OpenAI released its GPT-5.6 model lineup last week, but only to a handful of partners pre-approved by the U.S. government. The company made it clear this arrangement is not acceptable long-term, calling it a barrier that "keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

Source: TechCrunch.

The restricted rollout follows pressure from the Trump administration on AI companies to submit advanced models for government review before public release. A recent executive order requests voluntary pre-release submissions up to 30 days before launch. According to OpenAI, the administration specifically asked the company to limit GPT-5.6 access to partners "whose participation has been shared with the government."

The GPT-5.6 family includes three models: Sol, the flagship; Terra, a balanced mid-tier option; and Luna, optimized for speed and lower cost. All three remain under access restrictions despite varying capability levels.

What Happened With Anthropic

OpenAI is not the first company to face these restrictions. When Anthropic released its Fable 5 model earlier in June, the administration ordered the company to block access for any foreign national. Rather than implement selective blocking, Anthropic pulled the model entirely.

Anthropic's Mythos 5 model also faced effective prohibition from the administration this month. The repeated interventions have raised questions about how much control the government should exercise over commercial AI releases.

Dean Ball, a former White House AI adviser now joining OpenAI, argues the voluntary review process has become an involuntary licensing regime. He points out the government has not established clear safety standards, which could lead to indefinite delays. Those delays carry consequences beyond product timelines, potentially affecting billions in AI infrastructure investment and handing advantages to Chinese competitors.

GPT-5.6 Sol Capabilities and Pricing

OpenAI positions Sol as its most capable model to date, with improvements in agentic work across coding, biology, and cybersecurity. The model introduces two new reasoning modes: "max" for intensive problem-solving and "ultra" which coordinates multiple subagents for complex tasks. That ultra mode will drive up token costs fast.

On coding benchmarks, OpenAI claims Sol performs slightly better than Anthropic's Claude Mythos 5. The company also says Sol matches Mythos preview performance while using one-third the output tokens.

Pricing reflects the tiered approach:

• Sol: $5 per million input tokens, $30 per million output tokens

• Terra: Half of Sol's rates

• Luna: $1 input, $6 output per million tokens

OpenAI has also improved prompt caching to reduce costs for repeated queries.

Security Design Choices

OpenAI says it built security measures directly into Sol's core behavior rather than layering filters on top. This approach aims to avoid the problems Anthropic encountered with Fable 5.

When Fable 5 was briefly available, its safety classifiers would detect high-risk topics like cybersecurity or chemistry and route requests to an older, less capable model without telling users. This invisible downgrade caused false positives and user complaints before the model was pulled.

OpenAI claims Sol is hardened against adversarial attacks and optimized to favor defensive cybersecurity work over offensive exploits. In practical terms, the model should resist jailbreaking while helping users understand how to defend systems rather than how to compromise them.

Bottom Line

OpenAI plans broader availability for GPT-5.6 in the coming weeks as it negotiates a new framework with the administration. For now, business operators cannot access the models unless they are part of the approved partner group. The company's public pushback signals it will resist making government pre-clearance the standard process for future releases. Whether that resistance carries weight depends on how the administration responds and whether other AI companies join the objection.