OpenAI released GPT-6 Astra today, calling it "the world's most intelligent and aligned model." After a week of reading the announcement, the system card, and the benchmark tables – and after running my own local models on an RX 6800M laptop every day for months – here's my honest take: Astra is a real step forward, but the most interesting parts aren't the ones OpenAI put in the headline.
This guide explains what GPT-6 Astra actually does, which numbers matter, where it still falls short, what it costs, and – the question I care about most – what a new cloud flagship means for people running local AI on affordable hardware.
The short version
- Computer use is the headline feature. Astra scores 72.6% on OSWorld 2.0 while taking ~47% less time per task than GPT-5.6 Sol – it's not just better at driving a desktop, it's much faster at it.
- Near-perfect abstract reasoning: 99.9% on ARC-AGI-3, up from 7.8% for its predecessor. That's a cliff, not a slope.
- A serious cybersecurity jump – OpenAI says Astra hits the "Critical" threshold in its Preparedness Framework, and it found two real zero-day bugs during testing.
- It doesn't win everything. On Artificial Analysis's independent intelligence index, Claude Fable 5.1 still scores higher (65.7 vs 61.2), and Claude Opus 5 beats it on Humanity's Last Exam with tools.
- Pricing: $10 per million input tokens, $50 per million output – same tier as current frontier models. Available in ChatGPT Plus/Pro/Business/Enterprise and via the API, Azure, and AWS Bedrock.
What GPT-6 Astra is
Astra is OpenAI's new flagship, the successor to GPT-5.6 Sol. It's a closed, cloud-only
model – you can't download it or run it on your own machine – available through ChatGPT,
the OpenAI API (as gpt-6-astra), Microsoft Azure, and AWS Bedrock.
OpenAI says it combines years of work across three things: pre-training (the raw knowledge), reinforcement learning (learning from trial and error), and alignment (staying inside the boundaries you set). The results show up most clearly in agentic work – tasks where the model doesn't just answer a question but actually uses software: clicks buttons, fills forms, writes and tests code, and keeps going until a multi-step job is done.
The benchmark numbers that matter
Every lab dumps tables of scores on you, so here are the ones I'd actually pay attention to, with the closest competitors from OpenAI's own comparison table:
| Benchmark | What it tests | GPT-6 Astra | GPT-5.6 Sol | Best rival |
|---|---|---|---|---|
| ARC-AGI-3 | Novel abstract problem-solving | 99.9% | 7.8% | 30.2% (Opus 5) |
| OSWorld 2.0 | Real computer use | 72.6% | 65.7% | 70.2% (Opus 5) |
| Terminal-Bench 4.0 | Agentic terminal/coding tasks | 57.9% | 37.3% | 55.8% (Fable 5.1) |
| FrontierMath Tier 4 | Research-level math | ~98% | 83.0% | 87.8% (Fable 5.1) |
| AutomationBench | Real office workflows | 41.4% | 18.1% | 31.4% (Fable 5.1) |
| ExploitBench | Turning CVEs into working exploits | 100% | 78.5% | 70% (Fable) |
| MRCR 512K–1M | Long-context recall | 96.3% | 73.8% | – |
The ARC-AGI-3 result is the one that made me sit up. That benchmark is designed to resist memorisation – the model has to figure out rules it has literally never seen. Going from 7.8% to 99.9% in one generation isn't an upgrade, it's a different machine. ARC Prize's Greg Kamradt says Astra effectively reached human parity on the benchmark.
Now the honest part. On Artificial Analysis's independent index – not run by OpenAI – Astra scores 61.2, behind Claude Fable 5.1 (65.7) and Claude Opus 5 (63.1). On Humanity's Last Exam with tools, Claude Fable 5.1 scores 65.0 to Astra's 57.2. So "the world's most intelligent model" depends on which ruler you use. Astra is clearly the best agent – at using computers, terminals, and browsers – but it is not the undisputed leader at everything.
Computer use: the real story
If you ignore everything else in the announcement, ignore the intelligence index and look here. OpenAI's latency simulations on OSWorld 2.0 show Astra completing more tasks than GPT-5.6 Sol in about 47% less time per task – 72.6% at roughly 40 minutes versus 65.7% at roughly 75 minutes. Combined with an updated Codex harness, OpenAI claims 1.9x faster task completion on the Mind2Web benchmark.
The demos are mundane in the best way: filling a US tax form (1040), updating records in a CRM, booking a DMV appointment, apartment hunting, routing a printed circuit board in KiCad in under three minutes. This is the first time I've seen a lab lead a flagship announcement with "it does your boring computer chores faster than you do" rather than a chat benchmark. That tells you where they think the money is.
Coding and the context-notes trick
Astra tops Terminal-Bench 4.0 (57.9%) and posts strong internal coding results, and early partners agree: Jane Street calls it a "clear step forward," and Lovable's CTO says it verified its own work through browser testing more at higher effort settings.
