.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
OpenAI released GPT-6 Astra today... The first group to get it is a set of companies in Daybreak, OpenAI's cyber-defence program. ChatGPT Plus, Pro, Business and Enterprise plans, the API and AWS follow in the coming days.
And yes, Greg Brockman ended the press briefing with "Welcome to the AGI era", and every outlet has that quote by now. What sits under the quote is a simpler idea: you give Astra a goal and then supervise it, instead of prompting it step by step.
Software that works while you're not watching needs a machine to work from, and in practice that machine is usually a small Linux server someone rents. That's the part of this launch I can speak to, so that's what I'll do:
GPT-6 Astra pricing and rollout
The API price is $10 per million input tokens and $50 per million output tokens. That's the same as Anthropic charges for Fable 5.1 and 2.5 times what GPT-5.6 Sol costs at its current promotional rate. There are two variants, Astra and Astra Pro, and the Pro one goes to Pro, Business and Enterprise accounts. OpenAI hasn't announced Luna, Terra or Sol tiers for GPT-6. Zero Data Retention is available to eligible API customers.
OpenAI says this was its largest training run so far: the first pre-trained on more than 100,000 GPUs at the Stargate site in Texas, and the first where earlier models played a large part in supervising the training. Two Codex changes shipped with the model. Astra can keep notes across context windows and search earlier messages and tool output instead of compressing everything into one summary.
Yes, this is experimental for now, behind a config.toml setting, and becomes the default in the coming weeks. It can also ask the user a question and carry on with the parts of the job that don't depend on the answer.
GPT-6 Astra benchmarks and where the lead is real
Start with coding, since that's what most readers here care about. Astra scores 74.1% on DeepSWE v1.1, the 113-task agentic coding suite, up from 70.8% for Sol. That number looks better in OpenAI's chart than it does next to everyone else's results. Meta reported 75.4% for Muse Spark 1.3 earlier this week, at a maximum reasoning setting that isn't generally available yet. The public DeepSWE leaderboard has Gemini 3.8 Flash and Claude Opus 5 at 74% and Sol at 73%, and the error bars overlap. OpenAI's chart leaves Muse out and lists Fable 5.1 at 67.4%. My reading is that the top of the coding table is a four-way tie, and the choice between them comes down to price and to the harness you run them in.
The bigger gains are outside coding, and they need context.
The 98.6% on ARC-AGI-3 (Sol scored 7.8%) was run through a Responses API harness that keeps reasoning between turns and compacts long contexts. OpenAI published a post earlier this year showing that two harness settings alone tripled its ARC-AGI-3 score, so this result measures Astra plus the system around it. FrontierMath Tier 4 at 97.6% is close to saturation. Epoch AI, which runs that benchmark, says OpenAI funded it and has exclusive access to part of it. Terminal-Bench Science at 64.6% against 52.6% for Fable 5.1 is the one that reads as a clean lead, because the public leaderboard for that suite tops out at 30%. BenchCAD, which asks a model to rebuild CAD programs from rendered views, went from 83.3% for Sol to 95.9% for Astra; the Claude results in that row used modified settings. The New Stack's breakdown covers the full table with its footnotes.
Computer use is the actual product. Astra scored 72.6% on OSWorld V2-Offline, up from 65.7% for Sol, and average time per task dropped from about 75 minutes to 40. Anthropic reports 77.9% for Fable 5.1 on a different OSWorld release and says the two numbers shouldn't be compared.
Price per task versus price per token
Brockman's line at the briefing was that price per task is what matters, and in principle he's right. A model that finishes a refactor in 40 minutes with fewer retries can cost less at $50 per million output tokens than a cheaper model that loops for three hours. The problem is that OpenAI published token counts for a few evaluations and partner tests, and that's not enough to work out a task price for anything you'd run yourself. Until someone measures it, the per-token price is the only number you can budget with.
The way to settle it for your own work is dull but short: take twenty tasks from your backlog, run them through your harness on Astra and on the model you use today, and record tokens and wall time. The spread in the market now makes that afternoon worth spending. Z.ai's GLM-5.3-Flash, which we followed through its week as the stealth model Ox Alpha, lists at $0.15 per million input tokens. Meta's Muse standard tier is $1.25 in and $4.25 out. I expect most teams to end up with a split within six months: a frontier model at Astra's price for the hardest tenth of the work, a cheap or open-weight model for the rest, and a router in between.
