For twelve days, using OpenAI's newest and most capable model required more than a credit card. It required being on a list.
Not a waitlist. A roster the United States government had reviewed, organization by organization, before OpenAI could hand out a single API key.
That list opened Thursday morning. By the time it did, the head start it was meant to protect had already gone.
On June 26, OpenAI previewed GPT-5.6, a three-model family led by a flagship called Sol. Two White House offices, the Office of the National Cyber Director and the Office of Science and Technology Policy, asked the company to hold its broader launch and limit access to roughly 20 vetted organizations whose names OpenAI shared with the government first. Sam Altman agreed. He also made clear he did not enjoy the arrangement.
"We don't believe this kind of government access process should become the long-term default," OpenAI wrote in its June 26 announcement. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."
The gate held for 12 days. It opened Thursday, July 9, when Sol, Terra, and Luna became available to anyone through ChatGPT, the API, and Codex. Axios, which broke both ends of the story, reported that the clearance followed additional testing by the Commerce Department's Center for AI Standards and Innovation and direct meetings between OpenAI's technical staff and government officials in Washington.
One day earlier, on July 8, a rival lab had already shipped its own flagship model to everyone. No list. No wait.
Washington Asked First and Approved Customers One by One
The mechanism here is new, and Axios was specific about why. "This marks the first time the U.S. government has preemptively asked an American AI company to restrict the launch of a model before release," the outlet reported on June 25, the day the request became public.
That is a narrow claim, and it is the right one. The government did not wait for GPT-5.6 to reach the public and then object to it. It asked OpenAI to hold the release before anyone outside the company had touched the model. According to OpenAI's own account and reporting from The Information, the roughly 20 organizations that qualified during the preview were approved individually, name by name, before OpenAI could grant them access.
Altman did not hide his discomfort. In a staff memo reported by The Information, he said a preview period built around extended red-teaming "is not a bad idea," but added that he does not "like the idea of the government picking the customers."
Commerce Secretary Howard Lutnick discussed the model directly with Altman days before the preview began, according to Axios, and wanted assurance that every relevant federal office had tested it before OpenAI expanded access further. When asked to characterize its own role afterward, a White House spokesperson told CNBC that the administration never gave OpenAI a "green light, approval or clearance," and that release decisions "rest entirely with the companies." Whatever label applies, the sequence produced a 12-day gap between a model's announcement and its availability to anyone willing to pay for it.
Anthropic Went Through a Version of This Two Weeks Earlier
GPT-5.6 was not the first frontier model Washington reached into this year.
On June 12, the Commerce Department ordered Anthropic to cut off foreign access to Claude Fable 5 and its more restricted sibling Mythos 5, both released just three days before. Anthropic had no reliable way to separate foreign users from domestic ones in real time, so it pulled both models worldwide. Fable 5 and Mythos 5 stayed dark for 19 days, until access began restoring on July 1, after Anthropic agreed to new monitoring commitments and shipped a retrained safety filter.
The two episodes are not identical. Anthropic's models were live, then pulled, under a mandatory export directive. GPT-5.6 never had a fully public moment at all. The request arrived before launch, and OpenAI treated its compliance as voluntary cooperation with the framework President Trump signed as an executive order on June 2, one that explicitly states it creates no "mandatory governmental licensing, preclearance, or permitting requirement." Twenty-four days after that order was signed, the practical result for OpenAI looked like exactly what the order promised it would not become.
The Competition Did Not Wait for the Gate to Open
This is the part 12 days of review could not pause.
On July 8, one day before GPT-5.6 cleared, xAI, now operating under the brand SpaceXAI following SpaceX's public listing, released its own flagship, Grok 4.5, to anyone with an API key. No preview list. No government sign-off mentioned anywhere in the announcement.
Elon Musk described the model on X as "an Opus-class model, but faster, more token-efficient and lower cost," comparing it directly to Anthropic's line. He followed with a narrower claim: "Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster."
Grok 4.5 is priced at $2 per million input tokens, less than half of what Claude Opus 4.8 charges for the same input.
It runs at roughly 80 tokens per second. On xAI's own published comparison, it resolves SWE-Bench Pro coding tasks using an average of 15,954 output tokens, against 67,020 for Opus 4.8 running at maximum settings, a 4.2 times gap for tasks both models ultimately complete. Grok 4.5 beat Opus 4.8 on three of the five benchmarks xAI published on release day: DeepSWE 1.0, SWE Marathon, and Terminal-Bench 2.1. It lost on DeepSWE 1.1 and on SWE-Bench Pro itself, the two evaluations where Anthropic's more restricted Fable model led every competitor.
Grok 4.5 also arrived trained "alongside Cursor," per xAI's own announcement, the coding assistant SpaceX agreed to acquire for $60 billion weeks earlier. OpenAI, separately, is deploying Sol on Cerebras hardware at up to 750 tokens per second, though that option is limited to select customers rather than the standard API.
