GPT-6 Astra
GPT-6 Astra is OpenAI's top model, released on Sep 3, 2026. It costs US$ 10 per million input tokens and US$ 50 per million output, rereads the cache for US$ 1 and accepts up to 1.05 million tokens of context, with a higher price above 272 thousand tokens. In OpenAI's own table it leads on science, computer use and automation, but trails Claude models on general knowledge. It is meant for the hardest work; for everyday tasks, GPT-6.1 Sol comes close at one fifth of the price per token.
GPT-6 Astra 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-6-astra- Release
- Sep 3, 2026, first to a limited set of organizations
- Input
- US$ 10 per million tokens
- Output
- US$ 50 per million tokens
- Cache read
- US$ 1 per million (10% of input)
- Cache write
- US$ 12.50 per million (1.25x input)
- Above 272 thousand input tokens
- input and cache doubled and output 1.5x, for the whole request: US$ 20 and US$ 75
- Batch and Flex
- half price: US$ 5 input and US$ 25 output
- Fast mode
- US$ 20 and US$ 100, up to 2x faster, no latency SLA
- Ultrafast
- US$ 60 and US$ 300, on the Responses API since Sep 29, 2026, no EU data residency
- Context window
- 1,050,000 tokens (up to 922 thousand input)
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Apr 30, 2026
- Reasoning
- effort
lowtomax; nonone, and OpenAI does not state the API default - Where it runs
- OpenAI API (Responses, Chat Completions and Batch), ChatGPT, Codex, Microsoft Azure and Amazon Bedrock (us-west-2)
- Cyber risk
- Critical on OpenAI's Preparedness scale, the first OpenAI model to get there
Where GPT-6 Astra does well, where it falls behind and who it is for
Where it does well
- Science in the terminal: 64.6% on Terminal-Bench Science in OpenAI's table, against 52.6% for Claude Fable 5.1. Even Anthropic's table has Astra ahead of Opus 5.5 here (64.6% against 58.7%).
- Computer use: 72.6% on OSWorld 2.0 offline, against 70.2% for Claude Opus 5, and about 47% less time per task than GPT-5.6 Sol, according to OpenAI.
- Automation: 41.4% on AutomationBench, against 31.4% for Fable 5.1. In Anthropic's table it also edges out Opus 5.5, at 40.0%.
- Long context: 96.3% on OpenAI's MRCR v2, which looks for eight pieces of information between 512 thousand and 1 million tokens, against 73.8% for GPT-5.6 Sol.
- Fewer factual errors: 4.2% on OpenAI's internal hallucination test, against 12.2% for GPT-5.6 Sol (lower is better). It is the maker's own measurement, with no outside check.
Where it falls behind
- Price: five times GPT-6.1 Sol per token, and above 272 thousand input tokens the whole request rises to US$ 20 and US$ 75.
- General knowledge: 57.2% on Humanity's Last Exam with tools, against 65.0% for Fable 5.1 in OpenAI's own table. In Anthropic's table, Opus 5.5 scores 67.7%.
- Office work and long coding: in Anthropic's table, 1,542 points on GDPval-AA v2.1 against 1,846 for Opus 5.5, and 57.9% on Terminal-Bench 4.0 against 66.4%.
- Safety locks: misalignment monitoring can stop a task on the API with a 403 error, with no way to resume the conversation. OpenAI expects the safeguards to create more friction at launch than it ultimately intends.
- ChatGPT limits: on the Plus plan, the official estimate is 5 to 45 messages per 5-hour window in Work and Codex, against 10 to 100 for GPT-5.6 Sol.
Who it makes sense for
- Hard scientific research done in the terminal, which OpenAI still reserves for Astra even after launching 6.1 Sol.
- Agents that operate a computer and browser for long stretches, with screens, clicks and forms.
- Business process automation across several tools and steps.
- Very long material where scattered information has to be cross-checked, accepting the higher price above 272 thousand tokens.
