GPT-5.4
GPT-5.4 was OpenAI's main model from Mar 5, 2026, released in ChatGPT (as GPT-5.4 Thinking), on the API and in Codex. It was the company's first general-purpose model with native computer use, has a 1.05 million token window and costs US$ 2.50 per million input tokens and US$ 15 per million output. GPT-5.5 took its place on Apr 23, and Codex with ChatGPT sign-in retired it on Aug 31, 2026. It is still on the API with no shutdown date, but GPT-6.1 Sol costs less on both ends and has knowledge through April 2026.
GPT-5.4 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(snapshotgpt-5.4-2026-03-05)- Release
- Mar 5, 2026, in ChatGPT (as GPT-5.4 Thinking), on the API and in Codex
- Input
- US$ 2.50 per million tokens
- Output
- US$ 15 per million tokens
- Cache read
- US$ 0.25 per million (10% of input), no cache write charge
- Above 272 thousand input tokens
- input doubled and output 1.5x, for the whole session: US$ 5 and US$ 22.50
- Batch and Flex
- US$ 1.25 input and US$ 7.50 output
- Fast mode (formerly Priority)
- US$ 5 and US$ 30
- Context window
- 1,050,000 tokens
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Aug 31, 2025
- Reasoning
- effort
none(default),low,medium,highandxhigh - Where it runs
- OpenAI API (Chat Completions, Responses and Batch) and Amazon Bedrock, since Jun 1, 2026
- Starting limit (Tier 1)
- 500 requests and 500,000 tokens per minute
- Cyber risk
- High; the first general-purpose model with safeguards for that level
- Status
- replaced by GPT-5.5 on Apr 23, 2026; out of Codex with ChatGPT sign-in since Aug 31, 2026; no deprecation notice on the API
Where GPT-5.4 does well, where it falls behind and who it is for
Where it does well
- Computer use: 75.0% on OSWorld-Verified, above the 72.4% human performance OpenAI cites and the 47.3% for GPT-5.2.
- Office work: 83.0% on GDPval, the rate of wins or ties against professionals in 44 occupations, against 70.9% for GPT-5.2 and GPT-5.3-Codex.
- Spreadsheets: 87.3% on junior investment banking analyst modeling tasks, against 68.4% for GPT-5.2, on an OpenAI internal test.
- Many tools: with tool search, 250 MCP Atlas tasks with 36 MCP servers used 47% fewer tokens at the same accuracy, in OpenAI's tests.
- Fewer factual errors: according to OpenAI, claims 33% less likely to be false and responses 18% less likely to contain any error than GPT-5.2.
Where it falls behind
- Very long context: 36.6% on MRCR v2 between 512 thousand and 1 million tokens. In the April table, GPT-5.5 scored 74.0% on the same test, more than double.
- Terminal: 75.1% on Terminal-Bench 2.0, below the 77.3% for GPT-5.3-Codex, which costs less per token.
- Surcharge: above 272 thousand input tokens, the whole session moves to US$ 5 and US$ 22.50 per million.
- Older knowledge: it runs to Aug 31, 2025, against Apr 30, 2026 for GPT-6.1 Sol and GPT-6 Astra.
- A cheaper successor: GPT-6.1 Sol costs US$ 2 and US$ 10, 20% less on input and 33% less on output.
Who it makes sense for
- Production integrations already validated on
gpt-5.4: the snapshot stays fixed and there is no shutdown date on the API. - Apps tied to Chat Completions with function calling: the GPT-5.4 page lists both with no restriction, while on GPT-6.1 Sol tools only work through the Responses API.
- Low-latency flows without reasoning: the default effort is
none, and in that mode GPT-5.4 scored 64.3% on Tau2-bench Telecom, against 57.2% for GPT-5.2. - A staged migration: keep 5.4 while you measure GPT-6.1 Sol on the same tasks, since no shutdown date has been published.
When to pick another model
- New projects: GPT-6.1 Sol costs less per token and has more recent knowledge.
- Documents above 272 thousand tokens: the whole session gets more expensive and 5.4's long-context scores are low.
- Codex with ChatGPT sign-in: the model left on Aug 31, 2026, and OpenAI points to GPT-6 Sol instead.
- Simple volume work, such as classifying and extracting: GPT-6 Luna costs US$ 0.10 and US$ 0.50, 25 times less on input and 30 times less on output.
What OpenAI published at launch
GPT-5.4 brought the coding of GPT-5.3-Codex and the professional work of GPT-5.2 into a single model. OpenAI said it chose the name to mark that jump and to simplify model choice in Codex. Alongside it, OpenAI released tool search, built-in computer use through the computer tool and the 1 million token window on the API.
