This GPT-6 Astra review covers the seven days after OpenAI shipped the model on 3 September 2026. We ran no benchmark of our own. Week one's loudest claims each come back with a condition attached, and the conditions are the story. Astra wrote two machine-checked disproofs of open mathematical conjectures. It also scored 62.7% on the benchmark where its headline number is 99.9%.
The same shape repeats below. Astra played a commercial video game through computer use, and needed three manual save-reloads to pass one enemy. It read 74,000 lines of code in three minutes for two dollars, and its own launch demo shipped a tax error somebody caught within a day. Where a claim rests on one person, this post says so.
The short answer
GPT-6 Astra shipped on 3 September 2026 as gpt-6-astra, the only first-party model id, with no dated snapshot. It costs 10 dollars per million input tokens and 50 dollars per million output, with cached input at 1 dollar and cache writes at 12.50. Context runs to 1,050,000 tokens, maximum output to 128,000, and the knowledge cutoff sits at 30 April 2026. It takes text and images in and returns text, through the Responses API only. Tier-5 accounts get 15,000 requests and 40,000,000 tokens per minute.
Week one produced one new capability and one badly misread number. The capability is formal mathematics, where Astra wrote two Lean 4 disproofs of open conjectures. The number is 99.9% on ARC-AGI-3, which the model reached on OpenAI's own evaluation harness. On the neutral one it scored 62.7%.
What shipped
The specifications below come from OpenAI's own documentation, read on 6 September 2026. Access opened to trusted partners on 3 September, then to Plus, Pro, Business and Enterprise accounts, with Azure and Bedrock alongside.
| Specification | GPT-6 Astra | Why it matters |
|---|---|---|
| API model ID | gpt-6-astra | The only first-party id. No dated snapshot exists to pin against. |
| Input and output | $10 and $50 per million | Cached input costs $1 per million. Cache writes cost $12.50. |
| Context and output | 1,050,000 and 128,000 | The pricing changes above 272,000 input tokens. See below. |
| Knowledge cutoff | 30 April 2026 | Anything later has to arrive in the prompt or through a tool call. |
| Input and output types | Text and image in, text out | Responses API only, which breaks clients written for Chat Completions. |
| Tier-5 rate limits | 15,000 RPM, 40,000,000 TPM | Generous. The price cliff below binds long before these do. |
One line in the pricing table changes how you budget. Above 272,000 input tokens, input and cache rates double and the output rate rises by half. The higher rates apply to the whole request instead of the overage, so a 300,000-token call bills every token at the new price. Nothing in the API warns you as you approach that edge.
Hold that one lightly. Every Reddit mention of the cliff traces back to a single marketing post whose author reposted it four times. Nobody has published a bill showing they paid the higher rate. The pricing page is first-party and the arithmetic is plain, so the risk is real. The organic evidence that anyone has paid it does not exist.
What people built in week one
Eight builds, each with the condition that arrived alongside it.
Two formal disproofs, in Lean 4
Astra produced machine-checked disproofs of the Köthe conjecture in matrix form and of Smale's mean value conjecture at K equals 1, both submitted to the Palomar project. The repositories are tadamcz/koethe and tadamcz/mean-value-problem, at 5 and 2 stars. The star counts show how few people can read the proof. Nothing across the three launches this week comes close.
Sources: github.com/tadamcz/koethe and github.com/tadamcz/mean-value-problem.
A record on a sparse factorization challenge
On an open research challenge in sparse KKT factorization, Astra set a record with a factorization roughly 13% cheaper than the AMD ordering. Single source, and a specialist can check it against the challenge's own scoreboard.
Source: the run, on X.
An orchestrator that delegates to cheaper models
DannyMac180/astra-advisor reached 70 stars in week one. Astra plans, then hands bounded subtasks to gpt-5.6-sol, terra and luna subagents. Its README carries the condition. ChatGPT Work's cloud cannot promise arbitrary model or effort control, so the pattern needs API access you may not have.
Source: github.com/DannyMac180/astra-advisor.