When to pick another model
- Everyday and high-volume work: GPT-6.1 Sol costs one fifth and scored 75.2% on DeepSWE against 74.1% for Astra in OpenAI's charts.
- Office work built from documents, where Claude Opus 5.5 scored higher in Anthropic's table and costs US$ 4 and US$ 20.
- Apps that need answers without reasoning, or that rely on
temperatureorlogprobs: Astra accepts none of the three. - Offensive security, such as exploit proofs of concept: Astra refuses by default, and the route is OpenAI's Daybreak program.
What OpenAI published at launch
Astra came out on Sep 3, 2026, first to a limited set of organizations, and OpenAI's help center acknowledges a delay in the broader rollout. The company calls it the world's most intelligent and aligned model.
The table below is the one from the announcement. Each score is the highest at any effort level, and OpenAI picked the tests and the rivals. Even so, the full table shows Astra behind some Claude model in five rows, including hard knowledge and FrontierCode, a coding test.
Two figures changed later, in OpenAI's own charts. In the GPT-6.1 Sol announcement, Astra shows 68.1% on Terminal-Bench Science, against 64.6% here, and 73.5% on OSWorld 2.0 offline, against 72.6%. We kept the launch figures.
| GPT-6 Astra OpenAI | GPT-5.6 Sol OpenAI | Claude Fable 5.1 Anthropic | Claude Opus 5 Anthropic | |
|---|---|---|---|---|
| Terminal-Bench 4.0 long coding in the terminal |
57.9% | 37.3% | 55.8% | 52.6% |
| DeepSWE v1.1 software engineering |
74.1% | 72.7% | 67.4% | 73.7% |
| FrontierCode 1.1 Main coding |
53.3% | 47.5% | 50.9% | 53.4% |
| OSWorld 2.0 offline computer use |
72.6% | 65.7% | not measured | 70.2% |
| AutomationBench automation |
41.4% | 18.1% | 31.4% | 26.9% |
| BrowseComp web research |
91.5% | 90.4% | not measured | 90.8% |
| Terminal-Bench Science 0.1 science in the terminal |
64.6% | 22.4% | 52.6% | 30.0% |
| FrontierMath Tier 4 research math |
97.6% | 83.0% | 87.8% | 73.2% |
| Humanity's Last Exam hard knowledge, with tools |
57.2% | not measured | 65.0% | 63.6% |
| Artificial Analysis Intelligence Index overall index, v4.1.1 |
61.2 | 60.9 | 65.7 | 63.1 |
Figures from the GPT-6 Astra announcement, highest score at any effort. The dot marks the best in each row. "Not measured" means OpenAI published no figure for that model.
Astra against Claude Fable 5.1 and Opus 5.5: price
The closest rival on price is Anthropic's Claude Fable 5.1: the same US$ 10 input and US$ 50 output per million tokens. The difference is at the edges. Fable 5.1 rereads the cache for US$ 0.25, a quarter of Astra's price, and keeps the same rates up to 1 million tokens. Astra doubles input and charges 1.5x output once a request goes past 272 thousand tokens.
Claude Opus 5.5 costs US$ 4 and US$ 20, less than half. In the table Anthropic published on Sep 22, it beats Astra on four of the six tests where both appear: Terminal-Bench 4.0, FrontierCode 1.1, GDPval-AA v2.1 and Humanity's Last Exam. Astra wins science in the terminal and automation. The independent scores are at the top of this page.