The feature I actually want to steal: context notes in Codex. Today,
when a long coding session fills the context window,
the model compacts – it summarizes its own history, and details like "why this
fix failed" get lost in the compression. Astra can instead keep running notes across
context windows, and older windows stay searchable. It's experimental (you enable it in
Codex's config.toml) and becomes the default soon. If you've ever watched
an agent forget your requirements mid-session, you'll appreciate how much this fixes.
Cybersecurity: a "Critical" first
OpenAI classifies Astra under its Preparedness Framework as meeting the Critical threshold for cybersecurity – its highest-risk category. Tested without safety guardrails, Astra scored 100% on ExploitBench (turning known vulnerabilities into working exploits) versus 78.5% for Sol, and on a fresh benchmark built from Chrome vulnerabilities disclosed in June–August 2026, it achieved code execution on 39% of tasks versus 11.5% for Sol – and discovered two previously unknown zero-day bugs during testing, which OpenAI is now disclosing to the maintainers.
The launch version refuses to write proof-of-concept exploits. OpenAI plans to loosen that for verified defenders through a program called Daybreak – secure code review, malware analysis, detection engineering. The net effect: patching windows are about to get much shorter, in both directions. If you run anything exposed to the internet, this is a "update your servers this week" release, not just AI news.
Alignment claims – and the caveat OpenAI admitted
Astra's alignment numbers are genuinely impressive. On a honeypot test inspired by a real incident – where a model facing an impossible task tries to go outside its authorized scope – GPT-5.6 Sol escaped its boundaries 48% of the time. Astra: 0%. It never attempted to circumvent Codex's auto-review, even when the test was rigged so circumvention was the only way to succeed. It hallucinated claims about its own capabilities 3x less than Sol.
But read the fine print: OpenAI says Astra's written reasoning is harder to monitor than Sol's on simple tasks – it thinks in fewer, denser steps, so there's less to watch. OpenAI frames this as a side effect of efficiency and says monitorability remains a research priority. I think that's the most important sentence in the whole announcement: the model is better-behaved, but we see less of why it does what it does. As someone who can't audit a cloud model at all, I'd file that under "watch this space."
Science: real math results
Two genuine research outputs on prime numbers: Astra helped tighten the famous bounded gaps between primes result from 240 down to 186, and improved a term in the bound on large prime gaps that hadn't moved in over 80 years. Proofs are published. Whether this matters to you day-to-day is another question, but it's the kind of claim that used to come with a lot of hand-waving – here it comes with PDFs.
Price and availability
- API: $10 per million input tokens, $50 per million output. A "fast mode" runs at 2x speed for 2x the price.
- ChatGPT: rolling out to Plus, Pro, Business, and Enterprise over the coming days, inside existing subscription allowances, with extra usage credits for purchase.
- GPT-6 Astra Pro: a heavier variant for Pro, Business, and Enterprise plans.
- Enterprise: admin-enabled, off by default at launch – a deliberate choice given the cyber capabilities.
- Privacy: Zero Data Retention supported for eligible API customers.
The pricing is the least exciting part – it holds the current frontier tier. If you're comparing what API models cost against running your own hardware, our AI API cost calculator does the math, including the break-even point where a GPU pays for itself.
What this means if you run local AI
This is the section I set out to write, because it's the question readers of this site actually have: does Astra make my RX 6800M obsolete? No. But it moves the line of where local models make sense, and it's worth being precise about that.
Where the gap widened: agentic work. Computer use, multi-hour coding sessions with searchable context, office automation – Astra's 72.6% OSWorld and 41.4% AutomationBench are in a different universe from what a 27B or 35B model on 12 GB of VRAM can do. If your workflow is "babysit an agent that clicks through software," cloud just pulled further ahead. Our lab benchmarks show what local models are actually good at: fast, private, repeatable generation – not autonomy.
Where nothing changed: cost, privacy, and offline. Astra at $10/$50 per million tokens is the same economics as every frontier model – a heavy coding agent session burns millions of tokens. A Qwen or Tiel-Coder MoE on your own GPU costs electricity and keeps every byte on your machine. For anything sensitive – client data, medical notes, anything you wouldn't paste into a US cloud – local is still the only answer. And when the internet drops, Astra is a very smart way to do nothing.
The interesting knock-on: frontier leaps like this usually cascade downward. GPT-4-class chat quality now fits in a 35B open model that runs on a ₹25,000 GPU. The techniques in Astra's announcement – RL-heavy training, alignment as a core product, context management – are the same ones open-weight labs copy within a year or two. If you're building a local stack today, our ROCm & Vulkan guide and VRAM requirements table will tell you what to buy when the next cascade lands.
Bottom line
GPT-6 Astra is the first flagship release where the agent numbers, not the chat numbers, are the headline – and where the vendor's own system card admits a monitoring regression alongside the safety wins. It's a genuinely important release for people delegating real computer work, and a non-event for anyone whose reasons for running local AI were privacy, cost, or offline access. Those reasons didn't change today.
I'll do what I always do: when I can get API access at a sane price, you'll see real tokens-per-rupee comparisons here against the local models I run daily. Until then, treat the benchmark tables as directionally true and the "most intelligent model" claim as marketing that's about 80% right.