The table below is the one from the announcement, with OpenAI models only. GPT-5.4 leads 9 of the 11 rows and loses two: financial analysis, to GPT-5.2, and the terminal, to GPT-5.3-Codex. On OSWorld, 5.3-Codex shows 74.0%, measured with a new parameter that keeps the original image resolution; its own announcement had 64.7%.
| GPT-5.4 OpenAI | GPT-5.3-Codex OpenAI | GPT-5.2 OpenAI | |
|---|---|---|---|
| GDPval office work |
83.0% | 70.9% | 70.9% |
| FinanceAgent v1.1 financial analysis |
56.0% | 54.0% | 59.5% |
| Investment banking modeling spreadsheets, internal test |
87.3% | 79.3% | 68.4% |
| SWE-Bench Pro software engineering |
57.7% | 56.8% | 55.6% |
| Terminal-Bench 2.0 coding in the terminal |
75.1% | 77.3% | 62.2% |
| OSWorld-Verified computer use |
75.0% | 74.0% | 47.3% |
| BrowseComp web research |
82.7% | 77.3% | 65.8% |
| MCP Atlas tool use |
67.2% | not measured | 60.6% |
| GPQA Diamond PhD-level science |
92.8% | 92.6% | 92.4% |
| Humanity's Last Exam hard knowledge, with tools |
52.1% | not measured | 45.5% |
| ARC-AGI-2 abstract reasoning |
73.3% | not measured | 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.
GPT-5.4 against rivals, in the April table
The GPT-5.4 announcement did not include models from other companies. They appear in the GPT-5.5 table from Apr 23, which repeats the 5.4 scores next to Claude Opus 4.7, from Anthropic, and Gemini 3.1 Pro, from Google.
In the eight tests below, GPT-5.4 leads none. Against Opus 4.7 it is ahead on four (terminal, office work, web research and long context) and behind on four.
| GPT-5.4 OpenAI | GPT-5.5 OpenAI | Claude Opus 4.7 Anthropic | Gemini 3.1 Pro | |
|---|---|---|---|---|
| SWE-Bench Pro software engineering |
57.7% | 58.6% | 64.3% | 54.2% |
| Terminal-Bench 2.0 coding in the terminal |
75.1% | 82.7% | 69.4% | 68.5% |
| GDPval office work |
83.0% | 84.9% | 80.3% | 67.3% |
| OSWorld-Verified computer use |
75.0% | 78.7% | 78.0% | not measured |
| BrowseComp web research |
82.7% | 84.4% | 79.3% | 85.9% |
| Humanity's Last Exam hard knowledge, with tools |
52.1% | 52.2% | 54.7% | 51.4% |
| ARC-AGI-2 abstract reasoning |
73.3% | 85.0% | 75.8% | 77.1% |
| MRCR v2, 512 thousand to 1 million long context |
36.6% | 74.0% | 32.2% | not measured |
Table from the GPT-5.5 announcement (Apr 23, 2026), at xhigh effort for OpenAI models; competitor figures were published by OpenAI. The dot marks the best in each row. On SWE-Bench Pro, OpenAI notes that labs have seen evidence of memorization. "Not measured" means no published figure.
Is GPT-5.4 still worth using in 2026?
For a new project, no. Seven weeks after launch, OpenAI shipped GPT-5.5, at twice the price. The GPT-5.6 family came in July and the GPT-6 generation in September. Today OpenAI's model guide lists only three options: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna.
Price is what weighs most. GPT-6.1 Sol costs 20% less on input and 33% less on output than GPT-5.4, rereads the cache for US$ 0.10 and has knowledge through Apr 30, 2026. The caveat is cache writes, which cost 1.25 times input on GPT-6 models and are not charged on GPT-5.4.
On the API, gpt-5.4 has no entry on the deprecations page, and generally available models get at least six months of notice. In Codex with ChatGPT sign-in it left on Aug 31, 2026; with an API key it still works.
| GPT-5.4 Mar 2026 | GPT-6.1 Sol Sep 2026 | GPT-6 Astra Sep 2026 | GPT-6 Luna Sep 2026 | |
|---|---|---|---|---|
| Input | US$ 2.50 | US$ 2 | US$ 10 | US$ 0.10 |
| Output | US$ 15 | US$ 10 | US$ 50 | US$ 0.50 |
| Cache read | US$ 0.25 | US$ 0.10 | US$ 1 | US$ 0.01 |
| Cache write | not charged | US$ 2.50 | US$ 12.50 | US$ 0.125 |
| Above 272 thousand tokens input and output |
US$ 5 and US$ 22.50 | US$ 4 and US$ 15 | US$ 20 and US$ 75 | US$ 0.20 and US$ 0.75 |
| Knowledge cutoff | Aug 31, 2025 | Apr 30, 2026 | Apr 30, 2026 | May 18, 2026 |
| Accepts none effort | yes (default) | no | no | yes |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026.
