OpenAI made its GPT-5.6 model family — Sol, Terra, and Luna — available to the general public on July 9, ending roughly two weeks in which access was limited to government-approved enterprise customers. The wider release followed a review by the Department of Commerce's Center for AI Standards and Innovation (CAISI), which tested the models for safety, security, and standards alignment before granting broader clearance, according to reporting from Axios that CNBC and Engadget corroborated on July 8.

From government gate to open access

AgentsAI covered the original restriction on June 26: at the request of the Office of the National Cyber Director and the Office of Science and Technology Policy, OpenAI agreed to stagger GPT-5.6's rollout so federal officials could approve enterprise customers one by one during a preview period, citing the model's advanced cybersecurity-relevant capabilities. CEO Sam Altman said at the time that OpenAI viewed the arrangement as temporary and would push for a "more sustainable approach" to future releases. That process concluded after CAISI conducted additional testing and held further meetings with OpenAI, which the company says kept technical staff in Washington, D.C. to answer questions during the review.

What's actually shipping

The GPT-5.6 family launches in three tiers: Sol, OpenAI's flagship and most capable model; Terra, a mid-cost option OpenAI says performs close to GPT-5.5 at roughly half the price; and Luna, the fastest and cheapest tier. Per-million-token API pricing is $5 input / $30 output for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. OpenAI says the family advances performance on software engineering, computer use, professional knowledge work, scientific research, and cybersecurity tasks, and adds more predictable prompt caching with explicit cache breakpoints and a 30-minute minimum cache life.

The agentic angle

Sol ships with an optional "Ultra" mode that moves beyond a single reasoning chain: it decomposes a request and spawns parallel subagents that work on different parts of the task before a synthesis step recombines their output, at a token cost of several times a standard call. OpenAI cites Terminal-Bench 2.1, a benchmark for command-line, tool-using agentic workflows, as evidence of the approach's payoff.

Why the precedent matters

Beyond the model itself, the episode is a data point in a bigger story: a frontier lab's release timeline was directly shaped by a federal review process rather than the lab's own staged rollout. Whether CAISI screening becomes a standing step for future frontier releases — from OpenAI or its rivals — remains unresolved, but GPT-5.6's path from restricted preview to general availability is the first full run of that process from end to end.