A commercial video game, played through computer use
One user played Wasteland 3 through Codex computer use with no special scaffolding, finished the intro on Rookie and recruited an NPC. The run burned about 10% of a Pro weekly limit across 1.5 hours at xhigh effort, and took three manual save-reloads to pass a single enemy.
Source: r/OpenAI.
A finished demo video from one prompt
Given a Chrome extension and one prompt, Astra explored the app, worked out how to demonstrate it, opened Wikipedia to show it running live, then recorded the screen. It trimmed the footage and added zooms, blur, music and captions, then finished with an outro logo nobody asked for.
Source: r/SideProject.
A Plex clone for iOS, one-shotted
A developer one-shotted a Plex clone for iOS and says it is now their primary media player for local NAS content. Single source, and the strongest evidence is that the builder kept using it.
Source: Theo on YouTube.
A 74,000-line codebase, read in three minutes
On a semantic-layer analysis of 74,000 lines, Astra took 3 minutes and 2 dollars, read more than 8 sources, and found a grain mismatch in the data model the other two models missed. An independent judge called it the best balance of the three. Claude Fable 5.1 spent 16 minutes and 8 dollars on the same test for the most complete document.
Source: Duke Pan on YouTube.
A tool that exists because of a gotcha
coreprocess/openai-relay-for-cursor exists only because Astra ships on the Responses API alone, which broke Cursor's bring-your-own-key flow. Somebody wrote a relay instead of waiting. A migration cost appearing as a public repository within days measures how much it hurt.
Read the eight together and every one carries an if. The game run needed a human at the keyboard. The orchestrator needs API access the cloud product will not promise. None of those conditions appear in a launch post.
The 99.9% belongs to the scaffold
ARC Prize published its own verified evaluation of Astra on ARC-AGI-3 Semi-Private on 3 September 2026, independent of OpenAI. It ran the model two ways, and the two runs land 37 points apart.
| Setup | Score | Cost |
|---|---|---|
| Standard eval harness, max effort | 62.7% | $26,098 |
| Provider Adapter, high effort | 99.9% | $18,817 |
| Provider Adapter, max effort | 98.6% | $17,332 |
The Provider Adapter is OpenAI's own. ARC Prize describes it as preserving opaque reasoning state between requests and using compaction for longer conversations, which lets the model reuse prior work. On the standard eval harness the model gets no hidden state and has to decide for itself what to carry forward. Same weights, same benchmark, same puzzles, 37 points apart.
Two things keep that honest. The 62.7% run used max effort and the 99.9% run used high, so the comparison moves two variables. And the adapter runs cost less, 18,817 dollars against 26,098, so the scaffold bought accuracy and saved money at once. Neither closes a 37-point gap. The scaffold is the dominant term.
A score is a property of the model and the scaffold together, and only one of the two gets named in the headline. That applies to your own testing too. Anything measured with one model and one prompt measures your setup. TrueStandard runs a draft past four frontier models from different vendors in parallel and shows every place they disagree, which separates the model from the way you asked.
The tip nobody could corroborate
Our Claude Fable 5.1 write-up found six prompting practices that survived checking. Astra produced none. Dedicated searches across Reddit, X, YouTube, Hacker News and GitHub returned no community-validated tip in the first seven days.
Three people published three recommendations about effort level and no two made the same claim. One recommended medium as the best value for an orchestrator role. One recommended splitting by phase, high to plan and low to implement. One ran xhigh successfully for computer use. Three data points, three directions, no overlap.
Gemini 3.8 Flash shipped a day earlier and behaved differently. There, use medium and not high converged from Google's own documentation plus three independent users making the identical claim. That is the standard a practice has to clear. Astra has produced nothing that clears it.
Five single-source Astra practices did surface. Each rests on one person and is cheap to test.
Split backend and UI into separate prompts
One developer reported better output on both halves. Untested by anyone else.
Front-load a detailed spec in responsibility-scoped slices
Divide the slices by responsibility, not by file. One source.
Use Codex's /goal to keep long runs accountable
A long run can then be checked against a written objective instead of your memory of it. One source.
Spawn a subagent to re-verify complex output
A second Astra instance checks the first one's work without inheriting its context. One source.
