From government embargo to full availability

The GPT-5.6 story began dramatically. When OpenAI first introduced the model family on June 26, 2026, it was only a so-called "limited preview" — access was granted only to organizations approved by the Trump administration. The reason was cybersecurity concerns. The administration approved applicants on a case-by-case basis, and the whole process resembled export controls more than a standard software release. At the time, OpenAI commented that it "does not believe this type of government access process should become a long-term norm." After two weeks of intensive testing, which included extensive red teaming — including 700,000 A100e GPU hours of automated penetration testing — the green light came. On July 9, 2026, the Trump administration gave the models the go-ahead for public deployment. OpenAI CEO Sam Altman called GPT-5.6 "the best model we have ever built." And as if that weren't enough, OpenAI also announced a completely new product that same day — ChatGPT Work.

Three models at three prices: Sol, Terra, Luna

GPT-5.6 is not a single model, but three — each targeting a different type of use. In terms of pricing, it's a direct attack on Anthropic, whose Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. Sol is the flagship. At $5 per million input tokens and $30 per million output tokens, it offers what OpenAI claims is the highest performance across programming, cybersecurity, scientific research, and knowledge work. That's half the price of Claude Fable 5. On the Agents' Last Exam benchmark, which measures performance in professional tasks across 55 disciplines, Sol scored 52.7% — 13.1 points higher than Fable 5 with adaptive reasoning. Even in medium reasoning mode, Sol beats Fable 5 by 11.4 points at roughly a quarter of the price. Terra is the "golden middle" — at $2.50 per million input tokens and $15 per million output tokens, it delivers performance on par with GPT-5.5, the previous generation. According to OpenAI's measurements, Terra and Luna together outperform Claude Fable 5 at approximately one-sixteenth the price. That's a number that should interest anyone deploying AI models at scale. Luna is the cheapest and fastest variant — $1 per million input tokens, $6 per million output tokens. Ideal for routine tasks where you don't need top-tier performance but appreciate speed and low cost.

Max and Ultra modes: when you need more

OpenAI introduced two new modes for Sol. Max gives the model more time to think, explore alternatives, and revise its own conclusions. It's ideal for complex programming or analytical tasks where quality outweighs speed. Ultra goes even further — it runs four agents in parallel, working on different parts of the task simultaneously, and then synthesizes their results. On the Terminal-Bench 2.1 benchmark, which tests complex command-line workflows, Sol Ultra achieved a score of 91.9% — a new record. Pro and Enterprise users get Ultra in ChatGPT Work, while developers can build similar multi-agent workflows via the Responses API.

ChatGPT Work: An office assistant you let into your files

While the models themselves are an evolutionary step, ChatGPT Work is what could change the daily work of millions of people. OpenAI describes it as "a combination of ChatGPT and Codex for non-technical users." In practice, this means the agent can:

• Draw context from your apps, files, and workflows — you connect it to Slack, Gmail, Google Drive, calendars, and CRM systems through a unified plugin directory
• Create finished outputs: documents, spreadsheets, presentations, and web applications
• Work across applications — for example, take data from an email in Gmail, analyze it, and create a presentation from it

What only Claude Cowork from Anthropic (a combination of Claude + Claude Code) could do so far, OpenAI is now offering in its own version — and with a more aggressive price tag. The race to be the first to make AI agents truly useful for everyday people is heating up.

Availability: Desktop first, mobile and web gradually

OpenAI chose an unusual rollout strategy. Mac and Windows users — including those with a free account — get immediate access to ChatGPT Work and GPT-5.6 through the ChatGPT desktop app. On mobile and web, Pro, Enterprise, and Edu subscribers get priority, while Plus and Business users will get access "within a few days." The global rollout started immediately, and according to OpenAI, it will take approximately 24 hours for the models to reach all users. For Czech users, this means they should have access now — simply download the ChatGPT desktop app for Windows or Mac. Czech localization of ChatGPT has been available for a long time, so ChatGPT Work should work in Czech without issues.

Benchmarks: Where GPT-5.6 wins and where it loses

Beyond the already mentioned Agents' Last Exam, other numbers are worth noting. On the Artificial Analysis Coding Agent Index, Sol with maximum reasoning scored 80 — that's 2.8 points above Claude Fable 5, using less than half the output tokens and at roughly a third of the price. In cybersecurity, this is the biggest leap. On ExploitBench (creating exploits for the V8 engine), Sol reached 73.5% compared to 47.9% for GPT-5.5. On ExploitGym, it nearly doubled the previous generation's success rate. OpenAI emphasizes, however, that the model is stronger in defense than in offense — "it better helps people find and fix vulnerabilities than reliably carry out autonomous attacks." In the area of computer use, Sol set a new record on OSWorld 2.0 (62.6%) and BrowseComp (92.2% with ultra mode). This means the model can better control software, browse the web, and perform tasks that previously required human operation. On the Artificial Analysis Intelligence Index — a broad measure of intelligence covering agentic work, programming, scientific reasoning, and general abilities — Sol with maximum reasoning reached 58.9 points. That's just 1 point less than Claude Fable 5, but at half the price and with 61% shorter processing time.

Safety: 10× stricter blocking and a new reasoning monitor

This time, OpenAI paid unprecedented attention to safety — partly because the administration required it, partly due to Anthropic's recent issues with jailbreaking the Mythos model. The key innovation is the reasoning monitor — a system that analyzes conversations in real time and assesses whether there is a risk of misuse. Unlike traditional classifiers that simply flag a potentially dangerous request, this monitor "thinks" about the context. This allows it to block serious misuse while permitting legitimate defensive work — such as penetration testing or security research. Compared to previous models, GPT-5.6 Sol blocks roughly ten times more potentially harmful activity. OpenAI admits this sometimes creates friction even for legitimate use, and therefore offers the option to redirect a rejected request to a less capable model. To access the most sensitive cybersecurity capabilities, users must go through the Trusted Access for Cyber program under the OpenAI Daybreak initiative. This requires identity verification and hardware security keys (passkeys).

What this means for the Czech Republic and Europe

For Czech businesses and individuals, GPT-5.6 brings several concrete benefits: Price. Thanks to the new model family, it's now possible to run AI assistants significantly cheaper. Luna at $1/$6 per million tokens is affordable even for smaller Czech startups. Terra offers last year's flagship performance (GPT-5.5) at a fraction of the price. ChatGPT Work and Czech. ChatGPT has long supported Czech at a decent level. ChatGPT Work should work the same way — input in Czech, output in Czech. However, for more complex tasks (such as creating presentations), English will likely yield slightly better results. EU AI Act. While the US handles model regulation through executive orders, Europe is gradually implementing the AI Act. GPT-5.6, as a general-purpose AI model (GPAI), will need to meet European requirements for transparency and risk management. This doesn't change anything for users, though — the models are available through the ChatGPT app without geographical restrictions.

Broader context: The race for agentic AI intensifies

The launch of GPT-5.6 and ChatGPT Work fits into a broader trend. Google, Anthropic, Microsoft, and even Apple are competing to be the first to offer an AI agent that is truly useful for the mass user — not just for developers. The viral open-source agent OpenClaw showed that demand for autonomous AI assistants is enormous. Meanwhile, Anthropic is working on a public release of the Mythos model, which remains behind the gates of government access for now. With this double move — models plus an agentic product — OpenAI is signaling that it has no intention of losing its lead. But the race isn't just about technology. It's also about price. OpenAI has bet on an aggressive pricing policy: Sol at half the price of Fable 5, Terra and Luna at a fraction. At a time when companies are starting to count every token (Microsoft recently limited Claude Code, Uber exhausted its AI budget in 4 months), price may be the deciding factor.