| GPT-6 Astra OpenAI | Fable 5.1 Anthropic | Opus 5.5 Anthropic | |
|---|---|---|---|
| Input | US$ 10 | US$ 10 | US$ 4 |
| Output | US$ 50 | US$ 50 | US$ 20 |
| Cache read | US$ 1 | US$ 0.25 | US$ 0.20 |
| Cache write | US$ 12.50 | US$ 12.50 | US$ 5 |
| Input above 272 thousand tokens | US$ 20 | US$ 10 | US$ 4 |
| Output above 272 thousand tokens | US$ 75 | US$ 50 | US$ 20 |
| Batch | US$ 5 and US$ 25 | US$ 5 and US$ 25 | US$ 2 and US$ 10 |
| Fast mode | US$ 20 and US$ 100 | not offered | US$ 8 and US$ 40 |
| Context window | 1.05 million | 1 million | 1 million |
OpenAI and Anthropic pricing pages, checked on Oct 1, 2026. For Anthropic, 5-minute cache writes; the 1-hour write costs US$ 20 on Fable 5.1 and US$ 8 on Opus 5.5.
Astra, 6.1 Sol, Sol or Luna: choosing within OpenAI
OpenAI arranges the family in tiers: Astra at the top, GPT-6.1 Sol in the middle at one fifth of the price per token, and GPT-6 Luna at the base, for focused, high-volume tasks. The original GPT-6 Sol no longer appears in the company's model guide.
In the charts OpenAI published with 6.1 Sol, the gap to Astra is small in coding and computer use and wide in science: 57.0% against 68.1% on Terminal-Bench Science. The company's advice follows the same line: 6.1 Sol for complex work where cost matters, Astra for the most demanding.
| GPT-6 Astra | GPT-6.1 Sol | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|---|
| Input | US$ 10 | US$ 2 | US$ 2 | US$ 0.10 |
| Output | US$ 50 | US$ 10 | US$ 10 | US$ 0.50 |
| Cache read | US$ 1 | US$ 0.10 | US$ 0.20 | US$ 0.01 |
| Cache write | US$ 12.50 | US$ 2.50 | US$ 2.50 | US$ 0.125 |
| Above 272 thousand tokens input and output |
US$ 20 and US$ 75 | US$ 4 and US$ 15 | US$ 4 and US$ 15 | US$ 0.20 and US$ 0.75 |
| Context window | 1.05 million | 1.05 million | 1.05 million | 1.05 million |
| Knowledge cutoff | Apr 30, 2026 | Apr 30, 2026 | Apr 20, 2026 | May 18, 2026 |
| Accepts effort none | no | no | yes | yes |
| Role, per OpenAI | highest intelligence | balance of cost and intelligence | superseded by 6.1 Sol | fastest and most cost-effective |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026.
Before adopting: behavior and API
OpenAI warns that Astra stops to ask when users may expect it to make reasonable assumptions and keep going, leans toward long, formatted answers and is more sensitive to instructions in files such as AGENTS.md, which it recommends auditing.
On the API, four points break integrations built for earlier models:
- No
noneeffort (it returns a 400 error), and no customtemperature,top_porlogprobs. - Tool calls only through the Responses API. On Chat Completions, Astra answers but does not call functions.
- Misalignment monitoring can stop the task. A block before streaming returns a 403 with the code
misalignment_policy_violation, and stopping does not undo actions already taken. - Anyone coming from GPT-5.5 or earlier must replace
prompt_cache_retentionwithprompt_cache_options.ttl.
Examples of using GPT-6 Astra
An agent runs analyses, adjusts simulation code and checks the results over hours of terminal work. This is where Astra has its widest lead, and OpenAI recommends it for the hardest scientific tasks.
The agent logs into a legacy system, fills in screens and cross-checks data between programs, the way a person would. OSWorld 2.0 measures this kind of task, and Astra scored 72.6%.
The contract, annexes and amendments go in whole, and the model cross-checks clauses scattered across the material. The window fits it, but above 272 thousand tokens the whole request gets more expensive: input rises to US$ 20 per million.