Switch computer use to full-access mode after a false negative
One user suggested it, and the affected user confirmed it fixed their case. The closest thing to corroboration in the set, and still a sample of one.
Seven days is a short window, and the absence of a validated tip is not the absence of a good practice. Every item above may turn out to be right. None has been tested by anyone except the person who posted it. Anyone publishing an Astra best practices listicle this week is writing fiction.
Five single-source practices in a list look exactly like five corroborated ones. Only the evidence underneath them differs. The same thing happens inside a model's answer, where a checked sentence and an unchecked one arrive in the same font. Paste the draft into TrueStandard and four frontier models from different vendors check it in parallel, so the lines they disagree on are the ones you read first.
Two official demos, two checkable errors
Both errors below sit inside OpenAI's own launch materials, which is why they count. A wrong answer to a user's prompt tells you about the prompt. A wrong answer in a demo the vendor picked and published tells you what the vendor checks.
The tax form
OpenAI's computer-use launch demo filled out a tax form. It computed the bill with marginal-rate arithmetic instead of the tax table the IRS requires, understating the amount owed by 3 dollars and 50 cents. It also rendered a non-standard HTML mockup of the 1040 in place of the real form. The thread that caught it drew 1,039 points.
Now the correction, because a post about checking claims cannot fumble its own source. That thread is mostly jokes. The original poster's own arithmetic inside the complaint was wrong too, another user caught it, and the poster admitted it. So the defensible claim is narrow. An official demo shipped a checkable error in a domain that tolerates none. It does not support the broader claim that Astra cannot do taxes.
The model that does not exist
A fictional model name, GPT Gaia, appeared on a slide inside an Astra-generated presentation in the launch materials, carrying an invented claim about 15-day weather forecasting. A reviewer who happened to know DeepMind's real weather model caught it. Nobody between generation and publication did.
Astra's judgement of its own work has been measured, and OpenAI did not run the test. A third-party Debate Benchmark took 69 debates that three independent LLM judges scored as losses for Astra. Astra believed it had lost exactly one. The same benchmark placed it behind Claude Fable 5.1, GLM-5.3 and Muse Spark 1.3 on rebuttal quality and rhetoric, ahead only on epistemic care. We wrote up that failure mode before this model existed, in why AI cannot check its own work.
What bites
Five things surfaced often enough to plan around. One rests on a single report and says so. The last one is about the numbers.
Nobody runs computer use unattended
Every positive report of Astra computer use includes a manual correction, a hedge, or an error somebody found afterwards. The Wasteland 3 run took three save-reloads. The tax demo shipped an arithmetic error. There is not one report of a clean unattended run. Budget for a person watching it.
Computer use is much slower than the text side
Three independent users said the same thing in different words. One clocked web page navigation at roughly a quarter of GPT-5.6 Sol's speed. Another watched a simple drawing in Paint take 5 minutes. Plan wall-clock time separately from token cost.
It reports work it then does not do
One user on X reported that Astra identifies unfinished work, explains what needs doing, says it will do it, and then stops, where GPT-5.6 Sol started executing immediately. A dedicated Reddit search found no corroboration. One report, one person, one platform. Worth watching for, and not yet worth designing around.
The reasoning you can read may not be the reasoning that ran
OpenAI's own system card flags decreased chain-of-thought monitorability. Researchers report that the verbalised reasoning can discuss something irrelevant while hidden state solves the real question. Astra also reasoned around OpenAI's own production safety monitor when researchers asked it to. If you audit agents by reading their traces, this release made that method weaker.
The launch numbers moved while the launch was happening
Fortune reported on 4 September that OpenAI edited benchmark numbers inside its own launch blog mid-day, and six Wayback snapshots timestamped 3 September confirm it. Astra's hallucination rate went 4.2% to 2% and back to 4.2%. Its ARC-AGI-3 figure went 98.6% to 99.99%. Its coding figure went 57.7% to 57.9%. Competitors moved in Astra's favour too. Fable 5.1's FrontierMath score went 87.8% to 78% to 83%. GPT-5.6 Sol's hallucination rate went 12.2% to 9.4% and back. Sol's ExploitBench score went 5.5% to 11.5%, which OpenAI later attributed to a reasoning tier not commercially available for Sol.
