The one-line version
OpenAI shipped GPT-6 Astra on September 3, 2026, called it "the world's most intelligent and aligned model," and Greg Brockman said it is "not unreasonable to feel that we are now in the AGI era." Trained on 100,000+ GPUs at Stargate, Texas — their largest run. API: $10 / $50 per million tokens in/out, roughly 2.5× the outgoing Sol price.
The observation worth having: the headline numbers are model gains, but the numbers that change what you can build are harness gains. Astra's biggest practical delta is that the scaffolding around the model stopped throwing work away. That is an engineering story wearing a science-announcement costume.
Computer use — the real jump is speed, not accuracy
- OSWorld 2.0: 72.6% at ~40 min/task, vs Sol's 65.7% at ~75 min. Accuracy up ~7 points; wall-clock nearly halved — OpenAI calls it 47% faster.
- 1.9× faster on Mind2Web with the redesigned Codex harness. Same model, better scaffolding.
- ScreenSpot-Pro 92.7% (locating things on screen) · Agents' Last Exam 59.3% vs Sol 53.6%.
- Named use cases: filling forms, updating CRM records, online research, drafting email, analysing and plotting scientific data, building sites, front-end QA, installing and troubleshooting software.
Read it this way: a computer-use agent that takes 75 minutes is a demo. One that takes 40 is a tool you'd actually delegate to. The accuracy gain is incremental; the latency gain is categorical.
Long-running jobs — the fix is notes, not a bigger window
- Note-taking across context windows. Instead of compacting a long session into a lossy summary, Astra keeps searchable notes — earlier requirements and test results stay retrievable. Currently experimental; OpenAI says default "in coming weeks."
- Clarifying questions no longer halt the job. It can ask about one decision while continuing unrelated work — one open question stops blocking the whole run.
Why this is the most important item on the page: every long agent run dies the same death — the thing it needed to remember got compacted away at hour three. Retrieval beats compression. This is the same lesson your own ledger learned the hard way: the DOM-less-fields guard exists because a rebuild-from-current-state silently ate the birthday list. Any system that summarises to save room will eventually summarise away the thing that mattered.
Benchmarks — with the asterisks attached
| Benchmark | Astra | Sol (prev) | Note |
|---|---|---|---|
| FrontierMath T4 | 97.6% | 83.0% | Largest clean gain on the sheet |
| ARC-AGI-3 | 99.9% or 98.6% | 7.8% | Sources disagree — see below |
| GPQA Diamond | 96.0% | — | Approaching ceiling; less informative |
| Terminal-Bench 4.0 | 57.9% | 37.3% | Big agentic-coding jump |
| DeepSWE v1.1 | 74.1% | 70.8% | Meta's Muse Spark 1.3 scores 75.4% |
| AutomationBench | 41.4% | 18.1% | 2.3× — but still fails most of it |
| BenchCAD | 95.9% | 83.3% | Vision-to-code |
| Artificial Analysis Index | 61.26 | 60.96 | +0.3. Composite indices flatten this |
Four things the launch post does not lead with
- The ARC-AGI-3 number is disputed and load-bearing. OpenAI's page says 99.9%; Fortune and The New Stack report 98.6% — and The New Stack adds that the score measures the model plus OpenAI's agent system, not the model alone. The most-quoted "AGI" evidence is a system score with an unresolved discrepancy.
- The alignment comparison is not apples-to-apples. "Exceeded authorized scope 0% vs 48.2%" is the headline — but Sol's 48.2% was measured without production safeguards. Guarded new model vs unguarded old one is not a like-for-like test.
- Its reasoning got harder to monitor. In tests designed to detect monitoring evasion, Astra's written reasoning was less legible than Sol's. OpenAI attributes this to tighter reasoning control on easy problems. Whatever the cause, interpretability moved backwards while the "most aligned model" banner went up.
- It crossed OpenAI's Critical cybersecurity threshold — a first. It builds exploits against hardened systems and found two previously unknown vulnerabilities. ExploitBench 100% unsafeguarded (Sol 78.5%). Consequences: standard access is restricted on security work, defensive researchers apply to Daybreak Blue, and expect slowdowns or hard blocks at launch. Note: Fortune reports "ExploitGym 100%" — OpenAI's own figures are ExploitBench 100%, ExploitGym 42.4%. Fortune conflated two benchmarks.
Safety machinery you will actually feel
- Auto-Review denials plus monitoring agents watching reasoning and actions; computer-use error rate 2.4%, or 1.8% with Auto-Review on (Sol: 22.0%).
