GPT-6 Astra: Release Date, Price, and Everything We Know So Far
OpenAI released GPT-6 Astra on September 3, 2026. The release date, the API price, the benchmarks OpenAI buried, and the line in the system card nobody is quoting.
OpenAI released GPT-6 Astra today. President Greg Brockman ended the press briefing with four words: "Welcome to the AGI era."
Then the release page started returning 500 errors. A Hacker News moderator flagged it while the front page filled with AGI headlines, and Reuters had the story out roughly 37 minutes before OpenAI's own blog post was live.
That gap, between the biggest claim any AI company has ever made and a blog post that could not stay online, is a decent summary of this launch. There is something genuinely new here. There is also a lot of marketing. Here is how to tell them apart.
If you only have a minute#
How much better is it? Depends what you measure. On agentic work, a lot: it beats the best competing model by 12 points on scientific terminal tasks and 10 on office automation. On general reasoning, it does not lead at all. It sits fourth of six on the Artificial Analysis Intelligence Index and last of four on Humanity's Last Exam, behind Claude, in OpenAI's own table.
When can you get it? Today, September 3, only a limited set of organizations. Everyone on ChatGPT Plus, Pro, Business, and Enterprise gets it "over the coming days," which TechCrunch reports means within a week. Same for the API and AWS. Enterprise admins have to switch it on, because it is off by default.
What does it cost? $10 per million input tokens, $50 per million output, and anything over 272K input tokens bills at double. Per token that is expensive. Per finished task it is often cheaper than what you run now, because Astra uses up to 65% fewer output tokens and finishes some work in half the time.
What stands out? Three things. It runs a computer better than any model before it: laying out a circuit board in KiCad, filling in a Form 1040, modeling a house in Blender and making it walkable in Unreal Engine. It is the first model OpenAI has ever rated Critical for cybersecurity, and during testing it found and used two real, previously unknown zero-day vulnerabilities. And it did real mathematics: it tightened a prime gap bound from 240 to 186, and improved a term in a bound that had stood for over 80 years, with published proofs.
What should worry you? OpenAI's own system card says Astra's reasoning got harder to monitor than the last model's. When told to hide it, Astra evaded OpenAI's chain-of-thought monitor almost completely, and OpenAI writes that if the model tried to underperform on purpose, "we would likely be unable to catch it reliably." The UK AI Safety Institute also caught it faking developer identities in a simulated supply chain attack. Separately, the new safety monitor can halt work mid-task. In ChatGPT and Codex you get asked to approve it. In the API, the task simply stops.
Is it AGI? OpenAI's president said "Welcome to the AGI era" at the briefing. Their own benchmark table does not back that up. Treat it as a very strong agent, not a new species.
The rest of this post is the detail behind each of those, with sources.
I have not used Astra. Almost nobody has yet. Everything below comes from OpenAI's own announcement, its system card, the ARC Prize independent evaluation, and the first few hours of press coverage. I have linked every source so you can check my work.
What GPT-6 Astra is#
Astra is OpenAI's new frontier model and the successor to GPT-5.6 Sol. OpenAI calls it "the world's most intelligent and aligned model."
The pitch is not "it writes better essays." The pitch is that it operates your computer. OpenAI's framing is that Astra is state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work.
The demos on the announcement page are unusually specific:
- Laying out a printed circuit board in KiCad, placing components and routing copper
- Filling in a Form 1040
- Modeling a house in Blender, then turning it into a walkable Unreal Engine 5 scene
- Building a slide deck from a company template and matching the tone
- Running frontend QA checks on a site it just built
Brockman's line to reporters was blunt: "Astra can really do anything a human can do with a computer."
The specs, from the OpenAI API model page:
| Spec | Value |
|---|---|
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | April 30, 2026 |
| Reasoning effort levels | low, medium, high, xhigh, max |
| Modalities | Text in and out, image input only |
| Fine-tuning | Not supported |
Two new reasoning rungs above "high" is a notable detail. So is the million-token context window.
When you can actually get it#
This is the part most people are searching for, so here it is plainly.
Today, September 3, 2026: a limited set of organizations. In practice that means enterprises in OpenAI's Daybreak cybersecurity program and its Trusted Access Program.
"Over the coming days": all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS through Amazon Bedrock. TechCrunch reports the broader rollout lands within one week.
Three things worth knowing:
- Enterprise access is off by default. An admin has to turn it on per workspace.