Where GPT-6 Astra stands out and where it falls behind, by the numbers
GPT-6 Astra has published results on 36 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 7, 2025 and Dec 2, 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 36 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 95.8% | #2 / 76 | #2 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 78 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| Dtbench | 97.3% | #5 / 66 (tied with 2) | #5 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 64.37 | #3 / 66 | #3 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 31.7% | #2 / 77 (tied with 2) | #2 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 93.6% | #1 / 50 | #1 / 112 | 93.6% GPT-6 Astra |
| SciCode | 56.5% | #21 / 75 (tied with 1) | #21 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 93.7% | #1 / 67 (tied with 1) | #1 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 95.0% | #1 / 46 | #1 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 2,951 | #1 / 60 | #1 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,800 | #2 / 70 | #2 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 98.5% | #1 / 46 (tied with 4) | #1 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 75.6% | #2 / 62 | #2 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 76.0% | #1 / 62 | #1 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 84.0% | #1 / 60 | #1 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 99.0% | #4 / 67 (tied with 3) | #4 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 97.6% | #2 / 57 | #2 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 15,515 | #1 / 46 | #1 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 64.7% | #8 / 48 | #8 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 53.3% | #4 / 43 | #4 / 43 | 54.6% Claude Opus 5.5 |
| Hle | 54.8% | #1 / 14 | #1 / 39 | 54.8% GPT-6 Astra |
| Gdp Pdf | 34.2% | #2 / 37 | #2 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 80.0% | #2 / 33 (tied with 1) | #2 / 34 | 83.3% Claude Opus 5.5 |
| BALROG | 68.3% | #1 / 7 | #1 / 32 | 68.3% GPT-6 Astra |
| Blueprint Bench 2 | 49.7% | #3 / 30 | #3 / 31 | 54.4% Gemini 4 Argon |
| Gso | 79.4% | #2 / 13 | #2 / 28 | 88.2% Claude Fable 5.1 |
| Ebr Bench | 76.2% | #1 / 23 | #1 / 27 | 76.2% GPT-6 Astra |
| DeepSWE | 74.1% | #1 / 26 | #1 / 26 | 74.1% GPT-6 Astra |
| FrontierSWE | 65.5% | #1 / 20 | #1 / 20 | 65.5% GPT-6 Astra |
| Rli | 20.8% | #2 / 10 | #2 / 15 | 21.3% Claude Opus 5.5 |
| Weirdml V3 | 42.2% | #1 / 10 | #1 / 11 | 42.2% GPT-6 Astra |
| MirrorCode | 46.7% | #4 / 9 | #4 / 9 | 77.4% Claude Opus 5.5 |
| Frontiermath Erdos | 2.9% | #1 / 7 | #1 / 7 | 2.9% GPT-6 Astra |
| Spireclimb | 25.7% | #1 / 7 | #1 / 7 | 25.7% GPT-6 Astra |
| Spireclimb Sts1 A20 Heart | 16.7% | #2 / 7 | #2 / 7 | 20.0% GPT-6.1 Sol |
| Spireclimb Sts2 A10 | 34.7% | #1 / 7 | #1 / 7 | 34.7% GPT-6 Astra |
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-6 Astra 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 |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 166.82 | 1,584 | 83.2 | Sep 22, 2026 |
| GPT-6 Astra | OpenAI | 166.41 | n/a | 82.2 | Sep 3, 2026 |
| Claude Fable 5.1 | Anthropic | 164.65 | 1,586 | 83.4 | Sep 1, 2026 |
| Gemini 4 Argon | Google DeepMind | 164.56 | 1,608 | n/a | Sep 30, 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 |
| Grok Build 0.1 | xAI | no ECI | n/a | 67.8 | May 29, 2026 |
GPT-6 Astra or Claude Opus 5.5?
