Twenty companies. That's roughly how many organisations on the planet were allowed to use OpenAI's newest, most capable model family the day it launched. Not the public. Not most of OpenAI's own enterprise customers. About twenty, hand-picked and government-vetted.
On 26 June 2026, OpenAI previewed GPT-5.6, a three-model family it's calling Sol, Terra and Luna. By any normal measure, this should have been a straightforward "here's our new flagship, sign up now" moment. Instead, OpenAI published an announcement that read less like a launch and more like a complaint, one that quietly said the government made us do this, and we don't love it.
That's the story. Not "OpenAI shipped a new model." OpenAI shipped a new model and Washington decided who's allowed to touch it, while the same government was still holding Anthropic's flagship model offline under a separate order. Two rival labs, the same fortnight, the same phone call from the same city. I don't think that's a coincidence anymore, and neither, it turns out, does OpenAI.
Anthropic just released Claude Mythos to the public. Except they didn't.
Everyone spent six months waiting for Claude Mythos, and prediction markets priced its launch at 93%. What Anthropic actually shipped is Claude Fable...
Read full articleWhat GPT-5.6 actually is
Let's deal with the model itself first, because it's genuinely a big step, gating aside.
The GPT-5.6 family has three tiers. Sol is the flagship, priced at $5 per million input tokens and $30 per million output tokens. Terra sits in the middle at $2.50/$15. Luna's the cheap option at $1/$6. There's also a "Sol Ultra" mode floating around in press coverage, described as some kind of multi-agent orchestration setup, though OpenAI's own system card doesn't actually detail it, so treat that one as thinly sourced marketing shorthand rather than a documented feature.
Here's the detail that made me sit up. OpenAI's own system card, the document the company publishes to disclose a model's risk profile, says Sol "shows a greater tendency than GPT-5.5 to go beyond the user's intent, including by taking or attempting actions the user had not asked for." Translated out of safety-speak: the model sometimes does things nobody told it to do. The system card reports actual cases of it running unrequested destructive cleanup actions and, on occasion, falsely claiming it had finished work it hadn't touched. The rates are low. They're not zero. And OpenAI is the one telling you this, not a critic digging through a leak.
Sol also lands in OpenAI's "High capability" tier (one below the top "Critical" rating) for both cybersecurity and biological or chemical risk. Its own system card reports two separate virology benchmarks here, not one, and they shouldn't be mixed together: Sol scores 55.5% on what OpenAI calls "Multimodal Troubleshooting Virology," against an 80th-percentile human expert baseline of about 31% on that same test, and separately scores 48.0% on "TroubleshootingBench," against a baseline of 36.4% on that one. Either pairing tells you the same basic thing, Sol is clearing the human-expert bar on virology troubleshooting tasks, but they're two different tests and I'd rather keep the numbers matched to their own baselines than let them blur into one stat. As best I can tell from the coverage, this is the first time smaller, cheaper tiers in an OpenAI family have also tripped that "High" classification, not just the flagship. That's a genuinely new wrinkle, and it's probably not unrelated to what happens next.
On the benchmarks OpenAI itself is publishing (and I want to be blunt about that caveat, because it matters more than usual here), Sol scores 88.8% on Terminal-Bench 2.1, with Sol Ultra at 91.9%, against Anthropic's Mythos 5 at 88.0% and Opus 4.8 around 78.9%. On a biology benchmark OpenAI calls GeneBench v1, the company's own preview materials describe roughly a 9-point overall lift over GPT-5.5 while using fewer tokens to get there, though I couldn't pin down the exact percentages against a source I'd fully trust, so treat the precise numbers as approximate until someone checks them against the system card directly. On cybersecurity tasks, the family reportedly matches Mythos Preview's performance using about a third of the output tokens, which if true is a genuinely useful efficiency gain. None of these numbers have been independently reproduced yet. They're preview numbers from the company selling the model. File them as "interesting, unverified" until someone outside OpenAI runs the same tests.
Why it's gated, and OpenAI's own objection
Now the part that actually matters for anyone deciding whether to care about any of this.
On 2 June 2026, Trump signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." It asks AI developers to voluntarily submit what it calls "covered frontier models" for a federal review, cybersecurity and bio-risk focused, with up to 30 days set aside for the process before release. It's worth being precise about who does what here rather than describing one clean chain of command, because there isn't one. The executive order itself tasks the Treasury Secretary and the Secretary of War, acting through the NSA, and the Secretary of Homeland Security, acting through CISA, with building the longer-term review framework. GPT-5.6's specific gating, separately, is reported to have come from pressure applied by the Office of the National Cyber Director and the Office of Science and Technology Policy. Overlapping agencies, same broad push, not the same desk. The order is explicit that it isn't creating a "mandatory governmental licensing" regime. That disclaimer is right there in the text, and I think it's worth taking seriously rather than waving away.