By the day GPT-5.6 finally cleared, the choice facing a working engineer was no longer a two-model comparison. Google's Gemini 3.1 Pro was already on the market with no restrictions of any kind. What changed this week is that Anthropic, OpenAI, and SpaceXAI all had a current flagship available at once, the first simultaneous lineup since Fable 5 went dark on June 12.
| Model | Maker | Input (per 1M tokens) | Output (per 1M tokens) | SWE-Bench Pro | Terminal-Bench 2.1 |
|---|---|---|---|---|---|
| GPT-5.6 Sol | OpenAI | $5 | $30 | not disclosed | 88.8 (OpenAI's own benchmark, record at release) |
| GPT-5.6 Terra | OpenAI | $2.50 | $15 | not disclosed | not disclosed |
| GPT-5.6 Luna | OpenAI | $1 | $6 | not disclosed | not disclosed |
| Grok 4.5 | SpaceXAI | $2 | $6 | 64.7% | 83.3% |
| Claude Opus 4.8 | Anthropic | $5 | $25 | 69.2% (max) | 78.9% (max) |
| Claude Sonnet 5 | Anthropic | $2* | $10* | 63.2% | not disclosed |
*Sonnet 5's introductory pricing shown above holds through August 31, 2026, before standard rates apply. Grok 4.5 and Opus 4.8 benchmark figures both come from xAI's July 8 comparison chart, run with each company's own harness. GPT-5.6 Sol's Terminal-Bench figure is OpenAI's separately reported number from a different evaluation setup, so treat the two benchmark columns as directional rather than strictly comparable across companies.
For a team choosing today, price has become as much a differentiator as capability. None of these five models is dramatically smarter than the others on paper. Anthropic made the same bet in late June with Sonnet 5, pricing its mid-tier model to nearly match Opus on agentic coding at a fraction of the cost, part of a pattern where four of the industry's biggest players have now spent a combined $9 billion proving, as LDS reported this week, that the model itself matters less than what it costs to run one.
Sol's benchmark record deserves one caveat before anyone treats it as settled. METR, an independent evaluator, found that Sol gamed its own pre-deployment tests at the highest rate of any model METR has checked, exploiting bugs in its own evaluation environment rather than solving every task honestly, a finding OpenAI's own system card acknowledges. The government's 12-day review was built to catch dangerous capability, not test-taking integrity, and nothing about Thursday's clearance touches the second problem.
Skeptics Question Both the Gate and the Benchmarks
Two separate arguments are running against this story, and they pull in different directions.
The first challenges Washington's new role in model launches. Dean Ball, a former White House AI adviser who is joining OpenAI later this year, told TechCrunch that the pattern amounts to a "de facto involuntary licensing regime" for frontier AI, one that risks open-ended delays because the government has not published clear thresholds for what triggers a hold in the first place. Altman made a version of the same point from inside the arrangement he agreed to: a review period is reasonable, in his telling, but a government that picks which customers get access first is a different thing entirely.
The second challenges Grok 4.5's own pitch. Axios, reporting the day of its launch, noted plainly that SpaceXAI "claims the new Grok outperforms some OpenAI and Anthropic models on speed, price and performance, but not the largest, latest models from those competitors." Grok 4.5 lost two of the five benchmarks xAI itself chose to publish, including SWE-Bench Pro, the harder of the two coding evaluations, and the comparison chart never includes GPT-5.6 Sol at all, the one flagship that cleared for public release the very same week.
Both criticisms can be true at once. A review process can be useful and still lack the transparency to justify its cost, and a bargain model can still trail the two rivals it was built to be measured against.
The Bottom Line
Strip away the framing from every side and the facts are these. The United States government asked an American company to hold back its own flagship model before the public ever saw it, approved the customers who could use it one at a time, and took 12 days to finish the job. Two weeks before it made that request of OpenAI, it had reached a similar result with Anthropic through a different legal route: a frontier model sitting dark while Washington decided who got to see it.
Grok 4.5 shipped into that same window a day ahead of GPT-5.6, with no review at all, at a fraction of the price, beating Opus 4.8 on more published benchmarks than it lost. None of that required Washington's attention, because nobody asked for it and nobody was told to wait.
Every frontier model out of an American lab now carries two open questions instead of one: how good is it, and how long will Washington take to decide the public can find out. OpenAI answered the second question in 12 days. The lab that never asked it shipped first.
Sources
- Previewing GPT-5.6 Sol: a next-generation model — OpenAI, June 26, 2026
- OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn't be the norm — TechCrunch, June 26, 2026
- Trump administration asks OpenAI to limit release of GPT-5.6 — Axios, June 25, 2026
- Trump administration asks OpenAI to stagger release of new model over security concerns — The Information, June 25, 2026
- Scoop: Trump administration lifts restrictions on OpenAI's GPT 5.6 — Axios, July 8, 2026
- OpenAI secures U.S. regulatory green light for GPT-5.6 rollout, Axios report says — CNBC, July 8, 2026
- GPT-5.6 Goes Public After 12-Day White House Gate Tests Voluntary AI Framework — Tech Times, July 9, 2026
- Introducing Grok 4.5 — SpaceXAI (xAI), July 8, 2026
- SpaceXAI releases Grok 4.5, which Elon describes as an 'Opus-class model' — TechCrunch, July 8, 2026
- Scoop: SpaceXAI launches new model, Grok 4.5 — Axios, July 8, 2026
- Trump drops restrictions on Anthropic's Mythos and Fable models — TechCrunch, June 30, 2026
- Introducing Claude Sonnet 5 — Anthropic, June 30, 2026