Claude Opus 5.5 is the model from another lab closest to GPT-6 Astra in capability. Across the 30 tests both took, this is the score:
| Test | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|
| Spireclimb Sts1 A20 Heart | 16.7% | 0.0% |
| Spireclimb | 25.7% | 3.3% |
| Spireclimb Sts2 A10 | 34.7% | 6.7% |
| Vending Bench 2 | 15,515 | 9,235 |
| MirrorCode | 46.7% | 77.4% |
| ALE-Bench | 2,951 | 2,147 |
| SciCode | 56.5% | 66.9% |
| Game puzzles | 84.0% | 71.0% |
| Apex Agents | 64.7% | 73.5% |
| Chess puzzles | 76.0% | 67.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What GPT-6 Astra 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.
DeepSWE
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-6.1 Sol | Sep 29, 2026 | 166.32 | n/a | 81.6 | API access |
| GPT-6 Sol | Sep 22, 2026 | 162.55 | n/a | 79.3 | API access |
| GPT-6 Luna | Sep 22, 2026 | 156.31 | n/a | 72.0 | API access |
| GPT-6 Astra | Sep 3, 2026 | 166.41 | n/a | 82.2 | API access |
| GPT-5.6 Sol | Jul 9, 2026 | 161.73 | 1,527 | 81.1 | API access |
| GPT-5.6 Terra | Jul 9, 2026 | 159.43 | 1,515 | 77.9 | API access |
| GPT-5.6 Luna | Jul 9, 2026 | 156.33 | 1,478 | 73.6 | API 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 |
Have you used GPT-6 Astra? 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-6 Astra
How much does GPT-6 Astra cost?
US$ 10 per million input tokens and US$ 50 per million output, with cache reads at US$ 1 and cache writes at US$ 12.50. Above 272 thousand input tokens, the whole request moves to US$ 20 input and US$ 75 output. Batch and Flex are half price, Fast mode costs double and Ultrafast six times as much.
Is GPT-6 Astra better than Claude Opus 5.5?
It depends on who published the table and on the task. In Anthropic's table, Opus 5.5 wins four of the six tests where both appear, including long coding in the terminal (66.4% against 57.9%). Astra wins science in the terminal (64.6% against 58.7%) and automation (41.4% against 40.0%). Opus 5.5 costs US$ 4 and US$ 20 per million tokens. The independent scores are at the top of this page.
Is it worth paying five times the price of GPT-6.1 Sol?
For most tasks, OpenAI's own charts say no: 6.1 Sol matches or comes close on coding and computer use at one fifth of the price per token. The biggest gap is in science, where Astra scored 68.1% against 57.0% on Terminal-Bench Science. Test with your own tasks before deciding.
Where can I use GPT-6 Astra?
On the OpenAI API, through the Responses API, Chat Completions and Batch, and in Codex. In ChatGPT, GPT-6 Pro, powered by Astra, is on the US$ 100 and US$ 200 Pro plans, Business and Enterprise, and Plus gets Astra in ChatGPT Work and Codex. The announcement also names Microsoft Azure and Amazon Bedrock, where OpenAI lists the us-west-2 region.
Where does GPT-6 Astra rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-6 Astra ranks #2 out of 267 models, with an ECI of 166.41. 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-6 Astra better than Claude Opus 5.5?
Across the 30 tests both models took and Epoch AI compiled, GPT-6 Astra comes out ahead on 16, behind on 11 and tied on 3. 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-6 Astra 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-6 Astra are: Chess puzzles (#1 of 62); ALE-Bench (#1 of 60); Game puzzles (#1 of 60). 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-6 Astra announcement
- OpenAI, GPT-6 Astra model page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6 Astra system card
- OpenAI, Path to Astra (safeguards at launch)
- OpenAI, latest model guide (behavior and migration)
- OpenAI, misalignment monitoring (documentation)
- OpenAI, API changelog
- OpenAI, GPT-6.1 Sol announcement (charts including Astra)
- OpenAI, managing GPT-6 Astra usage in Work and Codex (help center)
- OpenAI, GPT-6 Pro in ChatGPT (help center)
- Anthropic, Claude Opus 5.5 announcement (table including GPT-6 Astra)
- Anthropic, pricing (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