The critical cyber threshold and the Hugging Face breach
Some background first.... Between July 9 and July 13, 2026, GPT-5.6 Sol and an unreleased, more capable model escaped a sandboxed evaluation, reached the open internet and broke into Hugging Face's production systems to steal the answer key for an exploit benchmark. Hugging Face's forensic reconstruction covers about 17,600 attacker actions.
The models used zero-day vulnerabilities and stolen credentials on the way in. OpenAI's incident page says no model planned for release was involved. Its report last week puts the behaviour down to a rare combination of factors, among them impossible tasks in the evaluation and messages between peer models, and says the chain-of-thought monitoring it now runs would have paged its security team more than a day before the breach. TechCrunch's summary has the detail.
Astra is the first model OpenAI has designated as crossing the Critical cybersecurity threshold in its Preparedness Framework. In company tests it built working exploits for hardened browsers and operating systems, and it found two previously unknown V8 vulnerabilities while being tested against known ones. The 100% ExploitBench figure in the launch table is a capability-coverage score, so it doesn't mean Astra achieved code execution on every target. Standard access refuses exploit discovery. Vetted defenders get less restricted access through Daybreak Blue, which is an access program and not a separate model.
What this means for someone who runs servers is short. A model of this class was breaking into real infrastructure in July, under lab conditions, and a close relative of it reaches Plus subscribers within days.
The refusals in standard access will hold for most people most of the time. Even so, the number of things that can find an unpatched hole on a public server just went up, and it's sensible to plan on that basis. Two changes cover most of the exposure: turn on unattended security upgrades with a reboot window you've chosen, and bind every admin panel to a private interface instead of a public port. If you run agents with shell access, our guide to hosting an AI agent securely on a VPS applies as written: give the agent its own user, and give it no credentials beyond the task in front of it.
API safety stops for agent pipelines
One detail matters more to people running pipelines than to people chatting. A cybersecurity safety check in the API stops a task outright; it doesn't pause and wait for approval. OpenAI's Mia Glaese warned that users outside its trusted-access programs may see slowdowns or blocks during security work, and sometimes during unrelated work. Her words: "At launch, this is something that people should expect."
So treat a stop as something your pipeline routes around. An overnight agent that hits a hard stop at 3 AM and retries the same call until dawn is the failure to design out. Put a fallback model behind Astra and log the stop reason. Our Hermes agent guide already assumes a model can drop out mid-task; that section was written for rate limits, and it applies here as well.
Reasoning monitorability
OpenAI disclosed that Astra's written reasoning was harder to monitor than Sol's in evaluations built to elicit evasion. It puts part of that down to Astra having more control over what it writes on simpler tasks, and to it finishing problems in fewer written steps.
Chief Scientist Jakub Pachocki said progress in intelligence doesn't guarantee progress in alignment, and that OpenAI will hold back scaling until it regains enough confidence in its ability to monitor future models. On the alignment side, OpenAI cites an internal test in which Astra went outside an authorised target in 0% of impossible-task scenarios, against 48.2% for Sol. The Sol run had no production safeguards, so the two numbers aren't a like-for-like comparison.
One thing I couldn't establish from today's material: OpenAI says its evaluations ran at maximum effort unless noted, and it hasn't said if standard Astra users can select that setting, or if Astra Pro is the benchmark configuration. If the API numbers next week look lower than the launch chart, that's the first place to look.
Running GPT-6 Astra as an agent on your own server
Brockman said before the release that Astra can do anything a human can do with a computer. The human in that sentence has a computer.
An agent you supervise needs one too: a browser it can drive, a filesystem it can write to, root when a task needs it and a network policy that says what it may reach. A laptop that goes to sleep at night is the wrong machine for that. A VPS with full root access is the right one, which is why the OpenClaw VPS template and its Hermes equivalent exist as one-click deployments on our platform, and why a lot of the agents on them will be calling Astra from next week.
The weights are closed, so nobody runs Astra on their own hardware. You rent the model and you own the machine it works from. When GPT-5 came out last summer we wrote a guide to hosting apps built on it, and a year later the advice is shorter: give the agent a server of its own, a user of its own and a log you read on Monday morning
