GPT-6 Astra landed over the weekend, and first impressions are all over the map: benchmark scores that initially tied its predecessor, jaw-dropping one-shot 3D games flooding X, and a computer-use capability that has power users going hands-free. The confusion isn't a bug — Astra is an opportunity model, not an efficiency model, and we don't have the tests or the habits to judge it yet.
Astra is an opportunity AI model, not an efficiency AI model — and that's why nobody can agree on it.
GPT-6 Astra is not about doing what you currently do better; it's about expanding what you can do. That's why the weekend takes felt so discordant: people judged a new thing with old criteria that didn't fit. Its 3D and computer-use capabilities follow the same pattern as image generation and AI coding before it — capability sets handed to people who never had access to them — and it carries a new default interaction pattern, where instead of clicking and typing we ambiently talk to a computer that works for us. Changes that immense don't package up in a weekend of testing.
— Sam Altman, announcing GPT-6 Astra. Astra started as a fundamentally different training run — a genuine answer to Anthropic's Mythos class, not a 5.5-to-5.6 style bump — and it belongs to the new generation of models that now get delayed before release on safety and security grounds. Altman's promise: worth the wait.
The Thursday announcement limited availability to partners in the cybersecurity-focused Daybreak program, leaving paid subscribers antsy. OpenAI's Tibo promised subscribers a banked reset for the wait, and by Friday night the model was in everyone's hands.
The three-minute announcement — 132 million views and over 100,000 saves by Tuesday — shows people sitting or pacing, talking to a laptop while Astra lists items on eBay, reviews contracts, and builds games, with food orders and tennis-court bookings running in the background. The demo isn't a feature list; it's a new way of working.
Dan Shipper called Astra the best writing model he's tried — easy to steer, very little slop — and incredibly good at computer use, able to "go for hours at a time using complicated apps to get work done." But it overcomplicates: ask for a simple interface and you get extra labels, buttons, and features, without Fable's knack for doing something delightful unprompted.
Artificial Analysis's intelligence index put Astra dead even with GPT-5.6 Sol, five points behind Fable 5.1, and one point behind Meta Mu Spark. The index skews toward fact memorization, with few tests for advanced coding and fewer still for advanced computer use — evidence that our old tests don't reflect what this model will actually do.
Version 4.2 adds more emphasis on agentic tasks via the AA briefcase test and reweights existing tests. Under the new scoring, Astra is still behind Fable 5.1 — but ahead of everything else. When the benchmark maintainers rebuild the benchmark days after a launch, the launch is telling you something.
To not bury the lede: GPT-6 Astra is not about doing what you currently do better — it's about expanding what you can do. That makes it extremely interesting and potentially extremely valuable, but also challenging: new capability sets don't come with established use cases attached.
On TerminalBench 4.0 Astra scored 57.6% vs Fable 5.1's 55.8% (and GPT-5.6 Sol's 37.3%); on DeepSui, 74.1% vs 73.7%. OpenAI now charts performance against cost, and pointedly highlighted that Astra hit those numbers a lot more cheaply than Fable 5.1.
On both TerminalBench and DeepSui, Astra scored best on high or extra effort settings and actually declined a little with effort set to max — a hint that cranking the dial can send the model down sidetracks instead of toward answers.
On Automation Bench, Astra scored 41.1% — far above Fable 5.1's 31.4% and GPT-5.6 Sol's 18.1%. The announcement post calls it explicitly the world's best computer use model, marking "a new frontier in the speed, accuracy, and safety of computer use."
Astra scored 64.6% on Terminal Bench science, beating Fable 5.1's 52.6% and demolishing GPT-5.6 Sol's 22.4%. On Frontier Math Tier 4 it hit a near-perfect 97.6%, up from Fable's 90.2%.
Astra scored 100% on Exploit Bench across all effort levels, suggesting it can build and execute exploits with little difficulty, and hit 39% on an internal benchmark of recently disclosed vulnerabilities versus 5.5% for GPT-5.6 Sol. Suddenly the cybersecurity-first Daybreak rollout makes a lot more sense.
