OpenAI is restricting the release of GPT-5.6 to a small group of trusted partners after a request from the U.S. government, the company said Friday, and used the same blog post to argue the arrangement should not become permanent. The limited preview covers all three new models: Sol, the flagship; Terra, the mid-tier; and Luna, the faster low-cost option. OpenAI said the preview is limited to partners whose participation has been shared with the administration. It is the second time in a month that Washington has gated a frontier model release.
The structure traces back to a recent Trump executive order asking certain AI companies to voluntarily submit their most advanced models for government review up to 30 days before release. Dean Ball, a former White House AI adviser who is joining OpenAI, has argued the order has created a de facto involuntary licensing regime for frontier AI. Ball's concern, laid out publicly, is that the government has not defined the safety standards it is reviewing against, which leaves release timelines open-ended.
OpenAI complied with the request but used Friday's announcement to register dissent.
“We don't believe this kind of government access process should become the long-term default,”— OpenAI, company blog post
Key facts
- 01OpenAI is limiting GPT-5.6 Sol, Terra, and Luna to a small group of partners after a U.S. government request.
- 02A recent executive order asks AI companies to submit frontier models for review up to 30 days before release.
- 03GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output tokens; Luna runs $1 and $6.
- 04OpenAI says Sol matches Anthropic's Claude Mythos 5 preview on coding while using one-third of the output tokens.
- 05Anthropic pulled Fable 5 entirely this month after the administration ordered foreign-national access be removed.
The company framed the limited preview as a short-term step on the path to broader availability in the coming weeks, while it works with the administration on a new cybersecurity executive order framework and what OpenAI called a repeatable process for future releases. The pricing has already been set for general availability: GPT-5.6 Sol runs $5 per million input tokens and $30 per million output tokens, Terra runs $2.50 and $15, and Luna comes in at $1 and $6. OpenAI also said it has improved prompt caching so repeated prompts come out cheaper and more predictable.
The context for the government's posture is Anthropic's Fable 5, released earlier and then pulled entirely this month after the administration ordered the company to remove access for any foreign national. In the brief window Fable 5 was live, its classifiers routed high-risk prompts on cybersecurity, biology, or chemistry to an older model rather than blocking outright, which produced false positives and user complaints. The administration also effectively banned Anthropic's Claude Mythos 5 this month, per OpenAI's own framing of the competitive field.
“It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”— OpenAI, company blog post
OpenAI is clearly trying to avoid Anthropic's outcome. The company said GPT-5.6 Sol's safety guardrails are built into the core model's behavior rather than bolted on as a separate filter, which is the design choice that caught Fable 5 out. Sol is hardened against adversarial attacks and optimized to favor defensive cybersecurity work over offensive exploits, meaning the model is tuned to help users defend systems rather than break into them. That positioning is aimed squarely at the review process the model has to clear.
On capability, OpenAI claims Sol is its strongest model to date, with gains in coding, biology, and cybersecurity agentic work. The model introduces a max reasoning effort mode and an ultra mode that coordinates subagents on highly complex tasks, with the token bill that implies. OpenAI says Sol slightly outperforms Claude Mythos 5 on coding workflows and is competitive with the Mythos preview while using one-third of the output tokens, which is the more interesting number for anyone paying API bills.
The broader access plan still points to ChatGPT, Codex, and the API once the administration signs off. That puts the eventual availability of Sol, Terra, and Luna on the same gating path that took Fable 5 off the market, which is the part OpenAI is implicitly asking the administration to fix. The company's argument is that withholding the best models keeps them out of the hands of cyber defenders and enterprise users who need them, not just the foreign-access cases the orders are designed to address.
There is a counter-case Ball himself has laid out: without defined safety standards, reviews can stretch indefinitely, which delays not just releases but the billions in AI infrastructure spending tied to model availability. That argument cuts the other way on China competition, where Washington's stated goal of maintaining a U.S. lead works against open-ended pre-release holds. Neither OpenAI nor the administration has published the standards the 30-day review is being run against, which leaves developers and enterprise buyers without a timeline they can plan to.
The pattern emerging across GPT-5.6 and Fable 5 is that the executive order has shifted launch dynamics for frontier U.S. labs from a product decision to a regulatory one, with the labs setting prices and benchmarks ahead of a release date they no longer fully control. For OpenAI, the immediate cost is a narrower launch window for its highest-margin product; the longer cost, if the process sticks, is that the API tier that generates revenue is gated on a review with no published clock. Expect both OpenAI and Anthropic to spend the next quarter trying to convert the current ad hoc reviews into a written framework with deadlines attached, because the alternative is a frontier-model market where pricing sheets are public months before the product is.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
