GPT-5.3 Codex
GPT-5.3-Codex is OpenAI's model built for agentic coding inside Codex, released on Feb 5, 2026 and opened on the API on Feb 24. It costs US$ 1.75 per million input tokens and US$ 14 per million output, works with 400 thousand tokens of context and was the first model OpenAI treated as High risk in cybersecurity. A month later, GPT-5.4 brought its coding into a general-purpose model, and since May 26, 2026 it no longer appears as a choice in Codex with ChatGPT sign-in. It is still active on the API, with no shutdown date.
GPT-5.3 Codex 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-5.3-codex(single snapshot, same name)- Release
- Feb 5, 2026 in Codex, for paid ChatGPT plans; Feb 24, 2026 on the API
- Input
- US$ 1.75 per million tokens
- Output
- US$ 14 per million tokens
- Cache read
- US$ 0.175 per million (10% of input)
- Fast mode
- US$ 3.50 and US$ 28, double the normal price
- Batch API
- not supported
- Context window
- 400 thousand tokens, with up to 272 thousand of input
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Aug 31, 2025
- Reasoning
- effort
low,medium,highandxhigh; nonone - Where it runs
- Responses API only; tools: functions, web search, hosted shell and skills
- Starting limit (Tier 1)
- 500 requests and 500,000 tokens per minute
- Cyber risk
- High, the first OpenAI model with that classification
- Status
- off the model picker in Codex with ChatGPT sign-in since May 26, 2026; no deprecation notice on the API
Where GPT-5.3 Codex does well, where it falls behind and who it is for
Where it does well
- Terminal: 77.3% on Terminal-Bench 2.0, against 64.0% for GPT-5.2-Codex and 62.2% for GPT-5.2. In the March table it came out ahead of even GPT-5.4 (75.1%).
- Computer use: 64.7% on OSWorld-Verified at launch, against 38.2% for GPT-5.2-Codex. Later, with a parameter that keeps the original image resolution, OpenAI published 74.0%.
- Freelance work: 81.4% on SWE-Lancer IC Diamond, against 76.0% for GPT-5.2-Codex and 74.6% for GPT-5.2.
- Price: US$ 1.75 and US$ 14, the same as GPT-5.2-Codex and below the US$ 2.50 and US$ 15 of GPT-5.4.
- Talking while it works: you can ask questions and change direction while the agent works, without losing context, according to OpenAI.
Where it falls behind
- One door only: it runs only on the Responses API, with no Chat Completions and no Batch API.
- Smaller window: 400 thousand tokens, with at most 272 thousand of input, against 1.05 million for GPT-5.4 and the GPT-6 models.
- Few tools on the API: functions, web search, hosted shell and skills. It does not list computer use, file search, code interpreter or
apply_patch. - Not for conversation: the safety report says the model is not intended for conversational use.
- Small gain on SWE-Bench Pro: 56.8%, only 0.4 points above GPT-5.2-Codex (56.4%).
Who it makes sense for
- Coding agents that run on the API with your own key and were already validated on
gpt-5.3-codex: the May retirement applied only to Codex with ChatGPT sign-in. - Long terminal tasks where you measured an edge over GPT-5.4: in OpenAI's figures it scores 77.3% against 75.1%.
- Comparisons with evaluations run between February and March 2026, using the model of that period.
- A staged migration: keep 5.3-Codex while you test GPT-6.1 Sol on the same repositories. OpenAI's policy gives at least three months of notice before retiring Codex variants.
When to pick another model
- A new coding agent project: for complex coding and agentic workflows, OpenAI now recommends GPT-6.1 Sol in Codex.
- Chat, customer support or end-user text: the model was not built for conversation.
- Repositories or documentation that exceed 272 thousand input tokens.
- Cheap overnight batches: there is no Batch API. On GPT-6 Luna, batch runs at US$ 0.05 and US$ 0.25 per million.
What OpenAI published at launch
According to OpenAI, GPT-5.3-Codex combined the coding of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2, and ran 25% faster in Codex. The company says it was the first model to help create itself: early versions debugged its own training and helped with the launch.
The announcement compared the model only with in-house predecessors. It leads five of the six rows and ties GPT-5.2 on GDPval. The biggest jumps are in computer use and the terminal; on software engineering, the gain over GPT-5.2-Codex was 0.4 points.
| GPT-5.3-Codex OpenAI | GPT-5.2-Codex OpenAI | GPT-5.2 OpenAI | |
|---|---|---|---|
| SWE-Bench Pro software engineering |
56.8% | 56.4% | 55.6% |
| Terminal-Bench 2.0 coding in the terminal |
77.3% | 64.0% | 62.2% |
| OSWorld-Verified computer use |
64.7% | 38.2% | 37.9% |
| GDPval office work |
70.9% | not measured | 70.9% |
| Capture the Flag cybersecurity |
77.6% | 67.4% | 67.7% |
| SWE-Lancer IC Diamond freelance software work |
81.4% | 76.0% | 74.6% |
Appendix of the GPT-5.3-Codex announcement (Feb 5, 2026), at xhigh effort, except GPT-5.2 on GDPval (high). The dot marks the best in each row; GDPval is a tie. "Not measured" means no published figure.