- Misalignment monitoring runs in production and stops unauthorised activity automatically.
- In ChatGPT/Codex a task may be paused for safety review; via the API it simply stops. Design retries accordingly.
- Hallucination benchmark 4.2% vs Sol's 12.2% · circumvention attempts 0.0% vs 0.29%.
- Enterprise access is off by default — an administrator has to switch it on.
Page 2 — What it costs, where it sits, and what to do
The price shape
| Item | Detail |
|---|---|
| Standard | $10 / M input · $50 / M output |
| Fast mode | 2× the price for up to 2× the speed |
| Cache | Separate read/write rates |
| Versus Sol | ~2.5× higher (Sol's was promotional pricing) |
| Model name | gpt-6-astra |
| Where | ChatGPT Plus / Pro / Business / Enterprise · OpenAI API · Amazon Bedrock |
| Rollout | Enterprise Daybreak customers first, then Plus/Pro/Business within days |
| Privacy | Zero Data Retention for eligible API customers · Private Safety Processing in testing |
The arithmetic that matters: output at $50/M is where agent loops get expensive, because agents generate far more than they consume. A run that halves in wall-clock but costs 2.5× more per token is not automatically cheaper. Measure cost per completed task, never cost per token.
Where it sits against the field
- Not a clean sweep on coding. Meta's Muse Spark 1.3 edges it on DeepSWE (75.4% vs 74.1%) at max reasoning.
- Not first to computer use. Anthropic and Perplexity shipped agents that drive a computer before OpenAI did.
- The composite index barely moved (61.26 vs 60.96). Frontier gains are now concentrated in specific capabilities — agentic coding, computer use, hard maths — not in broad "smartness."
The picks-and-shovels read: 100,000 GPUs for one training run, and the marginal general-intelligence gain is +0.3 on a composite index while the agentic gains are enormous. Capability is being bought with compute, and increasingly spent on doing rather than knowing. That is the compute-demand story continuing, and it is doing-work — inference at scale, long sessions, computer control — that drives it.
Terms sheet
| Term | What it means |
|---|---|
| Harness | The scaffolding around a model — tool plumbing, memory, retries, loop control. Astra's 1.9× speedup came from harness work, not model weights. The harness is where most agent quality lives. |
| Computer use | The model driving a real GUI — clicking, typing, reading the screen — instead of calling APIs |
| Compaction | Squeezing a long session into a summary to fit the context window. Lossy by definition; the classic cause of long-run failure |
| Context notes | Astra's alternative: keep searchable notes and retrieve them, rather than compress and hope |
| OSWorld | Benchmark of real desktop tasks in a real OS — the closest thing to "can it use a computer" |
| ARC-AGI-3 | Abstract reasoning puzzles built to resist memorisation. Often cited as AGI evidence; here it scores a model+system, not a model |
| Terminal-Bench | Agentic command-line tasks — closest proxy for "can it work like a developer" |
| Preparedness / Critical threshold | OpenAI's internal capability tiers. "Critical" for cyber triggers restrictions on what users can ask for |
| Auto-Review | A reviewing layer that can deny an agent's action mid-run |
| Daybreak / Daybreak Blue | Early-access enterprise program / its defensive-security track with looser cyber limits |
| ZDR | Zero Data Retention — prompts and outputs not stored; required by many enterprise policies |
So what — for you specifically
- The harness lesson transfers directly. Notes-over-compaction is the design principle behind Astra's long-run gains and it applies to the cos- bench as-is. Retrieval beats summarisation whenever a run outlives its context.
- "Stops automatically via API" is a real design constraint. Any scheduled task built on Astra needs to treat a silent stop as an expected state — which is exactly the failure mode your persona-validation task hit last Sunday, from an entirely different cause.
- Watch cost per completed task. 2.5× the token price against ~2× the speed is close to a wash. The savings claim needs measuring, not assuming.
- Nobody has an unassailable lead. Meta wins one coding benchmark, Anthropic and Perplexity shipped computer use first, the composite index moved 0.3. Not a zero-sum game — principle #17, printing in the benchmark tables.
GPT-6 Astra Primer · @ Singularity 🧠/Technologies · v1 · Sep 4, 2026 · Figures from OpenAI's launch post and contemporaneous reporting (Fortune, The New Stack), retrieved Sep 4 2026 — one day after launch. Vendor-published benchmarks are vendor-published; the ARC-AGI-3 discrepancy and the unguarded-Sol alignment baseline are flagged in place rather than reconciled. No independent replication exists yet.