Photorealistic bats built until the tokens ran out, a Zillow listing turned into a one-shot 3D house tour and promo video, a Tesla Model X exploded into 334 modeled website pieces, and a creepy VHS-style Backrooms — including one-prompt character rigging, which one user noted has never really worked in prior GPT versions. OpenAI's own post pitches use cases like game development, circuit board printing, and car transmission design.
Ethan Mollick had Astra turn the 1977 text adventure Zork into a full 3D action game in Three.js, keeping the original plot and puzzles. Derya Unutmaz — with no software engineering expertise — rebuilt a Doom-inspired game in hours, marveling that something that once required a legendary team can now be recreated and expanded by anyone. Even Altman weighed in: trivial relative to everything else, but making a fun little game and playing it minutes later "is so cool."
Theo's browser-based Fish Slop would have cost under $30 on a $200 subscription. The AI battle account built Sonic in 53 minutes on max settings using 4% of a Pro X5 account's weekly usage — or 25 minutes and 1% on medium. The turbo-AGI-machine-god demos are cheaper than they look.
Users had Astra build an interactive 3D model showing how Ozempic works, a full interactive ankle atlas with real motion axes and live ligament readouts to understand their own pain, a real-time soccer overlay tracking players and possession, and a tool that turns any image or idea into a buildable Lego set using official parts. The 3D capability isn't just games — it's a new medium for understanding.
Ethan Mollick's sharp observation: Astra being incredibly good at Blender-style visual work gives it a perception edge over Fable in the fight for social media attention, whether or not that was the goal — other kinds of outputs are harder to judge, but the visual stuff comes through. Cool, yes; but if you knew 3D design was Astra's superpower, would you even have a use for it?
A16Z's Martin Casado saw a meaningful step in computer use but not in coding for his work. Others reported weird Python slop and horrific unit tests when one step removed from normal code, shader work that blew them away next to front-end design that "crapped the bed" — with several saying Claude's noticeably better UI work might be the biggest reason to keep using it.
Claire Vo of the How I AI podcast had spent six months trying to build an architecturally complex product intelligence app: Fable did insane things with the architecture, GPT-5.6 got stuck on quality of insights — and Astra one-shotted it. The counterpoint to the saturation takes: on the hardest specific tasks, something real changed.
The truly transformative thread in the fuller reviews is computer use. Claire Vo, a self-described computer-use maxer, says Astra now navigates complex web UIs and CRM lead-routing workflows she never trusted to earlier models. Ali K. Miller's challenge captures it: imagine you recorded your screen for seven straight days — what would you hand off? Can you go mouse-free for a day?
Image generation was the first capability handed to people who'd never had it; AI coding — which turned non-coders into vibe coders after Opus 4.5 and GPT-5.2 landed in late 2025 — was the second. Astra's 3D modeling and design feels like the third. The open question: will it stay novelty, or make the jump agentic coding made into a thing broad swaths of knowledge workers use regularly?
NanoBanana wasn't notable for better images — it let you fix one part instead of regenerating the whole thing, and that single change unlocked enormous use cases. This summer's coding watchword has been loops: setting up structures where agents evaluate their own progress toward a goal rather than being prompted task by task. Astra brings that same kind of pattern shift to computers generally.
It's no accident the launch video is 130 million views of people completely hands-free, verbalizing their way through work while AI manages the interface. Voice mode has been the baby step; Astra's computer use makes ambient, spoken interaction the argued-for default. Don't expect to fully understand or take anywhere near full advantage of this model in the immediate term — the real work of the next few months is figuring out what new opportunities it unlocks, not where it swaps in for GPT-5.6 or Fable on everyday tasks.
— Tibo, OpenAI product lead. Astra was "probably our biggest competitive advantage" while it wasn't generally available, per OpenAI's product lead — with plans now shipping at Dev Day instead of mid-next-year. The question worth watching: was that just better agentic coding, or something else entirely?