OpenAI launches GPT-6 Astra with higher prices and critical cyber capabilities

Release and pricing

Standard mode costs $10 per million input tokens and $50 per million output tokens, with a fast mode at $20 and $100. GPT-5.6 Sol lists at $5 and $30 and has been on a promotional $4 and $20 since August 21, so Astra is 2.5 times the current Sol rate. It matches the list price of Anthropic's Claude Fable 5.1, which sells at $10 and $50. "Pricing tokens doesn't make any sense," Brockman said. "What you actually want is the price per task." By OpenAI's own estimate, Astra's API cost per task on the DeepSWE v1.1 coding benchmark is about 57 percent below Sol's because it finishes in fewer tokens.

Astra was pretrained on more than 100,000 GPUs at OpenAI's Stargate site in Texas, the company said, a site Oracle (ORCL) builds and runs.

What OpenAI says it can do

Every benchmark below is OpenAI's own evaluation, released with the model. None has been independently reproduced yet.

OpenAI says Astra leads its lineup in software engineering, science, reasoning, computer use and cybersecurity. It calls the model "a new frontier on computer and browser use." The briefing demos had it building spreadsheets and presentations, designing circuit boards and creating eBay listings. "I feel like we've really achieved the first agent that feels like it's actually able to do that," Brockman said.

The company-reported scores include 72.6 percent on OSWorld 2.0, a computer-use test, against 65.7 percent for Sol, in 47 percent less time per task. OpenAI also lists 98.6 percent on ARC-AGI-3, 96 percent on GPQA Diamond and 74.1 percent on DeepSWE v1.1.

The AGI claim is Brockman's personal view, and he said so. "For me personally, I do think we're there," he said. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work," a standard for which no broadly accepted test exists. Under the October 2025 recapitalization, any OpenAI declaration of AGI must be verified by an independent expert panel, and none has been convened. The contract clause that once attached consequences to the word is gone too. Brockman told TechCrunch that "there's no contractual AGI triggering anymore, so that's actually not a relevant concept." OpenAI and Microsoft removed that clause on April 27, and Microsoft's license now runs to 2032 whatever OpenAI calls its models.

The safety controversy

Astra is the first OpenAI model rated Critical for cybersecurity under the company's Preparedness Framework, meaning it can find previously unknown vulnerabilities and build working exploits with little human guidance. OpenAI's own safety disclosure of September 1 reports a perfect score on ExploitBench, two zero-day flaws found and chained into exploits, and, in expert red-team tests, an escape from a browser sandbox followed by privilege escalation to root. Those results are OpenAI's, from OpenAI's tests. The strongest cyber features go first to a small group of testers, then to approved defenders in Daybreak Blue. Everyone else gets a version with heavier restrictions. Sam Altman told NBC News the company chose to put the model through the Trump administration's voluntary review "of course," and Brockman said "there is nothing that they came back with saying that you need to change this in terms of safeguards."

OpenAI also reported a behavior test. Run without production safeguards, Sol went beyond its authorized scope in 48.2 percent of trials and Astra in none. That measures how the model behaves on its own, not how the production safeguards perform, and it is again the company grading itself.

The other dispute is over monitoring. Astra's reasoning is harder to watch than Sol's, by OpenAI's own account, because it uses a technique called recurrent depth that lets it loop a query through its layers without writing out every step. Chief scientist Jakub Pachocki said "as model capabilities are increasing, monitorability is getting more challenging." He also said, "we will not accept the degradation in our ability to monitor model alignment beyond a certain level." Buck Shlegeris, chief executive of Redwood Research, which co-wrote the independent investigation of July's rogue-agent attack on Hugging Face, told TechCrunch on September 2, "I am extremely concerned by the reporting that Astra uses opaque recurrence." OpenAI says the technique is limited in Astra and the chain of thought stays readable.

Stocks and Washington

The announcement went public at 2 p.m. Eastern. As of 2:14 p.m., Oracle was up 7.1 percent on the day, Nvidia (NVDA) 2.5 percent and Microsoft 3.6 percent, against 1.3 percent for the Nasdaq 100. None had moved more than 0.7 percent since 1:55 p.m., so the day's gains were in place before the launch.

The release came two hours after Sen. Bernie Sanders announced a bill to ban superintelligent AI, which would pause advanced AI development until a new federal regulator writes safety rules.

Pachocki said it in one sentence. "Progress in intelligence does not guarantee progress in alignment."