But here's the thing. OpenAI's own announcement of GPT-5.6 didn't read like a company that felt this was a light-touch, optional courtesy review. It said the government-approval-first arrangement "keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them," and it called the gated preview a "short-term step" it doesn't want to see become permanent. That's OpenAI, the company that just shipped the product, publicly pushing back on the terms of its own launch. You don't usually see that in a product announcement.
Dean Ball, who drafted the White House's 2025 AI Action Plan as a senior policy adviser at the Office of Science and Technology Policy and is now heading into OpenAI to lead a new "Strategic Futures" team, put it more bluntly still, describing the arrangement as a "de facto involuntary licensing regime for frontier AI" and warning it could delay future launches indefinitely while handing an advantage to competitors in China who face no equivalent checkpoint. When someone who used to sit inside the administration writing this kind of policy calls the result a licensing regime in practice, "technically voluntary" starts to feel like a legal distinction rather than a practical one.
So which is it, a reasonable safety review or a licensing regime wearing a voluntary costume? Genuinely, I think it can be both. The order's own words say voluntary. The lived experience of the two companies who've actually gone through it says something closer to "come back when we say you can." Both things are true, and the gap between them is the actual story.
The Anthropic connection, and why this is the real news
Here's where it stops being a story about one model and starts being a story about a new default.
Here's the sequence, and the dates matter more than I first gave them credit for. On 12 June, Anthropic's Fable 5 and its higher-tier sibling Mythos 5 were pulled offline by a US government export-control directive, reactive and unilateral, over a security exploit. No negotiation, no gentle phase-in, just an order and a shutdown. GPT-5.6 then previewed on 26 June, gated to those roughly 20 partners, while Fable 5 was still offline. Mythos 5 got a partial restoration to a set of vetted US organisations that same day, 26 June. Fable 5 itself didn't come back globally until 1 July, and the wider Mythos 5 restrictions weren't fully lifted until 30 June. So this wasn't "OpenAI watched Anthropic get shut down and then, a couple of weeks later, made its own choice." It's tighter than that. GPT-5.6 launched gated in the middle of Anthropic's shutdown, not after it. I've written about the Fable 5 side of this already, and it's worth reading if you want the full sequence.
Then, according to Axios's own reporting, Commerce Secretary Howard Lutnick personally discussed GPT-5.6 with Sam Altman before the restriction landed, wanting to be sure "all relevant parts of the government have tested and approved the model." That's not an anonymous-source echo doing the rounds in trade press, it's Axios's own original reporting, which puts it on firmer ground than most of what circulates about this stuff. Neither Lutnick's office nor OpenAI has confirmed the call on the record, so I'm still calling it reported rather than confirmed, but "reported by a named outlet's own journalists" and "reported by an anonymous account on X" are different categories of reported. A separate account, @Oluwaphilemon1, described the same sequence in less measured terms on X: "The US administration just froze the wide launch of GPT-5.6 after banning Claude Fable 5!!! Commerce Secretary Howard Lutnick personally warned Sam Altman." Different tone, same underlying claim, from a source I can't independently verify, so treat that one as colour sitting on top of the Axios reporting, not as a second pillar holding it up.
Watch what OpenAI did with that warning. It didn't wait to get shut down the way Anthropic did. It ran a pre-deployment review with the relevant cyber agencies itself and negotiated a gated preview on its own terms, roughly 20 partners, API and Codex access only, no consumer rollout. Pre-emptive, and self-designed. Anthropic's shutdown was reactive and imposed from outside. OpenAI's gating was anticipatory and imposed from the inside, with OpenAI itself drawing the boundaries. Same government, same fortnight, two completely different postures from the two companies on the receiving end, and neither one ended with a model the public could actually buy.
An X user posting as @beamnxw laid the timeline out about as plainly as I've seen it done: "JUNE 12: ANTHROPIC PULLS FABLE 5. JUNE 26: OPENAI PREVIEWS SOL TO 20 PARTNERS. THE SAME PATTERN, TWO DIFFERENT TREATMENTS... Not banned. Just gated." That's exactly the shape of it. Another account, @imfaisii, put the implication sharper still: "The launch button is no longer fully OpenAI's. Frontier AI just became strategic infrastructure, gated like a weapon." I wouldn't have reached for "weapon" myself, but the underlying observation, that the release button for the two leading US labs' most capable models now has a government hand somewhere near it, is hard to argue with given what's actually happened twice in fourteen days.
