OpenAI Releases GPT-5.6 in Limited Preview, Complies With Trump Administration Rollout Restrictions While Objecting to Them

Since the Trump administration's AI executive order earlier this month created a voluntary-but-not-really pre-release review process, the U.S. government has now effectively shaped two major AI product launches — Anthropic's and now OpenAI's — in ways that neither company requested.

On Friday, OpenAI unveiled GPT-5.6, a three-model suite named Sol, Terra, and Luna. Sol is the flagship. Terra handles "high-volume work" at roughly half Sol's cost. Luna is the budget option, priced below Terra. According to The Verge, Sol's pricing lands at $5 per million input tokens and $30 per million output tokens — nearly half the cost of Anthropic's competing Claude Fable 5 model, which runs $10 input and $50 output.

The release is a limited preview. OpenAI said it is only available to a "small group of trusted partners whose participation has been shared with the government," according to CNBC. The company did not disclose which partners are included.

What the Models Actually Do

OpenAI says Sol is its most capable model to date, with particular strength in coding, cybersecurity, and biology. The Verge reported that Sol also introduces two new operating modes: "max" for deeper reasoning, and "ultra" for deploying coordinated sub-agents on complex tasks. TechCrunch noted Sol outperforms Anthropic's Claude Mythos 5 on coding benchmarks, a model the Trump administration also effectively banned this month.

On the safety side, OpenAI dedicated the bulk of its Friday blog post to misuse risks. It claimed Sol is better at helping users find and fix vulnerabilities than at executing end-to-end attacks, and said the model does NOT cross the company's internal "critical" cybersecurity risk threshold, defined as providing "unprecedented new pathways to severe harm." OpenAI also said it committed roughly 700,000 A100e GPU hours to automated red-teaming, with third-party testing ongoing for at least two more weeks, according to The Verge.

The Government's Hand in the Release

The restrictions are not incidental. CNBC reported that two weeks ago Anthropic was ordered to disable access to two of its newest models under an export control directive, removing them entirely after finding it could not comply while keeping foreign nationals on the platform. Anthropic's models remain down as of June 26, with no public timeline for return.

The Trump administration's AI executive order, signed earlier this month, nominally asks for voluntary pre-release model reviews. In practice, as TechCrunch reported, Dean Ball, a former White House AI adviser who is joining OpenAI, argues this has become a de facto involuntary licensing regime for frontier AI. Ball's concern: without defined safety standards, the process could produce endless launch delays, hand an advantage to China, and threaten the billions already committed to U.S. AI infrastructure.

OpenAI's own blog post expressed the same discomfort more diplomatically. "We don't believe this kind of government access process should become the long-term default," the company wrote, according to CNBC. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

The company framed the restricted preview as a "short-term step" and said it is working with the Trump administration to develop a repeatable process for future releases. Friday's arrangement is explicitly not supposed to be permanent.

The Strongest Case for Government Review

The administration's position deserves a fair hearing. Frontier AI models with advanced cybersecurity capabilities represent a new risk category. A model capable of meaningfully accelerating biological research or identifying exploitable software vulnerabilities is not the same category of product as a search engine or a word processor. If the government has legitimate national security equities and classified threat assessments that the public cannot evaluate, a short pre-release review period is not obviously unreasonable. The question is whether "voluntary" review can remain functional without devolving into indefinite delay, and whether there is any defined threshold at which a model gets cleared.

Neither OpenAI nor the administration has publicly described what a completed review process looks like, or what would constitute approval. This is the unresolved structural problem Ball identified, and it has NOT been answered by Friday's launch.

An Industry Under Cost Pressure

Separate from the regulatory friction, OpenAI is releasing GPT-5.6 into a market that is actively pushing back on AI spending. CNBC reported that the era of "tokenmaxxing," where developers were incentivized to consume as much AI compute as possible, is winding down. Uber, for example, burned through its entire annual AI budget in four months, according to information revealed to The Information by Uber CTO Praveen Neppalli Naga in April. Uber has since capped individual AI tool spending at $1,500 per month as a baseline.

Lindy CEO Flo Crivello told CNBC he moved his company entirely off Anthropic's Claude models to DeepSeek, a Chinese open-weight alternative, to cut costs, calling it "a matter of survival."

Gil Luria, an equity analyst at D.A. Davidson, told CNBC that the current growth rates for OpenAI and Anthropic are "the fastest they will ever be" and that both companies may be timing their confidential IPO filings, both submitted in early June, ahead of enterprise clients pulling back on spending.

The pricing on GPT-5.6 reflects the pressure. Sol at $5/$30 per million tokens is a competitive move against Anthropic at $10/$50. If the limited preview expands to general availability in the coming weeks as OpenAI promises, the cost comparison will matter.

What remains genuinely open: whether the Trump administration will codify a clear, time-limited review standard before its current case-by-case approach causes another company to simply pull a product offline the way Anthropic did.