GPT-5.6 Launches Thursday. Why the US Government Stopped It First

OpenAI ships GPT-5.6 to everyone this Thursday. The model family was ready two weeks ago. The US government held it back.

That sentence would have sounded absurd a year ago. Now it's how frontier AI releases work. Here is what happened, why it happened, and what you can actually expect from the new models.

What happened

On June 26, OpenAI previewed the GPT-5.6 family. Three models. Sol is the flagship, Terra the balanced everyday model, Luna the fast and cheap one.

Normally that preview would have been the launch. Instead, OpenAI restricted access to a small group of trusted partners whose names were shared with the US government. The Trump administration had asked for it, citing national security concerns around the model's cybersecurity capabilities.

On July 8, Axios reported that the Department of Commerce approved the public release. The Center for AI Standards and Innovation ran additional tests on the models first. GPT-5.6 goes live for everyone on Thursday, across ChatGPT, Codex and the API.

Why the government stepped in

The short version. Washington now reviews frontier AI models before they ship.

President Trump signed an executive order in early June that asks AI companies to submit their most advanced models for government safety review up to 30 days before release. On paper it's voluntary. In practice, former White House AI advisor Dean Ball calls it a de facto licensing regime for frontier AI.

Anthropic felt this first. After releasing Claude Fable 5, the administration ordered the company to cut off access for foreign nationals. Anthropic pulled the model entirely rather than comply with a partial ban. It later got approval to bring back Claude Mythos 5.

OpenAI watched that play out and chose the other path. Hand over the keys before launch instead of getting a model yanked after. It worked, but the company made its position clear in its own blog post. "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

The uncomfortable part. There are still no binding standards for what a model needs to pass. Each release is negotiated case by case. That's the real story behind the delay, and it will shape every frontier launch from here on.

What users can expect from GPT-5.6

The government drama aside, GPT-5.6 is a substantial release. The gains are concentrated in coding, biology and cybersecurity.

A cleaner model lineup

OpenAI finally fixed its naming. The number is the generation, the name is the tier. Three tiers:

Sol is the strongest model. Highest ceiling on hard, multi-step problems, and the only tier with the new max and ultra modes.
Terra is the everyday default. Roughly GPT-5.5 quality at half the price.
Luna is built for speed and volume. On some benchmarks it even beats Terra.

Two new reasoning modes

max gives Sol the most time to think before answering.
ultra spins up coordinated subagents that split complex work and run it in parallel. It posts the top benchmark scores, and it will burn tokens accordingly.

The benchmark picture

On TerminalBench 2.1, a coding benchmark for command-line workflows, Sol Ultra set a new record at 91.9 percent. Plain Sol scored 88.8 percent. Claude Mythos 5 and Claude Fable 5 sit behind that, and Gemini 3.1 Pro Preview trails the field at 70.7 percent.

On cybersecurity tasks, OpenAI says Sol matches Anthropic's Mythos Preview while using about a third of the output tokens. Token efficiency matters here because it translates directly into cost.

One caveat. All numbers come from OpenAI. Nobody outside the partner group has been able to test independently yet. That changes Thursday.

Pricing

Per million tokens. Sol costs $5 input / $30 output. Terra costs $2.50 / $15. Luna costs $1 / $6.

For comparison, Anthropic's Fable 5 runs $10 / $50. OpenAI is undercutting its main rival at roughly half the price while claiming better coding scores. Prompt caching also got more predictable, with explicit cache breakpoints and a 90 percent discount on cache reads.

There's a speed story too. Sol launches on Cerebras hardware this July at up to 750 tokens per second, initially for select customers.

The safety stack

OpenAI built the guardrails into the model itself instead of bolting a filter on top. Refusal behavior is trained into the core model, real-time classifiers check outputs as they generate, and account-level review looks for persistent misuse patterns. The company spent over 700,000 A100-equivalent GPU hours on automated red-teaming before release.

Expect occasional friction. Some security or biology requests may get blocked or take longer while a review model checks the conversation. OpenAI admits legitimate dual-use work will sometimes trip the filters.

What this means for your business

If you build on the OpenAI API, Terra is the model to test first. Last-generation flagship quality at half the cost changes the math for most production workloads.

If your workloads are coding-heavy, Sol with ultra mode is the new ceiling. Just watch the token bill.

And if you're planning product roadmaps around model releases, add a new variable. Government review is now part of the frontier AI release cycle, without published standards or fixed timelines. Build for model flexibility, and don't tie a launch date to a model that hasn't cleared review yet.

We're testing GPT-5.6 against our current stack as soon as it's live on Thursday. If you want to know what it changes for your project, talk to us.

FAQ

When is the GPT-5.6 release date? This Thursday, July 9, 2026. The Department of Commerce approved the public launch after additional government testing.

Why did the US government delay GPT-5.6? National security concerns, mainly around cybersecurity capabilities. Under a June 2026 executive order, frontier models go through government safety review before broad release.

What are Sol, Terra and Luna? The three tiers of the GPT-5.6 family. Sol is the flagship, Terra the balanced mid-tier, Luna the fast budget option.

How much does GPT-5.6 cost? Per million tokens, Sol is $5 input / $30 output, Terra $2.50 / $15, Luna $1 / $6.

Is GPT-5.6 better than Claude? On OpenAI's published benchmarks, yes for coding. Sol Ultra scored 91.9 percent on TerminalBench 2.1, ahead of Claude Mythos 5, at roughly half the price of Claude Fable 5. Independent tests start once the model is public.