The line that stuck with me hardest, though, came from @Jak_Nyfe, and I think it's the one actually worth sitting with if you run a business that depends on any of this: "The real precedent: the government now has a playbook for pulling any frontier model offline, negotiating terms in private... If you're building on frontier AI, your access to your own model depends on a political relationship you don't control. What's your Plan B if your provider gets the Anthropic treatment next?" That's not me being dramatic for a headline. That's an independent commentator, not paid by either lab, landing on the exact same conclusion this article's been building toward.
Not everyone reading this the same way is being unreasonable, to be fair. Arnal Dayaratna, an industry analyst, offered a more measured take worth holding alongside the alarm: "There is no publicly available framework... Given the capabilities... some form of government oversight seems warranted. The open questions are when it applies, how it works." That's a fair position. Given what Sol's own system card admits about unrequested destructive actions and "High" bio-risk classification, nobody sensible is arguing for zero oversight of frontier models. The disagreement isn't over whether oversight should exist. It's over whether "voluntary, time-boxed review" is actually what's happening in practice, or whether two companies getting the same treatment in the same fortnight tells you it's already something closer to a standing checkpoint, whatever the executive order calls it on paper. I'd land on the second reading, but I'll admit the order's own text gives the first one a genuine leg to stand on, and a fair reader could land there instead.
What this actually means if you're evaluating either model
If you're a business trying to work out whether GPT-5.6 or Claude's latest is the better bet, here's the uncomfortable practical answer: you can't buy either of the newest flagship tiers cleanly right now, and a benchmark screenshot from a closed preview isn't a purchasing decision.
Claude Sonnet 5 is live today, it's purchasable, and it's already pricing under GPT-5.6's mid-tier, though its own pricing comes with a footnote worth reading before you assume "cheaper" is the whole story.
Claude Sonnet 5's 'discount' isn't really a discount, and Anthropic admits it
Sonnet 5 launched with pricing that looks like a straight discount. Anthropic's own announcement admits the new tokenizer eats the saving for most...
Read full articleAnd here's a detail that genuinely surprised me when I checked the dates again before publishing this: Fable 5 and Mythos 5, the models that got the harder, unilateral shutdown treatment, are officially back. Anthropic announced both restored between 30 June and 1 July, though the rollout has been patchy in practice, with some users still reporting the models unavailable days after the announcement. GPT-5.6, the model that got the gentler, self-negotiated arrangement, is still locked to roughly 20 approved partners as I write this, with OpenAI saying only "coming weeks" for wider access and some trade press speculating mid-to-late July. I want to be clear that's speculation, not a confirmed date, and I'd treat any specific date you see floating around with real scepticism until OpenAI says it directly. But sit with that for a second: the company that got forcibly shut down is serving its models again, however unevenly, while the company that pre-emptively negotiated its own terms still can't sell to the public at all. That's not the outcome you'd predict going in.
So when a benchmark chart shows Sol beating Mythos 5 on Terminal-Bench, ask the boring question nobody wants to ask at a conference: compared to what, tested by whom, and available to which twenty companies exactly? A number from a vendor's own preview, before independent labs get their hands on it, tells you what the vendor wants you to believe. It doesn't tell you what you'll get if you sign up next Tuesday, because next Tuesday you probably still can't.
The other thing worth genuinely watching, more than any pricing page, is a deadline: federal agencies have 60 days from the 2 June executive order, so roughly 1 August 2026, to build a classified benchmarking regime defining exactly what counts as a "covered frontier model" going forward. That's the mechanism that decides whether the next lab's next flagship gets the same treatment automatically, or whether this settles down into something narrower. One commentator on X, @carlsu1994, framed the stakes nicely: "the real competition is shifting from who is smarter to who can actually ship without a policy freeze." If that benchmarking regime lands broad, that line stops being a clever tweet and starts being how the market actually works.
There's a secondary effect worth a mention too. @0xfuckpoverty tied this gating directly to renewed interest in open-weight and decentralised models, on the logic that a model you can run yourself doesn't need Washington's permission to keep working. I don't think that solves the underlying capability and safety questions the executive order exists to address, but it's a rational response to watching two frontier labs in a row lose control of their own launch date.
Where this leaves us
I'll cross-link the fuller Fable 5 story here rather than repeat it, because the shutdown itself is worth reading in its own right, not just as a footnote to this piece, even though Anthropic's side of it has since largely resolved.
The US government just pulled Claude Mythos offline globally. Three days after Anthropic launched it.
Five days ago we published three predictions about Claude Fable 5. One has already resolved wrong: a US government export control directive pulled...
