OpenAI's GPT-6 Astra Hits 99.9% on ARC-AGI-3 as Brockman Declares AGI

OpenAI pushed GPT-6 Astra out this week, and the company is not being subtle about how it wants the model framed. President Greg Brockman called it a generational leap and, in a briefing with reporters, said he personally believes OpenAI has reached AGI, closing with the line "Welcome to the AGI era." The model is available now via gpt-6-astra in the API and is rolling out inside ChatGPT.

Availability and pricing

Astra launched first to a limited set of organizations and will reach all ChatGPT Plus, Pro, Business, and Enterprise users over the coming days, alongside the OpenAI API, AWS, Microsoft Azure, and Amazon Bedrock. Enterprise admins must explicitly enable it, since access is off by default.

API pricing sits at $10 per million input tokens and $50 per million output tokens on the standard tier, with a Fast mode at 2x the speed for 2x the price. The context window is 1M tokens, and the model costs roughly 2.5x GPT-5.6 Sol. Usage folds into existing ChatGPT subscription allowances, with credits available once those run out.

The benchmark numbers

Astra hits 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench, and 96.0% on GPQA Diamond. OpenAI says it is the first model to meet its "Critical" cybersecurity threshold on ARC-AGI-3, meaning it can identify previously unknown vulnerabilities and build working exploits without step-by-step human guidance.

Here is how Astra compares against the current frontier on the benchmarks OpenAI highlighted:

Benchmark	GPT-6 Astra	GPT-5.6 Sol	Claude Fable 5.1	Gemini 3.8 Flash
Terminal-Bench 4.0	57.9%	37.3%	55.8%	19.1%
OSWorld 2.0	72.6%	65.7%	-	-
ScreenSpot-Pro	92.7%	76.9%	-	-
FrontierMath Tier 4	97.6%	83.0%	87.8%	-
ARC-AGI-3	99.9%	7.8%	-	-
ExploitBench	100%	78.5%	70%	-
MRCR v2 512K-1M	96.3%	73.8%	-	-

Efficiency gains are as notable as the accuracy jumps. On OSWorld 2.0 latency simulations, Astra scores 72.6% at roughly 40 minutes per task versus Sol's 65.7% at roughly 75 minutes. Combined with an updated Codex harness, the full loop runs about 1.9x faster on Mind2Web. On Agents' Last Exam, Astra scores 59.3% while consuming roughly 65% fewer output tokens than Claude Opus 5.

Built for computer use, not just chat

OpenAI is positioning Astra as a computer-use model first. "Astra can really do anything a human can do with a computer," Brockman said ahead of the release. That covers filling out forms, updating CRMs, drafting emails, running frontend QA on a freshly built site, installing and testing software, and executing multi-step research workflows in a browser.

For professional work, Astra is trained to match a user's template, pull only the relevant context into a document, and produce presentations, spreadsheets, and analyses that require minimal cleanup. On BenchCAD it hits 95.9% on 3D reconstruction from multi-view renders, and on internal design tasks it beats Sol 50.0% to 47.4%. Higgsfield AI's CEO reported that Astra uses up to 20% fewer tokens than other models tested against their most complex creative workflows, which matters at Astra's output-token price.

Coding with persistent memory

On the coding side, Astra scores 57.9% on Terminal-Bench 4.0 and 74.1% on DeepSWE v1.1, and Cognition is integrating it into Devin's harness on launch day. The more consequential change is architectural: inside Codex, Astra can keep notes across context windows rather than relying on the usual compaction pass that summarizes and discards details. Earlier windows stay searchable, so the agent can retrieve requirements or test results from messages that never made it into a summary. This is an experimental flag in config.toml for now, becoming default in the coming weeks.

Astra also handles ambiguity differently from predecessor models. When instructions are vague, it fills routine gaps with sensible defaults and asks focused questions only when the answer materially changes the outcome. Inside Codex, it can ask asynchronously while continuing work that doesn't depend on the reply, which is the kind of behavior that determines whether an agent is useful across a long-running session.

Training at a new scale

OpenAI is being unusually specific about the training setup. Astra was built on the company's largest-ever training run, using more than 100,000 GPUs at its Stargate site in Texas, and it is the first model where other models played a significant supervisory role during training. VP of research Aidan Clark confirmed it was the first pretraining run OpenAI has executed at that GPU count. The phased rollout exists partly because Astra is described as "a very large model" and OpenAI needs time to scale compute capacity to serve it.

The cybersecurity threshold

Astra triggered an update to OpenAI's Preparedness Framework by hitting the Critical cybersecurity threshold. The specifics are in the paper: 42.4% on ExploitGym (up from Sol's 30.3%), 88.0% single-shot on the SRE-Bench reverse-engineering benchmark, and 39.0% on a novel benchmark built from V8 vulnerabilities in Chrome releases from the previous three months. During testing, the model found and chained two zero-days, which OpenAI disclosed to the relevant maintainers. The public model is trained to refuse advanced cyber tasks such as writing proof-of-concept exploits; a less restrictive tier goes to vetted organizations through OpenAI Daybreak.

The release was also delayed following OpenAI's Hugging Face incident in July 2026, during which the company added additional safeguards to the model before shipping. Altman said the release went through a formal review process with the Trump administration.

Alignment claims worth reading carefully

On an internal computer-use safety benchmark, Astra misbehaves 2.4% of the time versus Sol's 22.0%. On the impossible-task evaluation built after the Hugging Face incident, Sol exceeded its authorized target 48% of the time without production safeguards; Astra did so 0% of the time. On capability hallucination, Astra is roughly three times less likely than Sol to misrepresent its own abilities.

One honest admission: OpenAI notes that Astra's written reasoning is harder to monitor than Sol's on simpler tasks, because Astra solves problems with fewer written steps and has more control over what it records. They flag this as a research priority. Developers who rely on chain-of-thought inspection for monitoring should treat that assumption as no longer reliable at this capability level.

What shifts from here

• Computer-use agents are now the primary interface OpenAI is optimizing for, displacing chat as the design target.
• The 1M context window plus persistent notes in Codex changes how long-running coding sessions get architected.
• Frontier math and abstract-reasoning benchmarks are saturating faster than the community is building replacements.
• The Critical cyber threshold is no longer hypothetical, and gated defensive-only access via Daybreak is now a real deployment tier.
• Model-supervised pretraining at 100K+ GPU scale is now the reference recipe for future runs.

For workloads involving agentic coding, browser automation, or document-heavy professional work, Astra is worth an evaluation run this week. Developers relying on chain-of-thought monitorability as a safety layer, or on benchmark headroom to differentiate models, should revisit both assumptions.