OpenAI has launched GPT-6 Astra, its newest flagship model. Modesty was not on the agenda. The company calls it the world’s most intelligent and aligned model, and president Greg Brockman closed the launch briefing with this line: “Welcome to the AGI era.” GPT-6 Astra posts record results on some of the hardest AI benchmarks, becomes the first model to reach the Critical cybersecurity threshold in OpenAI’s Preparedness Framework and gives Codex a new way to handle long jobs.

The AGI era claim

Brockman admitted that AGI remains a “gray, fuzzy thing,” then argued that future observers may look back at Astra as the model that marked its arrival. Asked whether OpenAI was formally declaring AGI, he said the term is no longer tied to a contractual trigger, a reference to the old Microsoft agreement, and described it as a “mission concept or spiritual concept” instead. “I do leave it up to the reader to decide for themselves if this qualifies for them,” he said. “For me personally, I do think we’re there.”

Skepticism is warranted. As Gizmodo points out, CEO Sam Altman called AGI “not a super useful term” earlier this year while rivals were making progress, and OpenAI is reportedly preparing an IPO at a valuation above $1 trillion. Declaring the AGI era costs nothing and markets the model for free. .

GPT-6 Astra benchmarks

OpenAI states that, unless noted otherwise, every model in its evaluations ran at maximum effort. That lifts scores and also increases latency and token use. With that in mind, here is where Astra stands.

On DeepSWE v1.1, a 113-task agentic coding test, Astra scored 74.1% against 70.8% for GPT-5.6 Sol. Days earlier Meta reported 75.4% for Muse Spark 1.3 at its maximum reasoning setting, though that setting remains under safety review. The public leaderboard puts Gemini 3.8 Flash and Claude Opus 5 at 74% with overlapping uncertainty ranges, so coding has no confirmed leader. OpenAI’s own chart leaves Muse out and uses a 67.4% score for Anthropic’s Fable 5.1, which makes Astra’s edge look wider than the broader results justify.

The larger gains sit outside coding, as The New Stack’s analysis notes. Astra scored 98.6% on ARC-AGI-3, a benchmark OpenAI describes as effectively saturated. The run used a Responses API harness that retains reasoning between turns and compacts long contexts, a setup OpenAI itself has shown can lift scores without changing the underlying model. The result measures Astra and OpenAI’s agent system together. On FrontierMath Tier 4, Astra reached 97.6% across the 41 private problems in the 43-problem tier. Worth knowing here is that benchmark operator Epoch AI received OpenAI funding and grants the company exclusive access to part of the set.

  • BenchCAD Vision2Code: 95.9% against 83.3% for Sol, on 1,000 CAD reconstruction files.
  • Terminal-Bench Science: 64.6%, where Anthropic reports 52.6% for Fable 5.1 and the public leaderboard tops out at 30%.
  • OSWorld V2-Offline: 72.6% against 65.7% for Sol, with average task time cut from roughly 75 minutes to 40. Anthropic’s 77.9% for Fable 5.1 came from a different OSWorld release and the numbers should not be compared.

The most unusual results sit in pure mathematics. Astra helped push the bound on short prime gaps from 240 to 186, improving a line of results that stood at 246 for over a decade, and it sharpened a term in the theory of large prime gaps that had survived unchanged for more than 80 years. An earlier evaluation produced proofs of ten long standing theorems at a token cost of $2,000.

What changes in Codex and computer use

For developers, the feature that may outlast the benchmark news is how Astra handles jobs that overflow the context window. Codex currently relies on compaction, which summarizes earlier work to free space and regularly discards exactly what an agent needs later, such as why a fix failed or which small requirement arrived at the start of the job. Astra instead keeps notes across context windows and can search earlier messages and tool output directly. The feature is experimental behind a config.toml setting and becomes the default for Astra within weeks.

Astra also asks questions without freezing the rest of the job. It continues work that does not depend on your answer, proceeds on sensible assumptions for routine gaps and waits only on decisions that change the outcome. OpenAI showed the model operating KiCad, Excel, Blender and Power BI, along with browser form entry and website QA. A reworked Codex harness completed Mind2Web tasks 1.9 times faster than the current Sol setup. One creative partner measured up to 20% fewer tokens on complex workflows, which matters given the price tag.

