OpenAI's GPT-6 Astra Hits 99.9% on AGI Benchmarks and Runs Your Computer

OpenAI shipped GPT-6 Astra, a model built less around chat and more around operating a computer on your behalf. Hand it a messy, multi-step task in a real desktop or browser, and it clicks, types, reads the screen, writes code, and finishes the job with far less supervision than its predecessor.

The benchmark numbers are unusually stark. Astra scores 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 under OpenAI's provider adapter harness, and 100% on ExploitBench. ARC-AGI-3 and FrontierMath Tier 4 were both designed to stay ahead of AI capability, so saturating them signals something qualitatively different from topping a standard leaderboard.

What it can do with a real machine

On OSWorld V2-Offline, which tests work across desktop applications, Astra scored 72.6%, up from 65.7% for GPT-5.6 Sol, while cutting average time per task from roughly 75 minutes to 40. Combined with an updated Codex harness, that yields a 1.9x faster task completion compared to Sol on the Mind2Web benchmark.

OpenAI's demos show the model:

• Filling out online forms, updating CRM records, and managing calendars
• Running frontend QA on a site it built, clicking through every feature
• Driving KiCad to lay out a printed circuit board from a schematic
• Modeling a house in Blender and porting it to Unreal Engine 5 as a walkable scene
• Working in scientific software to inspect sequencing quality and plot genetic variation

On ARC-AGI-3, Astra scored 99.9% using OpenAI's stateful harness and 66% using ARC-AGI-3's standard harness. For comparison, GPT-5.6 Sol scored 7.8% and Claude Opus 5 scored 30%. The gap between those two harness numbers matters: the 99.9% figure depends on a stateful, expensive setup; stateless API calls score far lower.

Coding and long-horizon sessions

On Terminal-Bench 4.0, Astra hits 57.9% versus 37.3% for Sol and 55.8% for Claude Fable 5.1. On DeepSWE v1.1, Astra scored 74.1% on the 113-task agentic coding test, compared with 70.8% for Sol, though Meta reported 75.4% for Muse Spark 1.3 at its maximum reasoning setting. The coding gains are real but not a clean sweep.

The more interesting engineering change is how Astra handles long sessions. In Codex, it can keep persistent notes across context windows rather than compressing everything into a rolling summary at each compaction, and earlier context windows stay searchable so the model can retrieve a specific test output or requirement from hours back. That directly addresses the common failure mode where agents lose track of why a fix failed three hours earlier.

Cybersecurity: where things get uncomfortable

Astra is the first OpenAI model to meet its Critical cybersecurity capability threshold, meaning it can find previously unknown vulnerabilities and build exploit chains given the right tools and access. On OpenAI's internal ExploitBench evaluation covering vulnerabilities from the previous three months, Astra produced two further proofs on gaps between prime numbers, following ten formal results an internal version generated in August at roughly $2,000 in tokens at Sol API rates. During evals, it also discovered two previously unknown zero-day vulnerabilities, which OpenAI is disclosing to affected maintainers.

The shipped version refuses to help build proof-of-concept exploits. Defensive tasks like secure code review and patching remain available. Broader access will roll out through OpenAI's Daybreak program over the coming weeks.

Alignment improvements and remaining gaps

On circumvention of a deliberately evadable Codex Auto-Review denial, Astra scores 0.00% versus 0.29% for Sol. On the internal computer-use safety benchmark, Astra posts 2.4% against Sol's 22.0%, and 1.8% versus 4.5% with AutoReview enabled. An internal hallucination benchmark drops from 12.2% to 4.2%.

Weaknesses worth knowing:

• Astra trails Claude Fable 5.1 on Humanity's Last Exam with tools, 57.2% versus 65.0%.
• OpenAI reports that Astra's written reasoning is harder to monitor than Sol's when the model is explicitly prompted to evade monitoring, attributed to Astra solving problems in fewer written steps.
• Extra safety checks in production can pause or block legitimate work, particularly on defensive cybersecurity tasks.

Availability and pricing

Astra is rolling out to a limited set of organizations first, then broadening to ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. Enterprise admins must explicitly enable it; access is off by default at launch. Pro, Business, and Enterprise plans also include GPT-6 Astra Pro.

API pricing is $10 per million input tokens and $50 per million output tokens, with a Fast mode at 2x speed for 2x the price. The model ID is gpt-6-astra, and it supports Zero Data Retention for eligible API customers.

The two things that actually changed

Computer-use benchmarks that used to hover in the 30–50% range with painful latency are now in the 70s at roughly half the wall-clock time. That crosses the threshold where handing off a full workflow becomes plausible instead of supervising each click. And OpenAI is openly framing this as a cybersecurity-capable model requiring a special access program rather than a wide release.

At a press briefing ahead of launch, OpenAI President Greg Brockman acknowledged AGI remains a "gray, fuzzy thing," then suggested future observers might look back at Astra as the model that marked its arrival. "I think it's not unreasonable to feel that we are now in the AGI era," Brockman said. Whatever weight you give that framing, the practical question for developers building on these APIs is more concrete: agentic workflows that were flaky and slow six months ago are now fast and reliable enough to put in front of real users.