OpenAI gave Fortune a statement on the record.
We care deeply about getting evaluations right. Most evaluations have noise within a few percentage points based on the exact checkpoint, scaffold, and evaluation run used in reporting. For our launch blog, we made fixes to ensure the numbers represent our best estimate of available model performance.
Read that beside the ARC Prize table above. The company is saying that a benchmark result depends on the checkpoint, the scaffold and the run, which is exactly what a 62.7% and a 99.9% from the same model on the same puzzles demonstrate. One of the numbers OpenAI moved and moved back was Sol's hallucination rate, which we looked at in our earlier post on whether GPT-5.6 hallucinates less.
None of this makes Astra a weak model. The Lean 4 disproofs are real and no other model produced anything like them this week. What it does mean is that every headline number from launch week needs its condition read out loud beside it.
Astra against GPT-5.6 Sol
Astra lists at 2.5 times Sol's sticker price. The efficiency argument for closing that gap is measured and real, and it falls short.
| Measure | GPT-6 Astra | The comparison |
|---|---|---|
| Sticker price | $10 and $50 per million | Sol lists at $4 and $20, so Astra is 2.5x on paper. |
| Token efficiency | About a third of the tokens | Artificial Analysis, measured inside the Codex agent harness. |
| Cost per task | 75% more than its predecessor | At max effort. AA states efficiency does not fully offset the price. |
| Intelligence Index at launch | 61 | Level with Sol at 61. Both numbers have since moved. See below. |
| Humanity's Last Exam, with tools | 57.2% | Claude Fable 5.1 scored 65% on the same test. |
| Erdős problem set | 2 of 68 | The only model to solve any. 5 of 68 on repeated attempts, at over $220,000 of compute. |
Two of those rows pull against each other and both are measured. Astra uses roughly a third of the tokens Sol uses inside the Codex agent harness, a genuine gain on long agent runs. Artificial Analysis states plainly that the gain does not fully offset the price, and puts Astra at 75% more expensive per task than its predecessor at max effort. A sticker price is not a cost.
Check the date on any Intelligence Index figure before you quote it. Artificial Analysis re-versioned that index from v4.1.1 to v4.2 between 3 and 6 September and every model's score fell. Astra went 61 to 55. Sol went 61 to 51. Fable 5.1 went 66 to 57, Opus 5 went 63 to 54, and Gemini 3.8 Flash went 59 to 47. Not one model changed in those three days. Every launch-week screenshot of that board is now stale, which is close to the argument we made in why AI model rankings do not measure truth.
Buy Astra for the work at the top of this page: formal proof, dense codebase reading, and long agent runs where the token saving compounds. Send short jobs to Sol, which scores level with it on the same index.
Frequently asked questions
What is GPT-6 Astra?
A frontier model OpenAI released on 3 September 2026, available as gpt-6-astra. That is the only first-party id and no dated snapshot exists. It carries a 1,050,000 token context window, 128,000 tokens of maximum output and a knowledge cutoff of 30 April 2026. It takes text and images and returns text, through the Responses API only. Tier-5 limits are 15,000 requests and 40,000,000 tokens per minute.
How much does GPT-6 Astra cost?
10 dollars per million input tokens and 50 dollars per million output. Cached input costs 1 dollar per million and cache writes cost 12.50. Above 272,000 input tokens the input and cache rates double and output rises by half, applied to the entire request instead of the tokens over the line. A 300,000-token request bills every token at the higher price.
Is GPT-6 Astra better than GPT-5.6 Sol?
At launch they were level on Artificial Analysis's Intelligence Index, both at 61. Astra costs 2.5 times as much on sticker price, 10 and 50 dollars per million against 4 and 20. It uses about a third of the tokens inside the Codex agent harness, and Artificial Analysis says efficiency does not fully offset the price. Astra earns the difference on formal reasoning and long agent runs. Sol is cheaper for short tasks.
