GPT-5.6 lands as a three-tier model family, cleared by an opaque government review

OpenAI released GPT-5.6 on July 9, and the shape of the launch says as much as the model does. Rather than a single flagship, the company shipped three named tiers, Sol, Terra, and Luna, each priced for a different budget, with Sol positioned as its strongest coding model to date. The other notable detail is procedural: GPT-5.6 reached the public only after a federal safety review whose criteria remain undisclosed, even to some of the people asked to weigh in on it.

What shipped

GPT-5.6 splits into three variants that trade capability against cost. Sol is the workhorse at $5 per million input tokens and $30 per million output, and it carries the headline claim: OpenAI puts it at the top of the Coding Agent Index at a score of 80, about 2.8 points above Anthropic’s Fable 5, while using less than half the output tokens on the same tasks. Terra sits in the middle at $2.50 and $15, and Luna is the budget option at $1 and $6. All three are available now in ChatGPT, in Codex, and through the API, so the coding gains land directly in the agent workflow most teams already run.

OpenAI also leaned on security, calling GPT-5.6 its strongest cybersecurity model yet and pointing at threat modeling, code review, patching, and blue-team work as target use cases. Alongside the model, the company introduced ChatGPT Work, an enterprise surface, and named GPT-5.6 the preferred model inside Microsoft 365 Copilot. The same week, OpenAI began sunsetting its Atlas browser and folding those agentic-browsing features back into ChatGPT, a sign it is consolidating surfaces rather than spreading them.

The government review that stays opaque

The more consequential part of this launch may be how it cleared. GPT-5.6, and Sol in particular, went through a federal review before release, a process the administration had earlier signaled it might use to restrict the rollout over misuse concerns. The Department of Commerce’s Center for AI Standards and Innovation appears to be leading, with six cabinet agencies instructed to settle a final process by early August. Officials named in connection with the review include Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and National Cyber Director Sean Cairncross.

What the review actually tested is unclear. OpenAI previewed Sol with government officials and select users ahead of the public release, but the evaluation methods and pass criteria have not been disclosed. “Nobody knows what the requirements are to get licensed,” said Dean W. Ball, a former administration policy adviser now at OpenAI. A researcher at Georgetown’s Center for Security and Emerging Technology was blunter, saying there was not enough visibility into the process to judge whether it was adequate. For a working developer, the takeaway is not the politics but the precedent: frontier model availability is starting to route through a government gate whose rules are still being written.

Where this lands for developers

The three-tier structure is the same bet the rest of the market has been making, compressed into one release. Grok 4.5 and Claude Sonnet 5 both arrived recently arguing that most production work lives below the frontier and should be priced accordingly. GPT-5.6 makes that argument internally: Luna at $1 and $6 is cheap enough to leave running in agent loops and background jobs, while Sol is reserved for the runs where coding quality justifies flagship rates. Teams that route requests by task can now do so without leaving the OpenAI stack.

The coding claim is the part worth testing rather than trusting. A state-of-the-art score paired with lower token use is exactly the combination that changes real bills, but an index number and production behavior are not the same thing. The honest read is that Sol looks competitive with Anthropic’s current lineup on OpenAI’s own charts, which is a reason to benchmark it against your own workload, not a reason to switch on the strength of a launch post.

What’s worth watching

1. Independent coding benchmarks. The Coding Agent Index result is OpenAI’s own. Whether third-party evaluations reproduce Sol’s lead over Fable 5, and the token-efficiency claim, is the signal that matters for anyone paying per token.
2. The review as a template. If the Commerce-led process becomes the standard path for frontier releases, it will shape when and how models reach developers. The early-August deadline for a defined process is the near-term thing to track.
3. Codex adoption. Sol’s coding pitch only pays off if it shows up as better agent runs in Codex. Watch whether teams see fewer tokens and cleaner diffs, or just a new model id.

The plain read is that OpenAI matched the market’s pricing move and stapled a coding claim to the top of it, while quietly setting a precedent for how models clear to release. Stackmaven will revisit GPT-5.6 once independent coding benchmarks and the federal review’s defined process are on record, on or around October 8.

Sources cited

1. OpenAI: GPT-5.6 openai.com
2. TechCrunch: OpenAI launches its new family of models with GPT-5.6 techcrunch.com
3. TechCrunch: How did the government decide OpenAI's frontier model was safe to release? techcrunch.com