OpenAI launched GPT-6 Astra on Thursday with a two-track strategy: declare the arrival of the "AGI era" while gating the model's most potent capabilities behind Daybreak, the company's governed cybersecurity stack, and a $1 billion critical-infrastructure initiative4,7. The combination amounts to a bet that the safest way to monetize a model this powerful is to make safety itself the product.
Why it matters
Astra is not just another flagship model release. It is the first OpenAI model to reach the company's internal "Critical" cybersecurity threshold, scoring 100% on ExploitBench and 99.2% pass@4 on SRE-Bench9,14. ANALYSIS By routing initial access through Daybreak rather than a broad consumer rollout, OpenAI is establishing a precedent: models that cross a danger line ship first to defenders, not developers.
OpenAI president Greg Brockman called Astra "our most intelligent and, also very importantly, our most aligned model yet" and said it "brings together years of our research and big bets, with each breakthrough having built on the last". He ended the press briefing with a phrase that will define the news cycle: "Welcome to the AGI era".
The big picture
The benchmarks are striking. Astra scored 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, against 7.8% for its predecessor GPT-5.6 Sol on the latter test2,10. On OSWorld 2.0, a computer-use benchmark, Astra reached 72.6% versus Sol's 65.7%. OpenAI also claimed Astra was 1.9x faster than Sol on Mind2Web with Codex harness improvements.
But the ARC-AGI-3 result carries a caveat that matters. The 99.9% score was achieved using OpenAI's custom "Provider Adapter harness," which preserves opaque reasoning state between requests; the default ARC-AGI harness scored 62.7%. ANALYSIS The gap between 99.9% and 62.7% on the same benchmark, depending on harness, complicates any clean narrative about general intelligence arriving in a single model.
Brockman acknowledged the ambiguity. He described AGI as "a gray, fuzzy thing" and said the term was "no longer tied to a contractual trigger" with Microsoft, calling it instead "a mission concept or spiritual concept"13. "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". ◆ Decoupling AGI from its contractual definition gives OpenAI rhetorical freedom to claim the milestone without triggering governance consequences.
The timing is notable. Astra's launch comes "only weeks after a serious AI safety incident involving other models under its development caused international concern and a pause in Astra's training"3. ◆ Framing Astra as the "most aligned" model and leading with Daybreak positions the launch as a corrective narrative, not just a product release.
Between the lines
The $1 billion Daybreak for Frontline Defenders initiative extends subsidized access to Astra's cyber capabilities for organizations protecting water, electricity, local government, and banking systems. A pilot with the Multi-State Information Sharing and Analysis Center (MS-ISAC) will train and support state-level defenders. ◆ The structure converts a liability (a model that scores 100% on exploit benchmarks) into a public-good story, while creating a distribution channel that embeds OpenAI's stack in government infrastructure.
On the commercial side, Latent Space's early-access testing positioned Astra as "an automated AI Engineer you can hire for <$6 an hour," capable of commanding subagents, keeping pipelines saturated, and maintaining coherence over billions of tokens in a single agent thread. API pricing matches Anthropic's Fable 5 tier: $10 per million input tokens and $50 per million output tokens. Latent Space noted this was "the first time in their mutual history" that OpenAI outpaced Anthropic in launch engagement1. ◆ Pricing at parity while claiming benchmark leads on most self-reported metrics is a direct competitive play.
OpenAI said Astra was built on its largest-ever training run, using more than 100,000 GPUs at its Stargate site in Texas, and that this is its first model to use other models in a significant supervisory role during training. Chief scientist Jakub Pachocki said monitoring the reasoning process of a model was "a critical form of oversight".
What's next
Astra will roll out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS, over the coming days. The broader consumer rollout will test whether the Daybreak-first gating strategy holds as access widens. Anthropic's Fable 5 does not yet have a published ARC-AGI-3 result, and the competitive response from both Anthropic and Google DeepMind will shape whether "the AGI era" becomes an industry consensus or remains OpenAI's unilateral claim.