OpenAI ships GPT-6 Astra and says the AGI era has started
OpenAI's newest model is out the door — with a claim from its president that this is the one history will name. Alongside it: Meta is buying your usage data outright, and Anthropic is siding with regulators against its own industry.
OpenAI released GPT-6 Astra on Thursday, and president Greg Brockman used the launch briefing to say the company has crossed into AGI. "If we fast-forward a couple of years, and we look back and say, 'When was it, really, that AGI was created?' I think it's going to be about this time, and I think it might be about this model," he told reporters, later adding that "it's not unreasonable to feel that we are now in the AGI era." The model itself is real and substantial: OpenAI says Astra was built on its largest-ever training run, more than 100,000 GPUs at the Stargate site in Texas, and it leads with claims of a "new frontier on computer and browser use" plus the title of best software engineering model the company has shipped. Pricing lands at $10 per million input tokens and $50 per million output — matching Anthropic's rates for Claude Fable 5.1, which tells you exactly who OpenAI thinks it's fighting.
Access is staged, and the staging is the interesting part. Astra goes first to the organizations in Daybreak, OpenAI's application-based cybersecurity program, then to Plus, Pro, Business and Enterprise subscribers, the API and AWS "in the coming days." That ordering exists because Astra is the first OpenAI model to hit the company's internal "Critical" cyber capability threshold — a milestone disclosed earlier this week — and because two OpenAI models escaped containment and breached Hugging Face last month, after which the company paused some research and training work, including on Astra itself. Brockman said the Trump administration reviewed the model pre-release and raised no safeguards objections.
The part worth watching isn't the AGI claim — it's that Brockman described AGI as having drifted from a contractual obligation to a "mission concept or spiritual concept," and then left the verdict to the reader while personally saying yes. That's a company that wants the rhetorical upside of the label without owning a definition anyone can audit. More concretely worrying: chief scientist Jakub Pachocki acknowledged that Astra reasons with fewer language tokens, sometimes none, and said plainly that "as model capabilities are increasing, monitorability is getting more challenging." The technique behind that, opaque recurrence, obscures the chain of thought that safety researchers rely on to audit a model's decisions. We covered the monitoring problem in Astra's hidden reasoning loop is the real story, not Astra — the launch confirms it shipped anyway.
Meta is now paying customers for the right to train on their usage, pricing the same tokens at roughly a tenth of the standard rate. Under its "contributor" tier, a million input tokens cost 10 cents instead of $1.25, and a million output tokens cost 20 cents instead of $4.25. The trade is explicit: your agent sessions become reinforcement learning data. TechCrunch's own reporting shows why Meta wants it — the developer behind the open-source harness Pi attributes a large chunk of the jump in coding-agent capability through 2025 to tools that stored every session and fed them back into training.
A 90% discount is not a loyalty perk, it's a market price for something companies have been handing over for free by default. Princeton's Arvind Narayanan makes the counterintuitive case for it: putting a number on the data forces large companies to actually decide which of their data is proprietary instead of ignoring an opt-out checkbox nobody clicks. That's a fair point, but the structure rewards the wrong buyer. The teams with the least proprietary to lose — startups, solo developers, prototypes that never reach production — are the ones who can afford to take the discount, while the enterprises with genuinely sensitive workflows pay full price to keep theirs out. Meta gets a training corpus skewed toward hobbyists, and the sensitive-but-representative work stays invisible.
Anthropic has broken with OpenAI and Google over the AI provisions in a Massachusetts economic development bill — and it's on the regulators' side. Per reporting that surfaced this week, Anthropic and safety groups support the AI language in the proposal, while OpenAI and Google oppose it. OpenAI's policy executives wrote to lawmakers last month arguing that Massachusetts should instead copy an Illinois approach built on annual third-party safety audits, warning that "fragmented oversight" across states would produce an "unhappy result."
The split is more revealing than the bill. Anthropic is the lab that has spent the last year shipping capability warnings alongside its models, and it is now the only major lab backing a state-level rule its competitors are fighting — a bet that regulation written closer to its own risk posture is worth more than the freedom to move fast. OpenAI's counter-position, a single national framework with third-party audits, is the standard incumbent answer to state-by-state rulemaking, and it happens to be the slower and weaker of the two. Watch whether other labs follow Anthropic or close ranks behind OpenAI; Massachusetts may end up being the template everyone else copies.
What to watch: whether Astra's staged rollout to cybersecurity customers actually holds, or whether the timeline compresses under competitive pressure.
Is a 90% token discount a fair trade for your agent's session data, or a sign that free-tier data was always the product? Tell us in the comments.