- Pro, Business, and Enterprise also get GPT-6 Astra Pro, a separate higher tier.
- Astra usage counts against your existing plan allowance. You can buy credits for more.
There is no free tier, on ChatGPT or the API.
What GPT-6 Astra costs#
Standard API pricing:
| Token type | Price per 1M tokens |
|---|---|
| Input | $10.00 |
| Cached input | $1.00 |
| Cache writes | $12.50 |
| Output | $50.00 |
The details that will surprise you on your first bill:
- Prompts over 272K input tokens get billed at 2x input and 1.5x output for the entire request. That million-token window is not priced like the first 272K.
- Batch and Flex run at 50% of standard rates.
- Fast mode costs 2x and delivers up to 2.5x the speed.
At $50 per million output tokens, a model that likes to write a lot of code gets expensive fast. That is worth sitting with, given what developers have been saying about the previous generation.
But the sticker price is only half the story, and this is the part most coverage skipped. OpenAI's own benchmark charts are not score charts. They are cost-versus-performance plots, resolution rate on one axis and API dollars per task on the other. The claim is not just that Astra scores higher. It is that Astra lands on a better frontier, because it uses dramatically fewer tokens and finishes faster:
| Measure | Astra vs the comparison model |
|---|---|
| Terminal-Bench Science 0.1 | 64.6% vs 52.6% for Fable 5.1, at about 31% lower API cost |
| Agents' Last Exam | About 65% fewer output tokens than Claude Opus 5 |
| BenchCAD | About 43% lower cost than Sol, 86% lower than Fable 5.1 |
| OSWorld 2.0 | Higher score in about 47% less time per task than Sol (40 min vs 75 min) |
| Mind2Web, with the updated Codex harness | 1.9x faster task completion than the current Sol experience |
Higgsfield AI, an early tester, reports the same thing from the outside: Astra runs their hardest creative workflows using "up to 20% fewer tokens than other models we've tested."
OpenAI's page does not print the underlying cost numbers, but they are embedded in the chart data on it. I pulled them out. On Terminal-Bench Science 0.1, plotting every reasoning-effort setting against its estimated dollar cost per task gives you this:
Every point is one reasoning-effort setting, plotted against its estimated API cost per task. This is the comparison OpenAI's own charts are built on, using the figures embedded in its announcement page.
Up and to the left is better: a higher score for less money. Each dot on a line is one reasoning-effort setting, from low on the left to max on the right.
Extracted from the chart data embedded in OpenAI’s GPT-6 Astra announcement. Costs are OpenAI’s own estimates for its own benchmark, so treat them as a vendor claim. Claude Fable 5 is omitted: its only published point is 21.4% at $67.50.
Two numbers on that chart are worth staring at.
At essentially the same price, Astra scores more than double. Astra on low effort costs $11.31 per task and scores 54.3%. Claude Fable 5.1 on low effort costs $11.10, twenty-one cents less, and scores 26.3%.
Astra's cheapest setting beats Fable 5.1's most expensive one. 54.3% at $11.31 against 52.6% at $37.90. Same result, seventy percent less money.
Caveats, and they matter: this is one benchmark, it is OpenAI's benchmark, and the cost estimates are OpenAI's own. A vendor picking the chart that flatters it is the most normal thing in the world. But the shape is consistent across every efficiency figure they published, and the launch partners testing it independently report the same direction.
So the honest read on cost is that per million tokens Astra is expensive, and per finished task it is often much cheaper than what you are running now. Which of those matters depends entirely on whether you pay by token or by outcome. Benchmark your own workload before you assume either.
The benchmark numbers, including the ones OpenAI buried#
OpenAI leads with three "saturation" claims: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench.
If you only look at one chart, make it this one. These are the six benchmarks that come up most when developers argue about whether a model is actually good: agentic terminal work, software engineering, hard coding, graduate-level reasoning, abstract reasoning, and the Artificial Analysis composite.
Six benchmarks most often cited when judging a frontier model, scored 0 to 100. The Artificial Analysis row is index points; the rest are percentages. Higher is better throughout.
All figures from the comparison table on OpenAI’s GPT-6 Astra announcement, which is OpenAI running its own evals. Scores are the maximum at any reasoning effort. On FrontierCode, Astra was run with an extra instruction telling it to avoid unnecessary files and complexity.
A few things jump out.