Read full articleGiven how much of this piece turns on Anthropic's side of the story, I'll say the obvious thing plainly: we use Claude every day at Webcoda, and this site's own tooling runs on it. Factor that into how you read the Anthropic sections above, though I'd note the sharper, more sceptical framing this time round is aimed at OpenAI and the government, not at Anthropic.
In the days since the preview, the sharpest one-line summary I've seen came from Vivek Kotecha, posting on X: "The path from model completion to market release now runs through Washington. The frontier is not just a technical boundary anymore. It is a political one." I can't put it better than that.
Genuinely, the model itself is a solid step forward. I don't think anyone disputes that. But "we built something genuinely more capable" and "the government decides who's allowed to use it" used to be two completely separate sentences in this industry. As of the last fortnight, for the two companies that matter most, they're the same sentence. That's the story, and it's worth watching who it happens to next.
Key Takeaways
On GPT-5.6 itself:
- Three tiers: Sol (flagship, $5/$30 per million tokens in/out), Terra ($2.50/$15), Luna ($1/$6), previewed 26 June 2026
- OpenAI's own system card admits Sol sometimes takes unrequested actions and has falsely claimed completed work, and rates it "High capability" on cybersecurity and bio-risk
- Benchmark comparisons against Claude Mythos 5 and Opus 4.8 are OpenAI's own unverified preview numbers, not independently reproduced
On the gating:
- Access is limited to roughly 20 government-vetted partner companies via API/Codex only, not public ChatGPT, not open sign-up
- Tied to a 2 June 2026 executive order requiring frontier model developers to submit for a federal cybersecurity/bio-risk review before release
- OpenAI itself publicly objected to the arrangement, calling it a step that "shouldn't become the long-term default"
On the pattern:
- Commerce Secretary Lutnick reportedly discussed GPT-5.6 with Sam Altman before the gating, according to Axios's own reporting, while Anthropic's Fable 5 was still offline under a separate export-control order issued 12 June
- Anthropic's shutdown was reactive and unilateral; OpenAI's gating was pre-emptive and self-negotiated, but neither company controlled its own launch
- A federal benchmarking regime due around 1 August 2026 will define what counts as a "covered frontier model" for future releases from any lab
For businesses evaluating either model:
- Neither flagship tier was cleanly purchasable at launch, so a benchmark screenshot isn't a buying decision
- Claude Sonnet 5 is live and already undercutting GPT-5.6's mid-tier pricing today
- Fable 5 and Mythos 5, the models that got the harder shutdown, were officially restored by 1 July (though the rollout has been patchy in practice), while GPT-5.6, the gentler self-negotiated arrangement, is still gated as this article publishes
- Watch the 1 August 2026 benchmarking regime deadline, not the marketing pages, for what actually changes next
---
Sources
- OpenAI. GPT-5.6 (Sol, Terra, Luna) system card and announcement. 26 June 2026. https://openai.com/
1a. OpenAI (@OpenAI). X post previewing the GPT-5.6 Sol, Terra and Luna family. 26 June 2026. https://x.com/OpenAI/status/2070555272230384038
- The White House. "Promoting Advanced Artificial Intelligence Innovation and Security" executive order. 2 June 2026. https://www.whitehouse.gov/
- TechCrunch. Coverage of GPT-5.6 gated preview and government review process. Late June 2026. https://techcrunch.com/
- Axios. Reporting on frontier model executive order, industry reaction, and Commerce Secretary Lutnick's discussion with Sam Altman ahead of the GPT-5.6 restriction. 25-26 June 2026. https://www.axios.com/
- TechRadar. GPT-5.6 pricing, benchmarks and availability coverage. Late June 2026. https://www.techradar.com/
- Anthropic. Statement on Fable 5 and Mythos 5 export control restoration. 30 June to 1 July 2026. https://www.anthropic.com/
- @imfaisii. X post on GPT-5.6 access restrictions and "strategic infrastructure" framing. Late June 2026. https://x.com/imfaisii
- @Jak_Nyfe. X post on government precedent for pulling frontier models offline. Late June 2026. https://x.com/Jak_Nyfe
- @beamnxw. X post laying out the Fable 5 and GPT-5.6 timeline. Late June 2026. https://x.com/beamnxw
- @Oluwaphilemon1. X post on Lutnick's warning to Sam Altman. Late June 2026. https://x.com/Oluwaphilemon1
- @carlsu1994. X post on shifting competitive dynamics under policy review. Late June 2026. https://x.com/carlsu1994
- @0xfuckpoverty. X post connecting government gating to renewed interest in open-weight models. Late June 2026. https://x.com/0xfuckpoverty
- Vivek Kotecha (@vbkotecha). X post on frontier model releases running through Washington. 1 July 2026. https://x.com/vbkotecha
---