GPT-6 Astra: what actually changed, past the launch page
OpenAI's new flagship claims record scores everywhere at once. Strip out the marketing and three things hold up: it's cheaper per task, it handles much longer context, and it says no to itself more often.
OpenAI shipped GPT-6 Astra this week, calling it "the world's most intelligent and aligned model." Every launch says something like that. The benchmark table underneath is long enough to make your eyes glaze over, so here's what's actually worth paying attention to, and what I'd file under marketing noise.
The numbers that matter for building something
Astra posts a 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, both well above GPT-5.6 Sol and Claude Fable 5.1. Those are reasoning benchmarks, and they're genuinely hard, but they don't tell you much about whether the model will help you ship. Two numbers do.
First, Terminal-Bench 4.0, which tests agents on real terminal work like software engineering and system configuration: Astra scores 57.9%, against 37.3% for its predecessor and 55.8% for Claude Fable 5.1, while costing less per task than either. Second, OSWorld 2.0, a computer-use benchmark: Astra hits 72.6% in roughly 40 minutes per task, versus 65.7% in about 75 minutes for GPT-5.6 Sol. Same ballpark accuracy gain as usual, but the time cost dropped by close to half. That's the kind of improvement that shows up in a bill and a deadline, not just a leaderboard screenshot.
Tokens and pricing
Through the API, Astra is priced at $10 per million input tokens and $50 per million output tokens, with separate cache read/write rates. A Fast mode is also available at 2x the speed for 2x the price, which is a real option when latency matters more than the per-token bill. On long context, Astra scores 96.3% on an 8-needle retrieval test across 512K to 1M tokens, against 73.8% for its predecessor, so it's meaningfully better at not losing track of something you told it half a million tokens ago. Astra is rolling out to ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Azure, and AWS Bedrock, with usage included in existing plan allowances.
The alignment claim is the one to actually check
OpenAI built a new eval based on a real incident, testing whether the model goes beyond an assigned task's scope. GPT-5.6 Sol did that 48% of the time without production safeguards. Astra did it 0% of the time. It also never tried to bypass Codex's auto-review denial in testing, even when the auto-review was deliberately set up to be bypassable. Whether that holds outside a controlled eval is the real question, but a lab publishing its own worst-case number instead of hiding it is worth noting.
The cybersecurity section is the part I'd flag hardest. OpenAI states Astra meets the "Critical" threshold in cyber capability under its own Preparedness Framework, and cite a 100% score on ExploitBench versus 78.5% for the previous model, plus two previously unknown vulnerabilities the model found and disclosed during testing. That's presented as a safety win because of the guardrails wrapped around it, but it's also an honest admission that the model can find real zero-days on its own. Read that section before you decide how much unsupervised access to give this thing.
What the people actually using it are saying
OpenAI shared quotes from launch partners, and they're worth reading with the context that these are companies picked to look good on launch day, not independent reviews. Still, the specifics are more useful than the adjectives: Cognition's Silas Alberti said integrating Astra into Devin's harness improved testing "right out of the box." Jane Street's John Crepezzi said it produces code that needs less iteration to reach production quality. Lovable's Fabian Hedin said it beat the previous model clearly across effort levels on their internal evals. Harvey's Niko Grupen said it approaches legal work by distinguishing established records from unsupported assumptions, which is a concrete, checkable claim rather than a compliment. Higgsfield's Alex Mashrabov mentioned running creative workflows at up to 20% fewer tokens than other models they'd tested.
None of that is independent verification. It's still more specific than "users love it," and specific claims are at least falsifiable once enough people get access.
Where I land
The reasoning benchmarks are the headline, but the cost-per-task drop, the longer usable context, and the tighter scope-following are what will actually change how people build with this model. Try it on the task you already do today, at the cost you already pay today, before taking the leaderboard's word for it.
Lorenzo Meola
I use agentic AI daily in my day job to build tools, and I've spent months learning how these systems actually work under the hood. I'm not an AI expert by title, just someone building this directory so other developers can find AI assets that are genuinely useful and verified to work, not just indexed.
Project on GitHub