After GPT-5.4, which merged the two lines
On Mar 5, OpenAI released GPT-5.4 as the first mainline model with the coding of GPT-5.3-Codex built in, and said it chose the name to simplify model choice in Codex. In the current pricing table, 5.3-Codex is the only model in the Codex category.
In the March comparison, GPT-5.4 beats 5.3-Codex on six of eight tests, including office work (83.0% against 70.9%) and web research. 5.3-Codex stays ahead in the terminal, by 2.2 points, and costs less per token.
| GPT-5.3-Codex Feb 2026 | GPT-5.4 Mar 2026 | GPT-5.2 Dec 2025 | |
|---|---|---|---|
| GDPval office work |
70.9% | 83.0% | 70.9% |
| FinanceAgent v1.1 financial analysis |
54.0% | 56.0% | 59.5% |
| Investment banking modeling spreadsheets, internal test |
79.3% | 87.3% | 68.4% |
| SWE-Bench Pro software engineering |
56.8% | 57.7% | 55.6% |
| Terminal-Bench 2.0 coding in the terminal |
77.3% | 75.1% | 62.2% |
| OSWorld-Verified computer use |
74.0% | 75.0% | 47.3% |
| BrowseComp web research |
77.3% | 82.7% | 65.8% |
| GPQA Diamond PhD-level science |
92.6% | 92.8% | 92.4% |
| Price per million tokens input and output |
US$ 1.75 and US$ 14 | US$ 2.50 and US$ 15 | US$ 1.75 and US$ 14 |
| Context window | 400 thousand | 1.05 million | 400 thousand |
Scores from the GPT-5.4 announcement (Mar 5, 2026), at xhigh effort; prices and windows from OpenAI's documentation, checked on Oct 1, 2026. The dot marks the best in each score row. Here the 5.3-Codex OSWorld score is the newer measurement, with the original image resolution.
Is GPT-5.3-Codex still worth using in 2026?
For a new coding agent, no. Since May 26, 2026 the model no longer appears as a choice in Codex with ChatGPT sign-in, and OpenAI now recommends GPT-6.1 Sol in Codex for complex coding and GPT-6 Luna for focused, repeatable tasks, alongside Astra.
On the API, gpt-5.3-codex is still active with no shutdown date, and OpenAI's policy gives at least three months of notice before retiring Codex variants. Its predecessor, GPT-5.2-Codex, left the API on Jul 23, 2026, with GPT-5.6 Sol as the replacement.
On price, the natural switch is GPT-6.1 Sol: input gets 14% more expensive, but output drops 29%, cache drops 43% and the window grows to 1.05 million tokens. One caveat: OpenAI treats GPT-6.1 Sol as Critical in cybersecurity and applies Astra's safeguards to it. If you do vulnerability research, test those requests before migrating.
| GPT-5.3-Codex Feb 2026 | GPT-6.1 Sol Sep 2026 | GPT-6 Luna Sep 2026 | GPT-6 Astra Sep 2026 | |
|---|---|---|---|---|
| Input | US$ 1.75 | US$ 2 | US$ 0.10 | US$ 10 |
| Output | US$ 14 | US$ 10 | US$ 0.50 | US$ 50 |
| Cache read | US$ 0.175 | US$ 0.10 | US$ 0.01 | US$ 1 |
| Batch API input and output |
not supported | US$ 1 and US$ 5 | US$ 0.05 and US$ 0.25 | US$ 5 and US$ 25 |
| Context window | 400 thousand | 1.05 million | 1.05 million | 1.05 million |
| Knowledge cutoff | Aug 31, 2025 | Apr 30, 2026 | May 18, 2026 | Apr 30, 2026 |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026.
What changes when replacing GPT-5.3-Codex with GPT-6.1 Sol
On the API, switching to GPT-6.1 Sol touches these points:
- Identifier:
gpt-6.1-solinstead ofgpt-5.3-codex. Calls stay on the Responses API, the only one 5.3-Codex accepts. - Effort: GPT-6.1 Sol accepts
lowthroughmax, withmediumas the default. Themaxlevel did not exist on 5.3-Codex. - Tools: computer use,
apply_patch, code interpreter, file search and tool search come in, on top of the web search, hosted shell and skills 5.3-Codex already had. - Cache: on GPT-6 models, writes cost 1.25 times input, and lifetime is set with
prompt_cache_options.ttl, with a 30 minute minimum. - Context: the window grows from 400 thousand to 1.05 million tokens. Above 272 thousand input tokens, the whole request pays double input and 1.5x output.
- Knowledge: from Aug 31, 2025 to Apr 30, 2026.
