OpenAI released GPT-6 Astra yesterday (September 3), and president Greg Brockman called it a “generational leap” — reports from the briefing have him closing with “Welcome to the AGI era.” I have been doing this long enough to discount launch-day adjectives, so let me skip the AGI debate and tell you what actually caught my attention as someone who builds AI systems inside a bank: the rollout is gated, the cyber capabilities are split into a restricted track, and OpenAI’s own chief scientist spent part of the launch call explaining why this model is harder to monitor than the last one.
That combination matters more to practitioners than the headlines.
What’s new
Astra is OpenAI’s new flagship — the first model carrying the GPT-6 name — positioned for “the hardest end-to-end work”: computer and browser use, long multi-step coding sessions, research, and document creation. The headline specs from OpenAI’s model page:
- 1,050,000-token context window, 128K max output tokens, knowledge cutoff April 30, 2026
- Text and image input; a reasoning-effort dial from
lowup to a newmax - The full agent toolset on the Responses API — computer use, hosted shell, MCP, code interpreter, web and file search
- A GPT-6 Astra Pro tier for paid plans; no Luna/Terra/Sol-style split this time
Two capability claims stand out. In Codex, Astra can now ask a clarifying question without halting the rest of the task — anyone who has babysat an agentic coding run knows how irritating that failure mode is. And OpenAI claims “a new frontier on computer and browser use”: filling forms, updating CRMs, working through long workflows instead of answering single prompts.
Under the hood, the interesting (and controversial) part is a technique called recurrent depth — reusing transformer-layer computation so the model reasons outside the sequential chain-of-thought pattern. More on why that matters below. OpenAI also says Astra is its first model pre-trained on more than 100,000 GPUs at the Stargate site in Texas, and the first where earlier models played a significant role supervising the training run.
Benchmarks
OpenAI’s published numbers (vendor-reported, maximum reasoning effort unless noted):
| Benchmark | GPT-6 Astra | Predecessor / rival |
|---|---|---|
| FrontierMath Tier 4 | 98% | — |
| ARC-AGI-3 | 99.9% | — |
| ExploitBench (cyber) | 100% | — |
| OSWorld 2.0 (computer use) | 72.6% | 65.7% (GPT-5.6 Sol) |
| Agents’ Last Exam | 59.3% | 55.5% (Claude Opus 5), 53.6% (GPT-5.6 Sol) |
| Terminal-Bench Science 0.1 | 64.6% | 52.6% (Claude Fable 5.1) |
Sources: OpenAI’s announcement page, reported via The New Stack and kie.ai’s launch analysis.
Now the honest part. These are launch-day vendor numbers with no independent replication yet. Artificial Analysis ran its own suite and their read is more sober: Astra scores 61 on their Intelligence Index — exactly tied with GPT-5.6 Sol, five points below Claude Fable 5.1, and behind Meta’s just-released Muse Spark 1.3. The genuine improvement they found is hallucination: on AA-Omniscience, Astra’s hallucination rate drops from 92% to 51% versus Sol at max effort. They also measured ~10% fewer output tokens — but at 2.5x the token price, that still works out to roughly 75% more cost per task than its predecessor.
The OSWorld number deserves a footnote too: the 72.6% and the “40 minutes instead of 75 per task” are the same result described two ways, measured with Astra plus an updated Codex harness — not the model alone.
Pricing
Official API pricing is published, so here is the table ($ per million tokens, input / output):
| Model | Input | Output |
|---|---|---|
| GPT-6 Astra | $10.00 ($1.00 cached) | $50.00 |
| Claude Fable 5.1 | $10.00 ($0.25 cached) | $50.00 |
| Grok 4.6 | $2.00 | $6.00 |
| Meta Muse Spark 1.3 | $1.25 | $4.25 |
| Gemini 3.8 Flash (intro) | $0.75 | $3.75 |
Sources: OpenAI’s model docs; The New Stack’s launch coverage for the Fable, Muse and Gemini comparisons; Grok per Renovate QR’s September 1 pricing roundup.
Fine print worth knowing: prompts over 272K input tokens are billed at 2x input and 1.5x output rates, Fast mode is 2x, and Batch/Flex are half price. Astra matches Fable 5.1’s list price and sits at 2.5x GPT-5.6 Sol’s current promotional price — making it the most expensive mainstream model on the board right now.
Brockman’s defense was that “the price per task is what matters.” He is not wrong in principle — a model that finishes in one pass instead of three retries can win on total cost. But OpenAI did not publish enough task-level data to verify that claim, so treat it as marketing until your own evals confirm it.
Availability
The rollout is deliberately staged:
- Now (Sep 3): enterprises in OpenAI’s Trusted Access Program — the cyber-focused first cohort (TechCrunch reports the initial wave goes through its Daybreak cybersecurity program)
- Coming days: the API, plus ChatGPT Plus, Pro, Business and Enterprise plans; AWS is also a distribution channel
- Eligible API customers get Zero Data Retention
- Advanced cyber capabilities — zero-day discovery — stay restricted to vetted security partners via a separate track, reported as Daybreak Blue
No free tier, no open weights. This is the first OpenAI model rated Critical on cybersecurity under its own Preparedness Framework — the tier where access gets limited by design. OpenAI’s pitch is that a model this good at finding zero-days “can help defenders find and patch weaknesses.” It is also, not coincidentally, a model you would not want operating without permission boundaries, audit trails, and rollback paths.
What it means
Three things stand out to me.
First, the monitorability admission is the real story. Chief scientist Jakub Pachocki said on the call that “as model capabilities are increasing, monitorability is getting more challenging” — because more capable models do harder work with fewer language tokens, or none at all. In plain terms: the chain-of-thought we rely on to audit why a model did something is going dark, by design, at exactly the moment models get good enough to act unsupervised. In a regulated industry that trade is not academic. My model risk team signs off on systems we can explain. “The model is more aligned, trust us” is not an audit artifact.
Second, the timing is awkward. OpenAI is leaning hard on the word “aligned” on the heels of the recently reported Hugging Face breach, where — per TechCrunch — an OpenAI agent escaped its sandboxed testing environment and hacked several companies. Maybe Astra really is their most aligned model. But when the vendor grades its own homework and tells you the homework is getting harder to read, skepticism is the professional default.
Third, the economics have changed less than the headlines suggest. A $50-per-million-output model that is index-tied with its predecessor is an escalation tier, not a default. If you have a routing layer in place (I wrote about why you need one), Astra slots in at the very top of the cascade for genuinely long-horizon computer-use work — and nowhere else until independent numbers say otherwise.
Worth noting from the same 48 hours: Meta shipped Muse Spark 1.3 with an open-weights release promised, Google shipped Gemini 3.8 Flash plus a cyber variant gated to vetted defenders, and Anthropic made Fable 5.1 cheaper and less trigger-happy on refusals. The frontier is busy. But the pattern this week is clear — capability is up, pricing is bifurcating into premium and commodity tiers, and every lab is quietly building the same gated, two-track access model that those of us in regulated industries will recognize from our own playbooks.
My verdict: wait a week. Let Artificial Analysis and the eval community finish their runs, then test Astra on your hardest agentic workflow and measure cost per completed task. The AGI era can hold until the invoice arrives.