GPT-5.6 goes public: three models, three price tags, one story about government oversight

OpenAI will make the GPT-5.6 model family available to the general public this Thursday, which it first unveiled back at the end of June. Back then, however, only a handful of vetted partners got access — and the reason was unconventional. The U.S. government requested time to assess safety risks. The Trump administration signed an executive order in June that allows developers of "frontier" models to offer them to the government for up to 30 days of review before broader release. OpenAI has long disagreed with this approach — "we do not believe this government approach should become the standard," the company wrote in its announcement — but went along with it this time, saying it was the fastest path to public availability. The result is a trio of models with a new naming system. Sol is the flagship — the most powerful model OpenAI has ever built. On the Terminal-Bench 2.1 benchmark, which tests the ability to manage complex command workflows, it sets a new record. On ExploitBench, measuring cybersecurity capabilities, Sol is competitive with Anthropic Mythos — while consuming just a third of the tokens. It also comes with new reasoning modes: max gives the model the most time for deep analysis, ultra deploys sub-agents for parallel solving of complex tasks. Terra is the mid-range — according to OpenAI, it delivers performance comparable to GPT-5.5 but at half the price. And Luna is the cheapest and fastest variant for everyday tasks.

How much it costs

Prices per 1 million tokens (input/output): Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6. By comparison — GPT-5.5 cost $3.75/$15 per million tokens. Terra thus offers comparable performance at a significantly lower price, which is good news for companies working with AI at scale.

Safety first

OpenAI invested over 700,000 GPU-hours (converted to A100) in automated red-teaming testing for GPT-5.6. The goal? To find universal jailbreaks — attacks that work across different contexts, not just in one narrow scenario. The model has a multi-layer protection system: model-level refusal, real-time classifiers, account-level checks, and differentiated access based on sensitivity of capabilities.

xAI and Cursor: Musk's play for the coding community

Elon Musk certainly isn't planning to fall behind. His AI company xAI (sometimes referred to as SpaceXAI) is preparing to release a new model in collaboration with Anysphere — the company behind the popular AI developer tool Cursor — according to The Information. And possibly as soon as Wednesday. This is no coincidence. SpaceX (Musk's space company) recently bought Anysphere for $60 billion, making it clear that developer tools are one of the main battlegrounds of the AI race. The new model is said to be fast and, in some respects, competitive with Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. For developers, this means another option. While GitHub Copilot bets on the Microsoft ecosystem and Claude Code on safety and model "honesty," Musk's xAI apparently wants to win on pure speed and integration into Cursor, which has won over tens of thousands of programmers worldwide. Including Czech ones.

Meta Muse Image: generate images from your friends' photos. Even without their consent

Meta launched Muse Image this week — its first-ever image generation model, built at Meta Superintelligence Labs. At first glance, a classic generator: you write a prompt and get a visual. But there's a catch. Muse Image can use other users' public Instagram posts as a reference. All it takes is someone having a public profile, and anyone can "tag" them in a prompt and let AI create an image based on their photos. In the terms, Meta states that "people can create content with your Instagram content using Meta AI features" and that "you will not be notified about content created using Meta AI features." The only protection is the option to disable this usage in settings — opt-out, not opt-in. Which, in practice, means the vast majority of users don't even know about this feature, let alone actively disable it. For European users, this is an especially sensitive topic. GDPR requires explicit consent for processing personal data — and facial photographs undoubtedly fall into that category. It's not yet clear whether Meta plans to restrict this feature in the EU, or whether it's preparing for another round of disputes with European regulators. Given the company's history with GDPR (fines in the billions of euros), it would be naive to expect this to go unnoticed.

What this AI blitz means for average users and businesses

For Czech readers, there are several key points: GPT-5.6 will be available in ChatGPT even in the Czech Republic. OpenAI deploys models globally and Czech is among the supported languages. If you use ChatGPT Plus ($20/month) or Pro ($200/month), you'll be among the first to access the new models. Free users will likely get access to Luna — the cheapest variant. The xAI and Cursor model will be most interesting for developers. Cursor is popular among Czech programmers — according to informal surveys in Czech IT communities, roughly a third of developers working with AI assistants use it. The tie-up with Musk could mean aggressive pricing. Meta Muse Image raises the most questions. If Meta launches the feature in the EU in its current form, it will be a direct clash with GDPR. The Czech Office for Personal Data Protection (ÚOOÚ) and European supervisory authorities have already made it clear in the past that such practices will not be tolerated.

Broader context: the AI race is shifting from capabilities to accessibility

This week illustrates a fundamental trend of 2026: it's no longer just about who has the smartest model. It's about who can get it to people — and under what conditions. OpenAI had to wait a month for government approval. Anthropic had a similar experience with the Fable model (later Mythos). xAI is buying distribution channels (Cursor). Meta is using its existing user base (Instagram) as fuel for new AI features. It's a race in which regulation, distribution, and user trust play just as important a role as parameter counts or benchmark results. And for Europe, which bets on regulation as a competitive advantage (the EU AI Act is being phased in from February 2025), this means interesting times.