Examples of using GPT-5.3 Codex
The agent gets an open issue, reads the code, writes the patch and runs the tests in the terminal. That is the kind of task in SWE-Bench Pro and Terminal-Bench, where 5.3-Codex scored 56.8% and 77.3% at launch.
OpenAI says that with simple or underspecified prompts, 5.3-Codex started delivering sites with more functionality and more sensible defaults than GPT-5.2-Codex. In the announcement example, it showed the yearly plan as a discounted monthly price.
During a long task, Codex sends frequent updates and accepts questions and changes of direction without starting over. In the app, OpenAI turns this on in the settings (Settings, General, Follow-up behavior).
Where GPT-5.3 Codex stands out and where it falls behind, by the numbers
GPT-5.3 Codex has published results on 6 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between May 11, 2025 and May 6, 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 6 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| WeirdML | 79.3% | #2 / 54 | #10 / 112 | 93.6% GPT-6 Astra |
| WebDev Arena | 1,370 | #29 / 42 | #69 / 82 | 1,820 Claude Opus 5.5 |
| Vending Bench 2 | 5,940 | #7 / 32 | #21 / 57 | 15,515 GPT-6 Astra |
| Terminal-Bench | 78.4% | #6 / 33 | #6 / 33 | 84.7% GPT-5.5 |
| SWE-bench Verified | 74.8% | #12 / 25 | #15 / 32 | 83.5% Claude Opus 4.7 |
| METR: time horizon | 350 | #4 / 17 | #4 / 29 | 719 Claude Opus 4.6 |
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-5.3 Codex 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-5.3 Codex | OpenAI | 156.76 | n/a | n/a | Feb 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-5.3 Codex or Claude Opus 4.7?
Claude Opus 4.7 is the model from another lab closest to GPT-5.3 Codex in capability. Across the 5 tests both took, this is the score:
| Test | GPT-5.3 Codex | Claude Opus 4.7 |
|---|---|---|
| Vending Bench 2 | 5,940 | 10,937 |
| WebDev Arena | 1,370 | 1,567 |
| SWE-bench Verified | 74.8% | 83.5% |
| WeirdML | 79.3% | 76.4% |
| Terminal-Bench | 78.4% | 80.2% |
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 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 |
| GPT-5.2 | Dec 11, 2025 | 153.46 | 1,454 | 74.6 | API access |
Have you used GPT-5.3 Codex? 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-5.3 Codex
How much does GPT-5.3-Codex cost?
US$ 1.75 per million input tokens and US$ 14 per million output, with cache reads at US$ 0.175. In Fast mode the price doubles, to US$ 3.50 and US$ 28. There is no batch price, because the Batch API is not supported.
Does GPT-5.3-Codex still work?
On the API, yes, through the Responses API and with no published shutdown date. In Codex with ChatGPT sign-in it stopped appearing as a choice on May 26, 2026; Codex used with an API key was not affected.
What is the difference between GPT-5.3-Codex and GPT-5.4?
5.3-Codex was built for agentic coding; GPT-5.4 brought that coding into a general-purpose model, with a 1.05 million token window, computer use and Chat Completions. In OpenAI's figures, 5.4 wins six of eight tests, and 5.3-Codex stays ahead in the terminal (77.3% against 75.1%). 5.3-Codex costs less per token.
Is GPT-5.3-Codex suitable for security research?
It was the first model OpenAI treated as High risk in cybersecurity and the first trained directly to find vulnerabilities. Higher-risk requests may be routed automatically to GPT-5.2, and defensive researchers can apply for full access through the Trusted Access for Cyber program.
What should replace GPT-5.3-Codex?
In Codex, OpenAI now recommends GPT-6.1 Sol for complex coding and agentic workflows, and GPT-6 Luna for focused, repeatable tasks. On the API, GPT-6.1 Sol costs a little more on input (US$ 2) and less on output (US$ 10), with a larger window. The independent scores are at the top of this page.
Where does GPT-5.3 Codex rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-5.3 Codex ranks #21 out of 267 models, with an ECI of 156.76. 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-5.3 Codex better than Claude Opus 4.7?
Across the 5 tests both models took and Epoch AI compiled, GPT-5.3 Codex comes out ahead on 1, behind on 4 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-5.3 Codex 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-5.3 Codex are: WeirdML (#2 of 54); Terminal-Bench (#6 of 33); METR: time horizon (#4 of 17). 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-5.3-Codex announcement
- OpenAI, GPT-5.3-Codex model page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-5.3-Codex system card
- OpenAI, GPT-5.4 announcement (table with GPT-5.3-Codex)
- OpenAI, Codex changelog (deprecation on May 26, 2026)
- OpenAI, models in Codex and ChatGPT Work
- OpenAI, API deprecations
- OpenAI, API changelog (API release on Feb 24, 2026)
- OpenAI, GPT-6.1 Sol model page (documentation)
- OpenAI, GPT-6.1 Sol safety addendum
- OpenAI, prompt caching guide
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