapshot stays fixed. - Hard knowledge questions with search: in OpenAI's April table it posted the top score on Humanity's Last Exam with tools, above GPT-5.5 Pro.
- A staged migration: keep 5.4 Pro while you measure GPT-5.5 Pro and GPT-6 pro mode on the same questions, since no shutdown date has been published.
When to pick another model
- New projects: GPT-5.5 Pro costs the same, has more recent knowledge and higher scores on most tests.
- Volume, live chat and customer support: requests that take minutes and a low Tier 1 limit.
- Long instructions repeated on every request: with no cache discount, every call pays US$ 30 per million input tokens.
- Anyone who wants more reasoning for less: in GPT-6.1 Sol pro mode, tokens cost US$ 2 and US$ 10, although the mode uses more tokens than standard.
What "Pro" means in GPT-5.4 Pro
In its documentation, OpenAI describes GPT-5.4 Pro as the version of GPT-5.4 that uses more compute to think harder and give more consistent answers, meant for tough problems that may take longer and need deeper reasoning. The company did not detail the mechanism for this version. For GPT-5.5 Pro, the safety report explains that Pro is the same model using parallel compute at answer time.
The price shows the size of the gap: US$ 30 and US$ 180 per million tokens, 12 times GPT-5.4 on both input and output, with no cache discount. Pro runs only on the Responses API, defaults to medium effort and, according to OpenAI, some requests take several minutes.
In ChatGPT, it reached the Pro and Enterprise plans on Mar 5, 2026. OpenAI did not publish a separate safety report for Pro; the GPT-5.4 Thinking report treats the model as High in cybersecurity and in biology.
What OpenAI published at launch
The GPT-5.4 announcement compared Pro only with OpenAI's own models. It leads ten of the twelve rows below and loses two to regular GPT-5.4, on office work and investment banking modeling. There are no Pro figures for coding, computer use or long context.
| GPT-5.4 Pro OpenAI | GPT-5.4 OpenAI | GPT-5.2 Pro OpenAI | GPT-5.2 OpenAI | |
|---|---|---|---|---|
| GDPval office work |
82.0% | 83.0% | 74.1% | 70.9% |
| FinanceAgent v1.1 financial analysis |
61.5% | 56.0% | not measured | 59.5% |
| Investment banking modeling spreadsheets, internal test |
83.6% | 87.3% | 71.7% | 68.4% |
| BrowseComp web research |
89.3% | 82.7% | 77.9% | 65.8% |
| Frontier Science Research scientific research |
36.7% | 33.0% | not measured | 25.2% |
| FrontierMath Tier 1 to 3 advanced math |
50.0% | 47.6% | not measured | 40.7% |
| FrontierMath Tier 4 research math |
38.0% | 27.1% | 31.3% | 18.8% |
| GPQA Diamond PhD-level science |
94.4% | 92.8% | 93.2% | 92.4% |
| Humanity's Last Exam hard knowledge, no tools |
42.7% | 39.8% | 36.6% | 34.5% |
| Humanity's Last Exam hard knowledge, with tools |
58.7% | 52.1% | 50.0% | 45.5% |
| ARC-AGI-1 abstract reasoning |
94.5% | 93.7% | 90.5% | 86.2% |
| ARC-AGI-2 abstract reasoning, harder version |
83.3% | 73.3% | 54.2% | 52.9% |
Figures from the GPT-5.4 announcement (Mar 5, 2026), at xhigh effort unless noted. The dot marks the best in each row. "Not measured" means OpenAI published no figure for that model. On ARC-AGI-2, the GPT-5.2 Pro score was measured at high effort.
Is GPT-5.4 Pro still worth using in 2026?
Seven weeks after launch, OpenAI replaced it with GPT-5.5 Pro at the same price. In the GPT-5.5 table, the successor scores higher on seven of the eight tests where both appear; the exception is Humanity's Last Exam with tools, where 5.4 Pro leads by 1.5 points.
On the API, gpt-5.4-pro is still in the model index and has no entry on the deprecations page. In Codex, OpenAI retired gpt-5.4 and gpt-5.4-mini on Aug 31, 2026, without mentioning Pro. For ChatGPT, no official page we checked says whether it still appears.