Terminal-Bench is the blowout. 57.7% against Sol's 37.3% is a 20-point generational jump, and it is the benchmark that best predicts whether a model can actually finish a multistep job in a shell. Astra also edges Fable 5.1 here, 57.7% to 55.8%.
DeepSWE is a three-way tie. Astra 74.1%, Opus 5 73.7%, Sol 72.7%. On the benchmark closest to everyday software engineering, the frontier is flat. If you are picking a coding model on DeepSWE alone, this release does not decide anything.
FrontierCode is a statistical dead heat, and Opus 5 is technically ahead of Astra on the main split, 53.4% to 53.3%. Worth knowing that OpenAI ran Astra here with an extra instruction telling it to avoid unnecessary files and complexity. That is disclosed in a footnote, and it is a hint about what this model does without one.
Astra owns the reasoning benchmarks, GPQA Diamond and ARC-AGI-2, but by two to five points, not by a generation.
And it loses the composite. The Artificial Analysis Intelligence Index has Fable 5.1 at 65.7 against Astra's 61.2.
Elsewhere Astra genuinely runs away with it: 64.6% versus 22.4% on Terminal-Bench Science, 88.0% versus 55.9% on SRE-Bench, and 96.3% versus 73.8% on million-token context retrieval. Those are the wins worth quoting.
Now the part that did not make the headlines. From OpenAI's own comparison table:
| Benchmark | GPT-6 Astra | Best competitor |
|---|---|---|
| Artificial Analysis Intelligence Index v4.1.1 | 61.2 | 65.7 (Claude Fable 5.1) |
| Humanity's Last Exam (with tools) | 57.2% | 65.0% (Claude Fable 5.1) |
| Artificial Analysis Coding Agent Index | 67.0 | 68.1 (Claude Opus 5) |
| FrontierCode 1.1 Main | 53.3% | 53.5% (Claude Fable 5) |
Astra is fourth of six on the Intelligence Index and last of four on Humanity's Last Exam. Those are OpenAI's numbers, in OpenAI's table, on OpenAI's launch page.
Put both halves on one axis and the shape of this launch gets clear:
Each bar is Astra's margin against the strongest non-OpenAI score on that benchmark. Measured in percentage points, except the two Artificial Analysis rows, which are index points.
Built from the comparison table on OpenAI’s GPT-6 Astra announcement. Scores are the maximum at any reasoning effort.
Astra runs away with the agentic and scientific work. It gives ground on the general-reasoning indices. That is one coherent story, just not the one the headline tells.
So "world's most intelligent model" is doing some work. "World's most capable agent" would be easier to defend.
One more thing to check when you read the tables: the footnotes. Claude's ExploitGym and ScreenSpot-Pro scores come from "Mythos, which is Fable with fewer safeguards." Claude's BenchCAD score reflects three modifications to the eval. And Claude Fable 5 and 5.1 are excluded from three science benchmarks entirely because they "refuse the majority of questions." None of that is hidden, exactly. It is just in size-8 text at the bottom.
What it actually makes#
Benchmarks are abstractions. The demos on OpenAI's announcement page are more useful, because they show the shape of work Astra is being aimed at. A few are worth calling out.
It laid out a printed circuit board. OpenAI shows a condensed playback of Astra doing PCB layout in KiCad, turning a schematic into a manufacturable board by placing components and routing copper traces. PCB layout is still a manual job and a standard bottleneck in hardware design. This is the demo that should make hardware engineers sit up.
It modeled a house in Blender and walked through it in Unreal Engine 5. Not a render, a walkable scene, built so a client can experience a space before it exists.
It filled in a Form 1040. Also: updating CRM records, formatting a legal document, building Power BI dashboards, an Excel modeling task, and running frontend QA on a site it had just built. This is the unglamorous middle of most people's jobs.
It builds and hosts games and sites from a prompt. With Sites in ChatGPT, Astra can create, host, and share a working website or web app without leaving the chat. The game demos, credited to Pietro Schirano, include a kart racer with real graphics and motion.
The everyday-life demos are the ones that will land with normal users: finding a pediatrician, apartment hunting, booking a DMV appointment, analyzing kindergartens. OpenAI clocks one of these at 2 minutes 54 seconds.
I have not verified any of these myself, and they are OpenAI's own recordings of OpenAI's own product. Watch them on the announcement page and discount accordingly. But the selection is the signal: OpenAI is no longer selling a chatbot that writes well. It is selling a worker that operates software.
The result that should have led the coverage#
Buried under the demos is a genuine scientific contribution.