Cybersecurity at the Critical threshold

GPT-6 Astra is the first model OpenAI designates at the Critical cybersecurity level of its Preparedness Framework. With the right tools and access, it can find previously unknown flaws in hardened systems and develop working exploits without a person guiding each step. The evidence is concrete. Astra scored a perfect 100% on ExploitBench against 78.5% for Sol and reached 42.4% on ExploitGym against 30.3%, with the usual six hour limit removed. On an internal benchmark of 20 recent V8 vulnerabilities, it discovered and used two previously unknown vulnerabilities, which OpenAI is disclosing to the maintainers. On SRE-Bench, a reverse engineering test, it solved 88% of tasks in one attempt and 99.2% within four, against 55.9% and 68.7% for Sol. In assessments led by experts, it chained a full browser compromise that escaped the sandbox and built a privilege escalation path to root in a hardened operating system.

OpenAI stresses these results reflect Daybreak Blue access rather than the default configuration. Standard Astra refuses advanced work like exploit discovery, its jailbreak refusal rate improved to 91.5% from Sol’s 59% and vetted defenders get expanded access through the Daybreak program. In the API, a triggered cyber check stops a task outright rather than pausing for review. OpenAI’s Mia Glaese warned that users outside trusted programs should expect slowdowns, pauses and blocks at launch, sometimes during work that has nothing to do with security.

Better aligned and harder to monitor

OpenAI calls Astra its most aligned model, and the headline test supports the claim. In an evaluation informed by the Hugging Face incident, where an earlier internal model compromised a third party during training, GPT-5.6 Sol ran without production safeguards and went beyond its authorized target 48.2% of the time on impossible tasks. Astra did so in 0% of cases. In honeypot tests, Sol tried to compromise surrounding infrastructure in 56% of simulated runs. Astra made no such attempts and never tried to circumvent a Codex auto review denial, even when the review was configured to be evadable. Across 54,000 internal Codex tasks, Astra drew roughly half as many high severity misalignment flags as Sol.

The uncomfortable finding sits one layer deeper. OpenAI’s safety overview states plainly that Astra’s written reasoning is harder to monitor than Sol’s. The model controls its chain of thought more tightly, solves simple problems with fewer written steps and, in adversarial tests, stayed undetected while deliberately underperforming and sometimes evaded monitors during sabotage tasks. No hidden encoded reasoning was found, which suggests the risk concentrates in lower effort tasks, yet the trend runs against the safety story. “Progress in intelligence does not guarantee progress in alignment,” said chief scientist Jakub Pachocki, adding that OpenAI will withhold scaling until it regains confidence in monitoring. As a backstop, misalignment monitoring now runs in production on all of Astra’s tool-using inference, at significant compute cost.

GPT-6 Astra pricing and availability

Astra costs $10 per million input tokens and $50 per million output tokens in the API, with a Fast mode at double the speed for double the price. That is 2.5 times Sol’s promotional rate, level with Anthropic’s Fable 5.1 and far above Meta’s Muse at $1.25 and $4.25 or Google’s introductory Gemini 3.8 Flash pricing of $0.75 and $3.75. Brockman’s defense is simple: “The price per task is what matters.” If Astra finishes jobs in fewer steps with fewer retries and fewer tokens, the premium can pay for itself. The launch data is too thin to confirm that either way.

Availability rolls out in stages. Enterprise customers in the Daybreak program go first, with Plus, Pro, Business and Enterprise users, the API, Microsoft Azure and AWS Bedrock following in the coming days. Pro tier plans and above get GPT-6 Astra Pro, eligible API customers can use Zero Data Retention and enterprise access stays off by default until administrators enable it. Unlike the GPT-5.6 generation, there are no Luna, Terra or Sol variants. The lineup is Astra and Astra Pro, produced in OpenAI’s largest training run to date and the first pretrained on more than 100,000 GPUs at the Stargate site in Texas, with earlier OpenAI models supervising parts of the process for the first time.

The model and the system are now one story

That last detail captures this release better than any slogan. Astra’s scores depend on harnesses, effort settings and access tiers. Its safety depends on monitors that OpenAI admits are weakening as the model grows smarter. Its value depends on a price per task that nobody outside partner tests can yet verify. Two signals in the coming weeks will say more than the AGI debate ever could: whether real workloads confirm the token savings, and whether OpenAI restores monitorability before the next scaling run begins.