What changes when replacing GPT-5.4 with GPT-6.1 Sol
In the GPT-5.5 guide, OpenAI already asked developers to treat the new model as a family to tune for, not a drop-in replacement for gpt-5.4. Moving to the GPT-6 generation, these are the points that touch the integration most:
- Identifier:
gpt-6.1-solinstead ofgpt-5.4. In Codex, the replacement OpenAI recommends isgpt-6-sol. - Effort: GPT-5.4 defaults to
none, which GPT-6.1 Sol does not accept. Uselowto keep latency down; its default ismedium. If you neednone, GPT-6 Sol and GPT-6 Luna accept it. - Tools: on GPT-6.1 Sol, tool calls only work through the Responses API. With reasoning on, remove
temperature,top_pandtop_logprobs. - Cache: replace
prompt_cache_retentionwithprompt_cache_options.ttlset to"30m". Writes now cost 1.25 times input, and reads drop to 0.05 times. - Knowledge: moves from Aug 31, 2025 to Apr 30, 2026. Rerun your tests on questions about recent facts.
Examples of using GPT-5.4
An agent reads the screen of an old system and fills in forms with mouse and keyboard. That is what OSWorld-Verified measures, and GPT-5.4 scored 75.0% at launch.
An analyst asks for a three-statement model with formulas and formatting. On OpenAI's internal investment banking test, GPT-5.4 scored 87.3%, against 68.4% for GPT-5.2.
A company connects CRM, email, spreadsheets and a database through MCP. With tool search, the model loads only the definition it is about to use; in OpenAI's test with 36 servers, token use fell 47%.
Where GPT-5.4 stands out and where it falls behind, by the numbers
GPT-5.4 has published results on 39 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.
See all 39 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 93.3% | #5 / 62 | #18 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 97.8% | #5 / 62 | #29 / 168 | 100.0% GPT-5.5 |
| Dtbench | 94.4% | #7 / 59 | #22 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 51.97 | #7 / 56 | #19 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 23.4% | #4 / 46 | #18 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 77.7% | #5 / 52 | #13 / 112 | 93.6% GPT-6 Astra |
| SciCode | 56.6% | #2 / 42 | #19 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 78.6% | #5 / 37 | #19 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 74.0% | #6 / 36 | #17 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 1,607 | #2 / 47 | #12 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,463 | #16 / 45 | #49 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 93.7% | #6 / 36 | #19 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 45.1% | #16 / 36 | #39 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 48.0% | #7 / 35 (tied with 2) | #17 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 37.0% | #2 / 35 | #17 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 56.0% | #2 / 31 | #24 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 49.0% | #5 / 27 | #19 / 67 | 100.0% GPT-6.1 Sol |
| FrontierMath | 47.6% | #4 / 40 | #4 / 65 | 51.7% GPT-5.5 |
| ForecastBench | 59.50 | #22 / 31 | #33 / 62 | 62.50 o3 |
| Vending Bench 2 | 6,144 | #6 / 34 | #20 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 52.4% | #2 / 13 | #24 / 50 | 82.4% Gemini 4 Argon |
| FrontierMath tier 4 | 27.1% | #4 / 39 | #4 / 48 | 39.6% GPT-5.5 Pro |
| Hle | 36.2% | #5 / 18 | #9 / 39 | 54.8% GPT-6 Astra |
| EnigmaEval | 16.0% | #7 / 18 | #9 / 36 | 39.3% Claude Fable 5 |
| Furniture Assembly | 37.5% | #4 / 9 | #19 / 34 | 83.3% Claude Opus 5.5 |
| Terminal-Bench | 81.8% | #2 / 32 | #2 / 33 | 84.7% GPT-5.5 |
| SWE-bench Verified | 76.9% | #7 / 25 | #8 / 32 | 83.5% Claude Opus 4.7 |
| Blueprint Bench 2 | 27.1% | #3 / 12 | #20 / 31 | 54.4% Gemini 4 Argon |
| METR: time horizon | 342 | #5 / 13 | #5 / 29 | 719 Claude Opus 4.6 |
| Gso | 31.4% | #5 / 14 | #10 / 28 | 88.2% Claude Fable 5.1 |
| Ebr Bench | 25.4% | #3 / 14 | #15 / 27 | 76.2% GPT-6 Astra |
| DeepSWE | 51.8% | #3 / 6 | #20 / 26 | 74.1% GPT-6 Astra |
| DeepResearch Bench | 35.1% | #18 / 18 | #23 / 24 | 55.3% Claude Opus 4.6 |
| Gbaeval | 45.1% | #4 / 13 | #10 / 22 | 79.6% Claude Opus 5 |
| Cl Bench | 27.9% | #1 / 18 | #1 / 19 | 27.9% GPT-5.4 |
| Algotune | 1.85 | #3 / 11 | #3 / 15 | 2.05 GPT-5.2 |
| Cl Bench Life | 21.7% | #2 / 13 | #2 / 13 | 22.2% GPT-5.5 |
| Posttrainbench | 19.0% | #5 / 5 | #11 / 11 | 41.8% Claude Fable 5 |
| MirrorCode | 15.6% | #2 / 4 | #7 / 9 | 77.4% Claude Opus 5.5 |
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 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 |
|---|---|---|---|---|---|
| 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 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| GPT-5.4 | OpenAI | 156.72 | 1,519 | 78.0 | Mar 5, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
| Claude Opus 4.7 | Anthropic | 156.32 | 1,532 | 76.5 | Apr 16, 2026 |
| Grok 4.6 | xAI | 156.31 | 1,475 | 78.0 | Aug 12, 2026 |
GPT-5.4 or Muse Spark 1.3?
