The flagship, known as Sol, together with its mid-tier sibling Terra and the low-cost Luna, went out solely to round 20 organizations whose names have been individually cleared by the US authorities. A common launch is promised “within the coming weeks.” For now, essentially the most succesful mannequin OpenAI has ever constructed is one thing the general public can’t use, and the individuals who can have been chosen with Washington’s sign-off.
The launch arrives two weeks after the US authorities pressured Anthropic to drag its two strongest fashions offline worldwide. Put collectively, the 2 occasions describe a shift that the business has been bracing for and that’s now plainly underway: governments, beginning with the USA, are shifting to determine who will get entry to essentially the most succesful AI programs, mannequin by mannequin and buyer by buyer.
What GPT-5.6 Sol, Terra, and Luna are
GPT-5.6 is a three-model household. OpenAI describes Sol as its strongest model yet, with positive factors in coding, biology, and cybersecurity, and provides a “max” reasoning setting for lengthy issues and an “extremely” mode that spreads work throughout coordinated subagents. Terra performs near final technology’s GPT-5.5 at roughly half the value, and Luna is the most cost effective and quickest of the three. OpenAI has additionally modified the way it names fashions: the quantity marks the technology, whereas Sol, Terra, and Luna mark sturdy functionality tiers.
Functionality is the rationale for the controls. Underneath its Preparedness Framework, OpenAI is treating all three fashions as “Excessive” threat for each cybersecurity and organic and chemical functionality, although under the “Important” line, and under “Excessive” for self-improvement. On inside capture-the-flag safety testing, Sol scored 96.7 %, Terra 91.84 %, and Luna 85.19 %. OpenAI says Sol is healthier at serving to defenders discover and repair software program flaws than at operating an entire assault by itself, and that it didn’t produce a working full-chain exploit underneath the situations examined. The corporate additionally benchmarked Sol towards Anthropic’s Mythos Preview, reporting comparable outcomes on one exploitation check whereas utilizing a couple of third of the output tokens.
Pricing runs $5 per million enter tokens and $30 per million output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. Through the preview the fashions are reaching companions by means of the API and Codex, and OpenAI says broader availability by means of ChatGPT, Codex, and the API will observe.
Why the launch is gated: the cyber govt order
OpenAI didn’t select the restricted launch by itself. The Trump administration asked the company to stagger the rollout, the primary time the US authorities has preemptively pushed an American AI firm to limit a mannequin earlier than launch. In response to reporting cited by The Subsequent Net, the federal government accepted entry buyer by buyer throughout the preview.
Sam Altman defined intimately that the corporate will get the brand new mannequin to customers as quickly as it may possibly, supply: X
The request is the primary actual check of an govt order President Trump signed on June 2, which asks AI firms to voluntarily give the federal government as much as 30 days of pre-release entry to fashions with superior cyber capabilities. The order stops wanting necessary licensing. OpenAI framed its cooperation as a option to attain a wider launch sooner, nevertheless it used the announcement to register an objection. “We don’t consider this sort of authorities entry course of ought to grow to be the long-term default,” the corporate wrote, arguing that gatekeeping retains helpful instruments from builders, enterprises, and cyber defenders. OpenAI known as the preview a short-term step whereas it really works with the administration on a cyber govt order framework and what it described as a repeatable course of for future releases.
Additionally, the truth that the three fashions share names with alt cash was not misplaced on the crypto neighborhood. Sol, Luna and Terra? Severely?!
Solana famous the naming conference and its crypto similarities, supply: X
The Anthropic precedent: Fable 5 and Mythos 5 pulled offline
The rationale OpenAI’s pushback reads as cautious somewhat than defiant is what occurred to its fundamental rival. On June 12, the Commerce Division issued an export-control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any overseas nationwide, inside or exterior the USA, together with the corporate’s personal non-citizen workers. As a result of there is no such thing as a option to display screen customers by nationality in actual time throughout a base of a whole lot of thousands and thousands, Anthropic took each fashions down for everybody, worldwide, inside hours. BNC lined the shutdown because it occurred in Washington Pulls the Plug on Anthropic’s Most Powerful AI.
It was the primary time the federal government used export controls on AI fashions themselves somewhat than on the chips that practice them. The acknowledged set off was a reported technique for jailbreaking Fable 5. Anthropic stated it had obtained solely verbal proof of a slender, non-universal bypass that amounted to asking the mannequin to learn a codebase and repair the failings it discovered, and that different public fashions might floor the identical points with none jailbreak. The corporate stated governments ought to be capable of block unsafe deployments, however solely by means of a statutory course of that’s “clear, truthful, clear, and grounded in technical info,” and that this motion didn’t meet that check. Commerce Secretary Howard Lutnick delivered the order in a letter to chief govt Dario Amodei.
The friction didn’t begin there. The administration had already moved towards Anthropic earlier within the yr over a federal procurement dispute, and the corporate is contesting a Pentagon “provide chain threat” designation in court docket. Anthropic’s launch of Fable 5 and Mythos 5, and the weird choice to ship one mannequin as two merchandise separated by security classifiers, is the topic of an earlier BNC piece, Anthropic Splits the AI Frontier in Two.
What it means for entry, sovereignty, and enterprises
The quick impact lands on three teams. Enterprises that wire a single frontier mannequin into manufacturing now must deal with sudden lack of entry as an actual working threat, not a hypothetical one. Anthropic’s prospects realized {that a} mannequin could be switched off by a letter from Washington with no discover and no printed proof. OpenAI’s preview prospects are studying that entry could be granted, and presumably revoked, by means of an inventory the federal government helps keep.
The second group is the remainder of the world. Britain’s AI and on-line security minister, Kanishka Narayan, responded to the Anthropic shutdown by arguing that entry to superior AI is now a matter of nationwide functionality and a purpose for nations to spend money on their very own. Count on extra governments to succeed in the identical conclusion. If essentially the most succesful fashions could be minimize off on the US border, sovereign AI stops being a slogan and turns into a procurement choice.
The third group is the labs themselves, that are testing two totally different playbooks. OpenAI is cooperating in public whereas objecting to the precept, betting that flexibility buys a sooner path again to broad launch. Anthropic complied with its order however known as it disproportionate and warned that the identical customary utilized throughout the business would halt frontier deployments altogether. Each firms are heading towards public listings, which makes the query of who controls their core product greater than a coverage abstraction.
For the previous few years the working premise was that every new frontier mannequin can be broadly out there on launch. After Fable, Mythos, and now GPT-5.6, the safer assumption is that essentially the most succesful fashions arrive gated, and {that a} authorities decides when the gate opens.
Jason Jones Jason Jones Read More