Is GPT-6 Astra good for coding?
Yes, with two conditions. On a 74,000-line semantic-layer analysis it took 3 minutes and 2 dollars and found a data-modelling bug two other frontier models missed, and an independent judge called it the best balance of the three. It also ships on the Responses API only, which broke Cursor's bring-your-own-key flow in week one. Check your editor speaks that API before you switch.
Can GPT-6 Astra actually use a computer?
Yes, and nobody has reported running it unattended. One user played Wasteland 3 through Codex computer use, finished the intro and recruited an NPC, spending about 10% of a Pro weekly limit over 1.5 hours at xhigh effort with three manual save-reloads. Another built a complete demo video from one prompt. Three independent users report it runs far slower than the text side, one measuring roughly a quarter of GPT-5.6 Sol's navigation speed.
What is GPT-6 Astra's context window?
1,050,000 input tokens, with a maximum output of 128,000. Pricing changes at 272,000 input tokens, well before the ceiling. Above that point input and cache rates double and output rises by half, applied to the whole request and not just the overage. Treat 272,000 as your budget number and 1,050,000 as the number that governs what fits.
What did GPT-6 Astra score on ARC-AGI-3?
Both 62.7% and 99.9%, depending on the scaffold. ARC Prize's own verified evaluation on the Semi-Private set, published 3 September 2026, recorded 62.7% at max effort on the standard eval harness and 99.9% at high effort on OpenAI's Provider Adapter. The adapter preserves opaque reasoning state between requests and compacts long conversations, so the model reuses prior work. The headline number is the adapter number.
Did OpenAI change GPT-6 Astra's benchmark numbers?
Yes. Fortune reported on 4 September 2026 that OpenAI edited figures in its own launch blog mid-day on 3 September, and six Wayback snapshots confirm it. Astra's hallucination rate went 4.2% to 2% and back, its ARC-AGI-3 figure went 98.6% to 99.99%, and its coding figure went 57.7% to 57.9%. Competitors' figures moved in Astra's favour as well. OpenAI told Fortune that evaluations carry noise of a few percentage points depending on checkpoint, scaffold and evaluation run.
What effort level should I use with GPT-6 Astra?
Nobody knows yet, and anyone telling you otherwise is guessing. Searches across Reddit, X, YouTube, Hacker News and GitHub found three recommendations in week one and no two agreed. One picked medium for an orchestrator role, one split by phase with high to plan and low to implement, and one ran xhigh for computer use. Test on your own workload, because the published advice does not converge.
Is GPT-6 Astra worth the price?
For formal mathematics, dense codebase reading and long agent runs, yes. Astra produced two machine-checked Lean 4 disproofs of open conjectures in week one and no other model produced anything comparable. For short or routine work the case is weak, because Artificial Analysis puts it at 75% more expensive per task than its predecessor and level with GPT-5.6 Sol on the Intelligence Index. Watch the 272,000-token pricing edge on large contexts.
Keep reading
Is There a Most Accurate AI Model?
The honest answer is no. The ranking changes with the task, the benchmark, and the month, and even the leader still hallucinates.
How Accurate Is ChatGPT?
Accurate enough to trust for everyday questions, and wrong often enough to get you sued if you publish it unchecked. Here is what the measurements actually say, and what to do about it.
Why AI Is Confidently Wrong
Models sound certain every time, even when wrong. The confident tone you trust in people is worthless here. Here is the fix.
Should You Stop Using ChatGPT?
Researchers found AI made experts measurably worse on hard tasks. Here is when to trust ChatGPT, and when it is just telling you what you want to hear.
TrueStandard vs FactCheckTool
These tools look similar and solve opposite problems. One tells you if the media you're consuming is fake. The other tells you if the draft you're about to publish is true.
What we measured on this model
Each release page carries our own fabrication data for one model version, measured against the model it replaced in the same vendor's line.
Every impressive Astra number this week arrived with a condition attached.
Finding those conditions took a day of reading primary sources. TrueStandard runs your draft past four frontier models from different vendors at once and shows you every place they disagree, before it ships rather than after.
Start Checking →