Muse Spark 1.3 is the model from another lab closest to GPT-5.4 in capability. Across the 13 tests both took, this is the score:
| Test | GPT-5.4 | Muse Spark 1.3 |
|---|---|---|
| Game puzzles | 37.0% | 25.0% |
| Chess puzzles | 48.0% | 37.0% |
| WebDev Arena | 1,463 | 1,652 |
| CritPt | 23.4% | 26.0% |
| Apex Agents | 52.4% | 57.8% |
| Frontiermath Tier 4 V2 | 49.0% | 46.3% |
| Frontiermath Tiers 1 3 V2 | 78.6% | 74.4% |
| SciCode | 56.6% | 59.7% |
| ALE-Bench | 1,607 | 1,688 |
| Lmca | 51.97 | 53.87 |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What GPT-5.4 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
ARC-AGI 2
The score scale does not start at zero.
ARC-AGI
The score scale does not start at zero.
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? 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
How much does GPT-5.4 cost?
US$ 2.50 per million input tokens and US$ 15 per million output. Cache reads cost US$ 0.25, with no cache write charge. Above 272 thousand input tokens, the whole session moves to US$ 5 and US$ 22.50. Batch and Flex are half price, and Fast mode costs double. The price has not changed since launch.
Is GPT-5.4 still available?
On the API, yes: there is no shutdown date and no entry on the deprecations page. In Codex with ChatGPT sign-in it left on Aug 31, 2026; with an API key it still works. For ChatGPT, no official page we checked states the current status of GPT-5.4 Thinking.
Which model replaced GPT-5.4?
GPT-5.5, released on Apr 23, 2026 at US$ 5 and US$ 30. The GPT-5.6 family and the GPT-6 generation followed. In Codex, OpenAI points to GPT-6 Sol in place of gpt-5.4; on the API, the current guide lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna.
GPT-5.4 or GPT-5.3-Codex for coding?
In OpenAI's March figures, GPT-5.4 scored 57.7% on SWE-Bench Pro, against 56.8% for 5.3-Codex, and 75.1% on Terminal-Bench 2.0, against 77.3%. 5.4 has a larger window and computer use; 5.3-Codex costs less per token.
Is GPT-5.4 better than Claude Opus 4.7?
It depends on the test. In OpenAI's April table, GPT-5.4 is ahead of Opus 4.7 in the terminal (75.1% against 69.4%), office work and web research, and behind on software engineering (57.7% against 64.3%), computer use and abstract reasoning. The independent scores are at the top of this page.
Where does GPT-5.4 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-5.4 ranks #23 out of 267 models, with an ECI of 156.72. 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 better than Muse Spark 1.3?
Across the 13 tests both models took and Epoch AI compiled, GPT-5.4 comes out ahead on 4, behind on 9 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 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 are: Cl Bench (#1 of 18); ALE-Bench (#2 of 47); SciCode (#2 of 42). 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
- OpenAI, GPT-5.4 model page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-5.4 Thinking system card
- OpenAI, GPT-5.5 announcement (table with GPT-5.4 and competitors)
- OpenAI, latest model guide (migration to GPT-6)
- OpenAI, prompt caching guide
- OpenAI, GPT-6.1 Sol model page (documentation)
- OpenAI, Codex models (GPT-5.4 retirement)
- OpenAI, API deprecations
- OpenAI, API changelog
- OpenAI, models on Amazon Bedrock (documentation)
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