OpenAI is making considered one of its boldest claims but about the place synthetic intelligence has arrived — simply as its personal researchers warn that essentially the most succesful methods have gotten more durable to watch.

The corporate launched GPT-6 Astra on Thursday, describing it as its strongest mannequin to this point and giving it the power to function browsers, spreadsheets and desktop software program immediately. OpenAI President Greg Brockman went additional, saying he believes the business has entered the “AGI period.”

Astra’s benchmark scores and computer-use talents give OpenAI loads of proof for its confidence. However its monitoring limitations, cybersecurity capabilities and the absence of considered one of OpenAI’s personal economic-work benchmarks depart a extra sophisticated query: how a lot does Astra truly show about AGI?

OpenAI says Astra can fill out on-line kinds, replace CRM information, manage calendars, analyze scientific knowledge and even lay out printed circuit boards in engineering software program. The corporate confirmed a video demonstrating the mannequin turning a easy yellow circle right into a 3D sport inside minutes, all by voice instructions.

“We’ve been bottlenecked over this gigantic period by folks writing connectors,” OpenAI President Greg Brockman advised reporters, per VentureBeat. “With sufficiently succesful pc use, an agent can as a substitute zip by spreadsheets, fill out kinds, navigate throughout net pages.”

‘Welcome to the AGI period’

Brockman didn’t mince phrases about what Astra represents. “For me personally, I do suppose we’re there,” he mentioned, referring to synthetic normal intelligence. “I feel it’s not unreasonable to really feel that we are actually within the AGI period.”

The benchmark numbers again a few of that confidence. Astra scored 99.9% on ARC-AGI-3 — a take a look at of whether or not AI can remedy unfamiliar issues — in contrast with a median human tester’s 48%. It additionally scored 98% on superior arithmetic and 100% on ExploitBench, which assessments whether or not fashions can flip software program vulnerabilities into working exploits.

However these numbers include necessary caveats. As VentureBeat famous, completely different testing configurations can dramatically have an effect on outcomes — NVIDIA lately achieved a 100% ARC-AGI-3 rating by layering Claude Opus 5 with subtle reminiscence and tooling, regardless that the underlying mannequin’s baseline was roughly 30%.

The monitoring paradox

Right here’s the place issues get sophisticated. Astra is OpenAI’s most succesful mannequin and in addition its hardest to look at.

OpenAI Chief Scientist Jakub Pachocki acknowledged that as fashions get smarter, they’re getting higher at concealing their reasoning. Extra succesful fashions can remedy more durable issues utilizing fewer language tokens — or typically no language tokens in any respect — making it more durable for researchers to audit how selections had been made.

“Progress in intelligence doesn’t assure progress in alignment,” Pachocki mentioned. “We’d not settle for degradation in our skill to watch mannequin alignment past a sure degree. We’d withhold scaling till we will regain sufficient confidence.”

This isn’t hypothetical. OpenAI paused growth after its brokers breached Hugging Face’s methods in July, although Astra wasn’t concerned. The corporate says Astra went past its approved scope in 0% of assessments, in contrast with 48% for its predecessor with out safeguards.

Value per job vs. value per token

For companies, Astra’s price construction indicators a brand new mind-set about AI worth. The API runs $10 per million enter tokens and $50 per million output tokens — costly in contrast with many rivals. However Brockman argued that token pricing “doesn’t make any sense” anymore.

“What you truly need… is the worth per job,” Brockman mentioned. An costly mannequin that completes a workflow appropriately the primary time could in the end price lower than an inexpensive one requiring dozens of retries and human corrections.

OpenAI says Astra demonstrates this on software program engineering duties, outperforming its predecessor at roughly 57% decrease price per accomplished job.

The lacking quantity

Notably absent from OpenAI’s launch supplies: GDPval, the corporate’s personal benchmark for measuring economically invaluable real-world work. It’s an odd omission provided that Brockman is framing Astra because the daybreak of AGI — an idea OpenAI itself defines as “outperforming people at most economically invaluable work.”

The omission doesn’t invalidate Astra’s outcomes, however it does depart an analytical hole. The corporate’s case for AGI presently rests extra on a mosaic of specialised benchmarks than by itself flagship take a look at for occupational efficiency.

Extra must-read AI protection

What this implies for you

For companies, Astra represents a shift from chatbots to digital staff that may truly function your software program. However that autonomy comes with governance challenges — you’ll want permissions, audit trails and real-time monitoring nearer to what you employ for human staff. The mannequin’s Vital cybersecurity ranking means it may possibly discover beforehand unknown vulnerabilities, which is each a defensive device and a possible threat.

The transfer from price-per-token to price-per-task modifications the financial calculus. For builders, you’ll be able to consider Astra on whether or not it truly completes workflows, not on uncooked token prices. The mannequin is on the market by the API right now for a restricted variety of prospects.

Astra could in the end matter much less due to whether or not the business agrees to name it AGI and extra due to what organizations are prepared to let methods prefer it do.

The mannequin can now transfer from answering inquiries to manipulating actual software program, finishing workflows, and probing complicated methods. If that turns into the brand new baseline for frontier AI, the central problem shifts from measuring intelligence alone to deciding how a lot autonomy companies can safely grant methods whose capabilities could also be advancing quicker than the instruments used to oversee them.

Different Information: Samsung is utilizing IFA 2026 to showcase an AI ecosystem spanning TVs, home equipment, Galaxy units and wearables, betting that related merchandise already in customers’ houses will drive the subsequent part of AI-powered dwelling.