Since GPT-5.6, Pro is also a mode of the newer models, billed at each model's normal rates. If you are starting today, test GPT-5.5 Pro or GPT-6 pro mode.
| GPT-5.4 Pro Mar 2026 | GPT-5.5 Pro Apr 2026 | GPT-5.5 Apr 2026 | |
|---|---|---|---|
| GDPval office work |
82.0% | 82.3% | 84.9% |
| Investment banking modeling internal test |
83.6% | 88.6% | 88.5% |
| BrowseComp web research |
89.3% | 90.1% | 84.4% |
| GeneBench genetics data analysis |
25.6% | 33.2% | 25.0% |
| FrontierMath Tier 1 to 3 advanced math |
50.0% | 52.4% | 51.7% |
| FrontierMath Tier 4 research math |
38.0% | 39.6% | 35.4% |
| Humanity's Last Exam no tools |
42.7% | 43.1% | 41.4% |
| Humanity's Last Exam with tools |
58.7% | 57.2% | 52.2% |
| Price per million tokens input and output |
US$ 30 and US$ 180 | US$ 30 and US$ 180 | US$ 5 and US$ 30 |
| Knowledge cutoff | Aug 31, 2025 | Dec 1, 2025 | Dec 1, 2025 |
Scores from the GPT-5.5 announcement table (Apr 23, 2026), at xhigh effort; prices from OpenAI's pricing page, checked on Oct 1, 2026. The dot marks the best in each score row.
What changes when replacing GPT-5.4 Pro
OpenAI says existing Pro model IDs keep their behavior and pricing, so there is no rush. Anyone who decides to switch has two routes, and each changes different things:
- To GPT-5.5 Pro: change the identifier to
gpt-5.5-pro. The price is the same and the default effort rises frommediumtohigh. Code interpreter and hosted shell come in; computer use, tool search andapply_patchgo out. - To GPT-6 pro mode: call
gpt-6-astraorgpt-6.1-solwithreasoning.modeset to"pro". Tokens cost US$ 10 and US$ 50 or US$ 2 and US$ 10, with a cache discount, but each answer uses more tokens. - On GPT-6 models, caching uses
prompt_cache_options.ttlset to"30m", and writes cost 1.25 times input. - Knowledge moves from August 2025 to December 2025 on GPT-5.5 Pro and to April 2026 on GPT-6 models.
Examples of using GPT-5.4 Pro
An analyst asks the model to find and cross-check data scattered across many sources. BrowseComp measures this kind of persistent search, and GPT-5.4 Pro scored 89.3% at launch. Today GPT-5.5 Pro does the same kind of task at the same price, with 90.1%.
An expert-level question in medicine, law or engineering needs long reasoning and source lookups. On Humanity's Last Exam with tools, 5.4 Pro scored 58.7%, the top score in OpenAI's April table.
An older flow uses 5.4 Pro with the computer use tool to check screens before acting. Since GPT-5.5 Pro does not list that tool, the switch goes to GPT-5.5 or a GPT-6 model.
Where GPT-5.4 Pro stands out and where it falls behind, by the numbers
GPT-5.4 Pro has published results on 13 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Jun 8, 2025 and Jun 3, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
It is not in the bottom half of its generation on any test with 6 or more models compared.
See all 13 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 94.6% | #1 / 62 | #7 / 173 | 96.4% Gemini 4 Argon |
| CritPt | 30.0% | #2 / 46 | #9 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 57.4% | #19 / 52 | #35 / 112 | 93.6% GPT-6 Astra |
| Frontiermath Tiers 1 3 V2 | 82.5% | #3 / 37 | #14 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 83.3% | #3 / 36 | #14 / 83 | 95.0% GPT-6 Astra |
| ARC-AGI | 94.5% | #4 / 36 | #15 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 46.3% | #14 / 36 | #37 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 59.0% | #4 / 35 | #9 / 80 | 76.0% GPT-6 Astra |
| SimpleBench | 74.1% | #5 / 33 | #9 / 74 | 81.9% Claude Fable 5 |
| Frontiermath Tier 4 V2 | 58.5% | #3 / 27 | #16 / 67 | 100.0% GPT-6.1 Sol |
| FrontierMath | 50.0% | #3 / 40 | #3 / 65 | 51.7% GPT-5.5 |
| Hle | 44.3% | #2 / 18 | #6 / 39 | 54.8% GPT-6 Astra |
| EnigmaEval | 23.8% | #3 / 18 | #5 / 36 | 39.3% Claude Fable 5 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare GPT-5.4 Pro with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| GPT-5.4 Pro | OpenAI | 159.00 | n/a | n/a | Mar 5, 2026 |
| Claude Haiku 5.5 | Anthropic | 158.98 | n/a | 72.1 | Oct 7, 2026 |
| Claude Opus 4.8 | Anthropic | 158.26 | 1,507 | 76.2 | May 28, 2026 |
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Gemini 3.8 Flash | Google DeepMind | 157.20 | 1,551 | 75.8 | Sep 2, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
GPT-5.4 Pro or Claude Haiku 5.5?
