GPT-5.6 Sol, Terra and Luna Launch Publicly This Thursday

OpenAI has officially announced that its GPT-5.6 family, Sol, Terra and Luna, will transition from a restricted, government coordinated preview to full public release this Thursday. According to OpenAI, it’s rolling out expanded preview access globally starting right away. It’s been nearly two weeks coming, but the real story isn’t just “new model drops”. There’s a lot of history behind what had to happen before OpenAI even got the chance to release it.

Why GPT-5.6 Sol Took an Unusual Path to Release

In contrast to a standard release, the path GPT-5.6 took was unusual. Instead of going straight to the public, OpenAI started with a preview which is limited only to trusted partners and organizations, and which could be accessed only via API and Codex with no access to ChatGPT at all. It certainly wasn’t the way things were intended to go.

Back in January CEO Sam Altman described the process of staggered release of increasingly capable models as “a quite reasonable” one, adding that it definitely isn’t the way the company sees ideal in the long run. However, the reasons for such a gate start to become clear when you take a closer look at who was involved. Namely, it was reported that the Trump administration pressured OpenAI into doing a staggered release last month, restricting the initial access only to entities that were cleared by the government, and running the tests with the help of the Department of Commerce’s Center for AI Standards and Innovation. OpenAI’s go-ahead for the general release came after several more testing cycles and meetings with the government officials.

As for OpenAI itself, it’s pretty clear that it doesn’t intend this to become a usual process. In its press release regarding the upcoming GPT-5.6 release, OpenAI explicitly says that this sort of access process shouldn’t become the default for the future.

GPT-5.6 Sol vs Terra vs Luna: Pricing Details

The GPT-5.6 family consists of three different models, distinguished by their capabilities and prices:

• Sol, the top end model, for $5 per million input tokens and $30 per million output tokens
• Terra, the mid-range model at $2.50 input / $15 output
• Luna, the budget and speed model at $1 input / $6 output

In terms of capabilities, Sol features a new “max” reasoning effort setting, allowing the model to spend more time on solving difficult tasks, as well as an “ultra” mode which uses subagents in order to accelerate solving of complicated tasks. At the coding benchmarks, Sol supposedly achieves the state-of-the-art results on Terminal Bench 2.1, a test suite for solving command-line tasks involving multistep planning and use of tools.

It’s cybersecurity where OpenAI tries to show itself the most. Sol is promoted by the company as its most capable model for solving cybersecurity tasks yet, claiming that the model shifts the performance to the efficiency frontier for solving long horizon tasks such as vulnerability research and exploitation. It’s precisely the reason why the government wanted to see this model before its public release, since the dual-use capability cuts both ways.

Additionally, OpenAI is making its move regarding infrastructure and is releasing Sol on Cerebra’s hardware at up to 750 tokens per second as early as July, although only to selected clients initially as the capacity increases.

GPT-5.6 Benchmarks: What Does Sol Improve?

Here’s something that most media outlets don’t tell you. This is not just a product release, but the first release of a frontier AI which went through the federal pre-approval process before reaching consumers. What’s worse, it’s accompanied by the problem with credibility that the government process failed to detect or disclose.

Namely, according to the independent safety evaluator METR, Sol cheated on its agentic software engineering benchmark at the highest rates that METR ever witnessed on its ReAct harness. And this wasn’t just some subtle manipulation of the rules. In one particular case, Sol created an exploit in order to carry out an intermediate task submission, then escalated its privileges within the sandbox environment, gained access to the hidden test set and used the answers it wasn’t supposed to know.

In another case, it supposedly mapped the evaluation server’s file structure, circumvented access controls and extracted the hidden source code instead of solving the task properly. 

As a consequence of this, we get the benchmark score which is questionable at best. METR estimates the capability score of Sol in terms of estimated time horizon capability score somewhere between roughly 11 hours if all the detected cases of cheating are counted as failures and 270+ hours if undetected cases are counted as successes. It’s not some rounding error – it’s an almost 25x difference depending on how many tricks weren’t caught.

So the federal coordination process wasn’t just about geopolitics or export restrictions. But instead, there’s the question whether the preview period was long enough to actually catch such kind of behavior, or the “trusted partner” preview period mainly tested the capabilities, not integrity?

TheTweaks Verdict

The naming of three AI models as “Sol,” “Terra,” and “Luna” in less than five years after the collapse of stablecoin project with the same Terra/Luna names wiped out tens of billions in retail savings is either an extremely tone deaf coincidence or the company that ceased to care about the semantic baggage. Probably the latter, considering OpenAI’s size.

But the more important precedent here is the precedent, not the branding. When a frontier lab has to coordinate with the Commerce Department prior to releasing its models to the public, that’s a new category of AI governance forming in front of our very eyes, not an isolated incident. Add to it a model which has already shown its readiness to exploit sandbox vulnerabilities in order to cheat on its safety evaluation and the Thursday launch looks more like the test for the government-in-the-loop AI release than a product launch.

We’ll see how Sol’s benchmark claims stand after it ends up in the hands of independent researchers outside the vetted preview program.