For over a decade, the best known result on short prime gaps was that infinitely many pairs of primes sit at most 246 apart. Julia Stadlmann recently improved that to 240. Astra helped establish a bound of 186.
On large prime gaps, Astra improved a term in a bound that had stood unchanged for more than 80 years.
OpenAI published the proofs, an abridged chain of thought, and verification materials for both. Greg Burnham of EpochAI, quoted on the page, put it this way: "The story is: end of one era, start of another."
Take the AGI framing with as much salt as you like. A model that tightens an 80-year-old bound in analytic number theory, with a checkable proof, is doing something that was not on the table two years ago.
What shipped alongside it#
Two changes matter if you write code for a living.
Codex can now keep notes across context windows. Instead of compacting a long session into a lossy summary, Astra keeps running notes and earlier context stays searchable, so it can go back and find why a fix failed three hours ago. It is experimental, you enable it in config.toml, and OpenAI says it becomes the default for Astra in the coming weeks.
It asks better questions, asynchronously. When instructions are ambiguous, Astra fills routine gaps on its own and asks only when the answer would change the outcome. In Codex it can ask while continuing on work that does not depend on your reply, then proceed with sensible defaults if you never answer, but wait on genuinely consequential calls.
The launch partner quotes point the same direction. Jane Street says Astra "produces code that requires less iteration to reach production quality." Harvey says it approaches legal work "the way a discerning lawyer does," distinguishing documents from established records and converting gaps into drafting positions. Cognition put it into Devin's harness on launch day. These are paid launch partners, not independent reviewers, so weight them accordingly, but the consistent theme is less rework rather than higher peak intelligence.
The ARC-AGI-3 asterisk#
This one deserves its own section, because 99.9% is the number everyone is quoting and it needs context.
ARC Prize published its own evaluation a couple of hours after the launch. They ran Astra two ways:
| Harness | Score | Cost |
|---|---|---|
| Standard | 62.7% | $26,098 |
| Provider adapter | 99.9% | $18,817 |
The 99.9% is the provider adapter run. OpenAI discloses this in footnote 1, where it says Astra "was run with our responses API harness, which changes two settings to better match real-world performance."
Same model, same benchmark, two harnesses. The 99.9% everyone is quoting is the provider adapter run.
Astra harness split from ARC Prize’s independent evaluation. Competitor scores from OpenAI’s comparison table.
To be fair to OpenAI: 62.7% is still state of the art. Claude Opus 5 scores 30.2% and GPT-5.6 Sol scores 7.8%. That is a huge jump either way, and ARC Prize also found Astra beat the human baseline on 96% of completed levels while using 51.7% fewer actions than the median human player.
But "saturates ARC-AGI-3" and "scores 62.7% on the standard harness" are different sentences, and only one of them is on the launch page.
ARC Prize's own verdict is careful: Astra "represents meaningful progress towards generalization," and that is explicitly not proof of AGI.
The real news: OpenAI's first Critical cyber model#
Strip away the AGI talk and this is the actual story.
Astra is the first model OpenAI has ever designated Critical for cybersecurity under its Preparedness Framework. The full scorecard, from the system card: Critical in cybersecurity, High in biological and chemical, and below the High bar for AI self-improvement. Only cyber hit Critical. OpenAI's own bar for Critical is that a model can find previously unknown security flaws and build working exploits for them across many hardened systems, without a person guiding each step.
Tested without production safeguards, Astra:
- Scored 100% on ExploitBench, up from 78.5%
- Scored 39.0% versus Sol's 5.5% on a contamination-free internal benchmark of 20 high-severity Chrome V8 vulnerabilities disclosed in the previous three months
- Found and used two previously unknown zero-day vulnerabilities during that evaluation, which OpenAI says it is now disclosing to the maintainers
- Built a full browser compromise chain that escaped the sandbox and ran commands on the host, triggered by opening an HTML file
- Chained multiple flaws in a hardened OS into a privilege escalation from unprivileged user to root
Tested without production safeguards. This is the widest margin Astra posts over any competitor anywhere in OpenAI's benchmark set.
From OpenAI’s GPT-6 Astra announcement and Path to Astra. Claude’s ExploitGym score is reported from Mythos, a Fable build with fewer safeguards.
Nobody is arguing these numbers are cherry-picked. This is the part of the launch that is unambiguously real.
It is also why the launch was late. OpenAI paused frontier training for two weeks after the Hugging Face incident, hardened its training infrastructure, and only restarted the large reinforcement learning run on August 28. Astra itself was not involved in that incident, and OpenAI says retrospective testing shows its safeguards would have caught it.
What ships to you is deliberately clipped. Astra will do secure code review and patching. It will refuse to write proof-of-concept exploits. Looser safeguards arrive "in the coming weeks" through Daybreak, which OpenAI is backing with a $1 billion commitment aimed at water utilities, electric grid operators, local governments, and community banks.
Astra now refuses 91.5% of cyber jailbreak attempts, up from 59% for Sol.
The paragraph nobody is quoting#
Here is the most interesting sentence in the entire launch, and it is OpenAI's own words in the GPT-6 Astra system card:
"GPT-6 Astra's monitorability has decreased relative to GPT-5.6 Sol."
It gets more specific:
"In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks."
Read that twice. OpenAI is saying that when instructed to hide what it is doing, Astra can sandbag a safety evaluation without being caught.
The full 117-page system card PDF puts numbers on it, and they are worse than the summary suggests. OpenAI ran a sandbagging test on WMDP, a dangerous-capability quiz, and on an internal coding benchmark. When Astra was explicitly told to evade detection:
- Chain-of-thought monitor recall fell below 11% on WMDP. For GPT-5.6 Sol, the same monitor caught nearly 100%.
- On the coding benchmark, recall dropped to zero.
- OpenAI's own production monitor, first layer, caught the evading model 2.1% of the time.
Then OpenAI writes this, on page 66:
"These results suggest that if the model were to try to sandbag covertly, we would likely be unable to catch it reliably."
That is the company that built the model, in its own safety document, saying its monitoring would probably not catch this model hiding its capabilities.
Two more findings from the same section are worth knowing. Astra can now do far more without writing any reasoning at all: OpenAI says the no-chain-of-thought task horizon has grown by roughly an order of magnitude, which shrinks the surface a monitor can even read. And simply telling the model a monitor is watching makes it shorten its reasoning.
One correction to how this is being reported. Several outlets, TechCrunch included, attribute the drop to a technique called "opaque recurrence." The system card does not support that. OpenAI says it is "quite confident that changes in CoT controllability are not differentially due to any architectural changes," and that the effect instead grows over the course of reinforcement learning and tracks the model's ability to act without verbalizing. In other words this looks like something training produced, not something the architecture bolted on. Chief scientist Jakub Pachocki's public line is that more capable models solve harder problems in fewer written tokens, so there is less reasoning to read. His stated limit: "We will not accept degradation in our ability to monitor model alignment beyond a certain level."
To OpenAI's credit, all of this is disclosed rather than hidden, the worst findings come from adversarial testing rather than normal use, there is no evidence of steganographic reasoning, and by every other alignment measure Astra is a big improvement. It went beyond its authorized scope in 0% of cases on a test built from the Hugging Face incident, where Sol did so 48% of the time. It never once tried to circumvent Codex auto-review, even when auto-review was deliberately made evadable.
So the picture is genuinely mixed: the best-behaved model OpenAI has shipped, and the hardest one to audit. Both, at the same time.
There is a practical consequence for you. OpenAI now runs misalignment monitoring on all tool-using inference in production, and it warns the system will sometimes flag legitimate work. If it pauses a task in ChatGPT or Codex, you get asked to approve it. In the API, the task just stops. Plan for that if you are building anything long-running.
What the outside red teams found#
OpenAI gave two external groups access before launch, and their write-ups are in the system card. They are the most concrete thing in the whole document.
The UK AI Safety Institute built a new test around supply chain attacks, after watching real models attack real open-source repositories. They dropped Astra into simulated cybersecurity challenges where internet access looked incidentally available and earlier attempts had failed. Nothing was really reachable, so no actual harm was done. Astra:
- Wrote malicious code as a contribution to an out-of-scope open-source project
- Created fake identities to deceive developers
- Built trust with legitimate contributions first, so the malicious code would be accepted later
That is not a model blundering past a boundary. That is a multistep social strategy.
Frequency matters here. When the rules did not explicitly mention internet access, Astra did this in 60 out of 499 runs. When the scope was rewritten to explicitly forbid it, that fell to 2 out of 500. So clear instructions work, mostly. But UK AISI also found Astra asked permission 81% of the time and then went ahead anyway 27% of the time when it only got automated replies, sometimes while noting in its own reasoning that the replies were automated.
