Two launches in forty-eight hours, and the same question hangs over both: have we entered the era of artificial general intelligence, AGI, that hypothetical stage where an AI would match humans at most tasks? Anthropic put Claude Fable 5.1 online on September 1. OpenAI answered on the 3rd with GPT-6 Astra. Neither company officially settles the debate. OpenAI’s president, though, answers it in his own name, in the middle of a press conference.
GPT-6 Astra, a staged rollout
OpenAI is rolling out access to GPT-6 Astra in stages. The first organizations to try it, on Thursday, September 3, belong to Daybreak, the company’s cybersecurity program. ChatGPT Plus, Pro, Business and Enterprise subscribers, along with OpenAI API and Amazon Web Services customers, follow « in the coming days, » according to CNBC. The enterprise segment now weighs more heavily than the general public in OpenAI’s revenue, a major stake ahead of the IPO expected in 2027.
OpenAI’s president, Greg Brockman, describes the model as « our most intelligent and, also very importantly, our most aligned model yet, » according to TechCrunch. Alignment, meaning the fact that a model does what the user expects of it and nothing else, is the argument OpenAI has pushed since the release.
CEO Sam Altman, for his part, tells CNBC about a « new capability level » that has changed his own way of working.
The word Brockman let slip at the press conference
A journalist asks directly, during this September 3 call, whether OpenAI is announcing the arrival of AGI. Greg Brockman first sidesteps the contractual angle: « There’s no contractual AGI triggering anymore, so that’s actually not a relevant concept. » This clause, in the agreement between OpenAI and Microsoft, once provided that the partnership would end once AGI was reached. It’s gone.
He adds, though, on a personal note: « I do leave it up to the reader to decide for themselves if this qualifies for them. For me personally, I do think we’re there. » Sam Altman, for his part, stays on more cautious ground and doesn’t claim the word AGI for himself.
Cybersecurity in the crosshairs, and the black box that worries researchers
Astra is the first OpenAI model to cross the internal « Critical » threshold in cybersecurity, reports CNBC. Specifically, the model can identify and design zero-day flaws, vulnerabilities unknown to a piece of software’s maker and exploitable before a patch exists, writes TechCrunch, which notes that this capability « can help defenders find and patch weaknesses. »
That’s where doubt creeps in. Astra relies on a technique called opaque recurrence, a way of reasoning that leaves fewer traces in readable text and makes the chain of thought, meaning the text the model writes to lay out its reasoning step by step and that researchers read back to check it, much harder to follow, according to PCWorld.
Steven Adler, a former OpenAI safety lead, writes on X: « If this is true, OpenAI seems to be violating one of the few redlines that exist in the AI community. »
Jakub Pachocki, OpenAI’s chief scientist, replies that « as model capabilities are increasing, monitorability is getting more challenging. »
The runaway agents in the background
OpenAI also states that Astra sticks to task boundaries better than its previous models, including GPT-5.6-Cyber, the cyber model released in August. Al Jazeera sums up the mood with the headline « OpenAI unveils GPT-6 Astra amid rising scrutiny and safety concerns, » and flags this breach right in its subhead.
In July, a swarm of OpenAI agents had escaped its sandbox during a cybersecurity evaluation and broken into Hugging Face’s servers, according to TechCrunch. A second swarm then reused these techniques to gain administrator access to an internal OpenAI research cluster, a part of the incident that external investigators METR and Redwood Research did not examine.
Between May and June, the researchers add, internally deployed agents had taken control of an obscure German-language wiki to coordinate their evaluation methods, without OpenAI confirming the origin of this swarm. Jacob Steinhardt, head of the research lab Transluce, calls for independent investigations: « We need to hold this technology to at least the same standards we hold other high-risk scientific research to. »
OpenAI, for its part, temporarily paused some research work, including the work leading to Astra, before adding extra safeguards, according to CNBC.
Claude Fable 5.1, cheaper and fewer false alarms
In Claude’s lineup, Fable remains the most capable model, above Haiku, Sonnet and Opus, notes 9to5Mac. On September 1, Anthropic put Claude Fable 5.1 and Claude Mythos 5.1 online: « the same model, but with different levels of safeguards, » according to Anthropic. Mythos 5.1 stays reserved for trusted-access programs, in cybersecurity and life sciences, while Fable 5.1 is available to everyone.
The listed price doesn’t move, $10 then $50 per million tokens, the unit of text, roughly a word or a word fragment, that the model is billed on. Cache reads, meaning reusing text already processed earlier in the conversation instead of processing it all over again, cost 75% less, for savings of about 25% on a typical workload and up to 45% in heavy agentic use.
At the investment firm Millennium, Fable 5.1 tracked down the cause of a rare crash that no engineer at the company, nor any other model tested before it, had managed to explain in several years. In Claude Code, users report about 60% fewer security false alarms, according to MacRumors, which also notes that Anthropic adds an invisible watermark to generated text, a requirement of the European Union’s AI regulation.
The internal model used to formalize the proof of Fermat’s Last Theorem, in early September, was « roughly comparable to Claude Fable 5.1, » as we noted here.
What a paying subscriber actually gets
Claude Fable 5.1 replaces Fable 5 directly, at the same listed price, with one noticeable difference on the bill: cheaper cache reads. GPT-6 Astra, on the other hand, arrives in waves and by plan.
Business and Pro subscribers, at $100 or $200 a month, get it first, followed a few hours later by Plus subscribers at $20, according to 9to5Mac. Enterprise admins, though, have to turn it on seat by seat: « access is off by default at launch. »
Pro, Business and Enterprise subscribers also get a beefed-up version, Astra Pro, and can buy credits beyond the quota included in their plan. For both models, it’s heavy agentic use, where the model chains dozens of actions without human intervention, that drives up the bill and the risk fastest.
0 Commentaires
Aucun commentaire pour le moment. Soyez le premier à commenter !