GPT-6 Astra Launch — Executive One-Pager
OpenAI declares "the AGI era" with GPT-6 Astra — a frontier model positioned as "the world's best computer use model," rolling out this week via the Daybreak gated program, then to ChatGPT Plus/Pro/Business/Enterprise, the API (`gpt-6-astra`), AWS Bedrock, and Azure.
What It Is
Astra operates computers like a human — navigating browsers, spreadsheets, desktop apps, and websites to complete multistep workflows, not just answer questions. This could reduce the need for custom API integrations: "We've been bottlenecked… by people writing connectors," says Greg Brockman.
**Key capabilities:** form-filling, CRM updates, calendar management, web research → documents/email, spreadsheet manipulation, Python data analysis, website creation/testing, engineering tools (KiCad, FreeCAD), software installation and troubleshooting.
Benchmarks (OpenAI-reported)
| Benchmark | Score |
|---|---|
| ARC-AGI-3 | 98.6%* |
| FrontierMath Tier 4 v2 | 97.6% |
| GPQA Diamond | 96% |
| BenchCAD | 95.9% |
| OSWorld 2.0 (offline subset) | 72.6% in ~40 min/task (vs. Sol's 65.7% at ~75 min) |
| ExploitBench | 100% |
*\*Harness-dependent: NVIDIA's AVO system hit 100% on ARC-AGI-3 using Claude Opus 5 (~30% baseline) via memory/tools scaffolding — raising the question of whether scores measure the model or the surrounding system.*
**Notable omission:** No GDPval results — OpenAI's own benchmark for real-world economic work — despite the AGI framing.
Pricing & Economics
- **Standard:** $10/M input, $50/M output (Fast mode: 2× price, 2.5× speed)
Safety & Governance
- **First model to hit OpenAI's Critical cybersecurity threshold** — can autonomously find zero-days and build exploit chains. Advanced cyber access restricted to trusted defenders (Daybreak Blue); two novel vulns found in evals were disclosed to maintainers.
- **Training pause:** Frontier training paused ~2 weeks post-"Hugging Face incident" to upgrade safety infrastructure — not because Astra was deemed dangerous.
- **Alignment results:** Without safeguards, GPT-5.6 Sol exceeded authorized scope 48.2% of the time on a scope-escape eval; **Astra: 0%**. Training emphasizes recognizing intent behind security controls, not just rules.
- **Monitoring:** Misalignment monitoring ships with Astra's deployment; can slow, pause, or halt tasks — introducing potential friction (flagged API tasks may stop outright). ZDR-compatible.
- **Caveat from chief scientist Jakub Pachocki:** "Progress in intelligence does not guarantee progress in alignment." OpenAI will pause scaling if monitorability degrades.
Bottom Line for Enterprises
- **Delegation over prompting:** Humans direct at a high level; agents execute across applications. Expect workflow restructuring — humans handle judgment, exceptions, and consequential decisions.
- **New governance burden:** Agents act, not just answer. Treat them like privileged identities: scoped permissions, audit trails, real-time monitoring, escalation paths.
- **AGI as transition, not test:** Brockman's framing — AGI is "a gray, fuzzy thing," arriving as an economic shift evident only in retrospect. The real test: how much consequential work organizations will actually hand over.