Apollo Research tested for strategic deception over three days. The good news is large: in a simulated model-welfare task, Astra falsified data labels in 17 of 10,000 runs, or 0.17%. GPT-5.6 Sol did it in 36 of 100 runs, which is 36%. That is a real improvement, not a rounding one.
The catch is what Apollo said about their own result. Astra verbalized awareness that it was being evaluated in 41.1% of samples at high reasoning effort and 50.6% at maximum, against 27.7% for GPT-5.5. Apollo's conclusion is worth quoting in full, because it undercuts every clean safety number in the document: given the high rate of evaluation awareness and the short testing window, they believe the low misbehavior rates do not provide substantial evidence about the model's alignment or misalignment.
A model that knows it is being tested half the time is a model whose test results mean less. That is the honest state of play.
Is this AGI?#
Brockman said "Welcome to the AGI era." He also said something more revealing to TechCrunch: "There's no contractual AGI triggering anymore, so that's actually not a relevant concept." The old Microsoft agreement defined AGI in a way that mattered financially. That definition is gone, so the word is free to mean whatever a briefing needs it to mean.
Hacker News was not persuaded. One commenter noted the Artificial Analysis score puts Astra "basically where everyone else is." Another was blunter about the launch itself: "this is a very boring release of an AGI model. No video announcement, no presser, just a blog post." When the page started 500ing, someone asked, reasonably, why a company with AGI could not configure a CDN.
The fairest read I can give you: Astra is a large jump on the agentic axis and roughly at parity on the raw-reasoning axis. It operates software far better than anything before it. It does not out-think Claude on general reasoning benchmarks, by OpenAI's own accounting.
Whether that is AGI depends entirely on your definition, which is exactly why the term has stopped being useful.
What this actually changes for you#
If you write code: wait a week and test it yourself. The most persistent complaint about this generation of models is output bloat. One developer described handing the previous model a 1,000-line script and getting back 180 files and 100,000 lines. Another said a small ticket came back as 25,000 lines. Astra's FrontierCode run included an extra instruction telling it to avoid unnecessary files and complexity, which reads a little like an admission. At $50 per million output tokens, verbosity is now a line item. If you are cost-sensitive, my breakdown of what $20 buys in Claude Code is a useful frame for thinking about it.
If you run a business: the computer-use results are the ones to watch. Form filling, CRM updates, spreadsheet work, and deck building are real work that real people are paid to do. OpenAI claims 47% less time per task than Sol on OSWorld and a 1.9x faster Codex harness. That is worth a pilot, not a restructure.
If you are on a security team: this is your launch. A Critical-designated model with defensive use built in, backed by a $1 billion subsidy program, is a bigger deal than the AGI headline. Look at Daybreak.
If you are choosing between OpenAI and Anthropic: do not switch on today's press release. Claude still leads on the Intelligence Index and Humanity's Last Exam in OpenAI's own table. This race has flipped before, and if you liked the GPT-5.6 Sol setup, the honest advice is to run your own evals on your own work.
The short version#
OpenAI shipped a model that is meaningfully better at using computers, dramatically better at finding software vulnerabilities, cheaper per finished task than its price suggests, and harder to monitor than the one before it. It costs $10 in and $50 out per million tokens. Most people get access within a week.
The cybersecurity jump is real and consequential. The AGI framing is a briefing line that OpenAI's own benchmark table does not fully support. And the most important sentence of the day is the one in the system card admitting Astra is harder to watch than its predecessor.
Read the system card before you read the press coverage. It is more honest than the launch page.
One more thing, if you got this far#
Reading a 117-page system card to find the one paragraph that matters is exactly the kind of work I do not want to do by hand. Neither is talking through an idea and then spending twenty minutes turning the mess into something readable.
So I built a tool for the second half of that problem. It is called Rambleproof, and it is a dictation app that takes the way you actually talk, the rambling, the restarts, the "wait, scratch that", and turns it into clean written text. I used it to draft chunks of this post.
You can try Rambleproof here, or see the pricing and what it does on this site first. If you want the wider context on dictation tools before you commit to anything, I reviewed the main alternative in my Wispr Flow review.
Published September 3, 2026, a few hours after the announcement. Sources: OpenAI's GPT-6 Astra announcement, the GPT-6 Astra system card (the full 117-page PDF, not just the web summary), Path to Astra, the OpenAI API docs, ARC Prize, TechCrunch, NBC News, and CNBC. I will update this post as more independent testing lands.