Claude Haiku 5.5 is the model from another lab closest to GPT-5.4 Pro in capability. Across the 7 tests both took, this is the score:
| Test | GPT-5.4 Pro | Claude Haiku 5.5 |
|---|---|---|
| SimpleQA Verified | 46.3% | 23.8% |
| Hle | 44.3% | 30.1% |
| CritPt | 30.0% | 22.6% |
| Frontiermath Tier 4 V2 | 58.5% | 46.3% |
| Frontiermath Tiers 1 3 V2 | 82.5% | 75.1% |
| Chess puzzles | 59.0% | 55.0% |
| GPQA Diamond | 94.6% | 89.6% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
Other models from OpenAI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| GPT-5.5 Instant | May 5, 2026 | 142.45 | 1,456 | n/a | API access |
| GPT-5.5 Pro | Apr 23, 2026 | 161.97 | n/a | n/a | API access |
| GPT-5.5 | Apr 23, 2026 | 158.95 | 1,533 | 80.2 | API access |
| GPT-5.4 Mini | Mar 17, 2026 | 148.77 | 1,446 | 66.4 | API access |
| GPT-5.4 Nano | Mar 17, 2026 | 146.12 | 1,404 | 69.6 | API access |
| GPT-5.4 Pro | Mar 5, 2026 | 159.00 | n/a | n/a | API access |
| GPT-5.4 | Mar 5, 2026 | 156.72 | 1,519 | 78.0 | API access |
| GPT-5.3 Codex | Feb 5, 2026 | 156.76 | n/a | n/a | API access |
| GPT-5.2 Pro | Dec 11, 2025 | 156.03 | n/a | n/a | API access |
Have you used GPT-5.4 Pro? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about GPT-5.4 Pro
How much does GPT-5.4 Pro cost?
US$ 30 per million input tokens and US$ 180 per million output, with no cache discount. Above 272 thousand input tokens, the whole session moves to US$ 60 and US$ 270. The price has not changed since launch.
What is the difference between GPT-5.4 and GPT-5.4 Pro?
Pro uses more compute to think harder on each answer, according to OpenAI. It costs 12 times more per token, runs only on the Responses API and takes longer. In the launch figures it beats GPT-5.4 on web research, math and hard knowledge, but loses on GDPval and investment banking modeling.
Has GPT-5.4 Pro been replaced?
Yes. On Apr 23, 2026, OpenAI released GPT-5.5 Pro at the same price, with higher scores on seven of the eight tests where both appear. Since GPT-5.6, Pro also exists as a mode of the newer models, billed at each model's normal rates.
Is GPT-5.4 Pro being shut down?
On the API there is no published date and no entry on the deprecations page. In Codex, OpenAI retired GPT-5.4 and GPT-5.4 mini on Aug 31, 2026, without mentioning Pro. For ChatGPT, no official page we checked states its current status.
Is GPT-5.4 Pro better than GPT-5.5?
It depends on the test. In OpenAI's April table, 5.4 Pro is ahead of regular GPT-5.5 on web research (89.3% against 84.4%), FrontierMath Tier 4 and Humanity's Last Exam, and behind on office work (82.0% against 84.9%), investment banking modeling and ARC-AGI-2. GPT-5.5 costs US$ 5 and US$ 30, one sixth of the price. The independent scores are at the top of this page.
Where does GPT-5.4 Pro rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-5.4 Pro ranks #13 out of 267 models, with an ECI of 159.00. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is GPT-5.4 Pro better than Claude Haiku 5.5?
Across the 7 tests both models took and Epoch AI compiled, GPT-5.4 Pro comes out ahead on 7, behind on 0 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is GPT-5.4 Pro best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of GPT-5.4 Pro are: GPQA Diamond (#1 of 62); CritPt (#2 of 46); FrontierMath (#3 of 40). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- OpenAI, GPT-5.4 announcement (includes GPT-5.4 Pro)
- OpenAI, GPT-5.4 Pro model page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-5.4 Thinking system card
- OpenAI, GPT-5.4 usage guide
- OpenAI, GPT-5.5 announcement (table including GPT-5.4 Pro)
- OpenAI, GPT-5.5 Pro model page (documentation)
- OpenAI, reasoning guide (pro mode)
- OpenAI, Codex models (GPT-5.4 retirement)
- OpenAI, API deprecations
- OpenAI, ChatGPT release notes (help center)
- OpenAI, API changelog
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking