OpenAI released GPT-6 Astra on Thursday, calling it the company's most capable model yet and, in the words of co-founder and president Greg Brockman, the clearest signal so far that artificial general intelligence has arrived. "Welcome to the AGI era," Brockman told reporters at a briefing in San Francisco, capping a rollout that began with enterprise customers on OpenAI's Daybreak program before extending to ChatGPT Plus, Pro, Business and Enterprise subscribers.
The headline feature is not a chat window but a set of hands. Astra can operate browsers, spreadsheets and desktop software directly, filling out online forms, updating customer-relationship records, organizing calendars, analyzing scientific data sets and even sketching circuit-board layouts from voice instructions. "We've been bottlenecked over this gigantic era by people writing connectors," Brockman said, framing the model's computer-use ability as the removal of that bottleneck.
Benchmarks that don't quite agree
OpenAI's own numbers are eye-catching: a 99.9% score on ARC-AGI-3, a benchmark designed to be difficult for narrow pattern-matchers; a perfect 100% on ExploitBench, a test of vulnerability-exploitation skill; 97.6% on FrontierMath Tier 4; and 96% on GPQA Diamond, a graduate-level science exam. Researcher Mia Glaese said the model gives "users incredible capabilities at their fingertips" that "seemed very far away less than a year ago."
But independent observers, writing in VentureBeat's account of the launch, note the ARC-AGI-3 score depends heavily on which system harness is wrapped around the model — Nvidia's AVO architecture, for instance, reportedly reached a perfect score using Anthropic's Claude Opus 5 as its base, a model that scores closer to 30% on its own. That, critics argue, suggests long-horizon task performance is emerging from elaborate agent scaffolding as much as from the underlying model. OpenAI's launch materials also omit GDPval, the company's own 2025-vintage benchmark for real-world occupational value — the exact metric that would most directly support a claim about AGI's economic significance.
For me personally, I do think we're there. I think it's not unreasonable to feel that we are now in the AGI era.
Greg Brockman, OpenAI co-founder and president
Alignment, pricing and the AGI hedge
Brockman's own chief scientist struck a more cautious note. "Progress in intelligence does not guarantee progress in alignment," said Jakub Pachocki, who added that OpenAI "will not accept the degradation in our ability to monitor model alignment beyond a certain level." That tension is not abstract: more capable models tend to solve problems using fewer interpretable reasoning steps, making it harder for engineers to audit why a system did what it did — what one researcher called the defining infrastructure problem of the coming agent era.
Brockman also used the launch to argue that OpenAI's per-token pricing model is becoming obsolete. "Pricing tokens doesn't make any sense," he said. "What you actually want is the price per task." Even so, the company published conventional token pricing for developers: $10 per million input tokens and $50 per million output tokens for GPT-6 Astra through the API, alongside availability on Microsoft Azure and Amazon Bedrock.
Pressed on what "AGI" actually means in practice, Brockman reframed it as an economic transition rather than a single threshold crossed on a test: if enterprises eventually restructure workflows around autonomous agents that can operate software the way employees do, he argued, the shift will look obvious only in hindsight. That framing lets OpenAI claim the moment without pinning itself to a falsifiable definition — a hedge that reflects how contested the term has become even inside the company that popularized it.
For now, GPT-6 Astra's real test will not be a leaderboard but adoption: whether enterprises trust a model that can click, type and transact on their behalf, and whether the safety monitoring Pachocki described can keep pace with a system built explicitly to need less human supervision, not more.
