What exactly can GPT-5.5 do and how is it different

GPT-5.5 isn't just an improved chatbot. OpenAI is building it as a model meant to take over a larger portion of work without requiring detailed supervision. "Instead of carefully managing every step, you can give GPT-5.5 a messy, multi-part task and trust it to plan the approach, use tools, check its work, and keep going," states OpenAI in its announcement.

In practice, this means several concrete improvements. The model can more quickly understand what the user wants and handle tasks that require planning, using external tools, checking outputs, and navigating ambiguous situations. OpenAI highlights four main areas where GPT-5.5 raises the bar most significantly:

• Agentic coding — writing and debugging code, including working with large codebases.
• Computer control — interacting with web applications, browsing pages, taking screenshots, and iterating based on what it sees.
• Knowledge work — creating spreadsheets, documents, online research, and working across different programs.
• Early scientific research — analyzing large datasets, for example in genomics or pharmaceutical research.

Interestingly, OpenAI promises better performance without losing speed and without higher token consumption. In fact, GPT-5.5 is supposed to be more efficient compared to the previous version GPT-5.4 — in Codex, it should return better results to most users with fewer tokens. This is especially important for developers who pay for every token used.

Technically, the model runs on NVIDIA GB200 and GB300 NVL72 systems and uses custom heuristic algorithms — written by artificial intelligence itself — to distribute and balance work across GPU cores. Thanks to this, OpenAI claims token generation speed has increased by more than 20%.

Availability and price: who gets access to the model and for how much

GPT-5.5 is being rolled out gradually. In ChatGPT, the model is available to subscribers of Plus, Pro, Business, and Enterprise plans. The GPT-5.5 Pro version, designed for the most demanding tasks with higher precision, is being made available to users of Pro, Business, and Enterprise plans. In Codex, GPT-5.5 is also available for Edu and Go plans, showing that OpenAI is targeting students and smaller teams as well.

For regular ChatGPT subscribers, there is no fundamental price change — they pay their standard monthly fee (Plus $20, Pro $200 per month). But for developers and companies calling the model via API, a noticeable price hike is coming. GPT-5.5 is twice as expensive as GPT-5.4 in the API:

OpenAI tries to compensate for the higher price with better efficiency. In the enterprise tool Codex, a Fast mode is also coming, which generates tokens 1.5x faster, but at 2.5x the price. API access to GPT-5.5 is not yet publicly open; OpenAI promises it will come "very soon."

Benchmarks: OpenAI returns to the top, but Claude Mythos remains a dangerous rival

The market for the most advanced American language models dramatically intensified in April 2026. Exactly a week ago, Anthropic released Claude Opus 4.7, which briefly topped the leaderboards. OpenAI now responds with GPT-5.5 and in a number of tests returns to first place among publicly available models.

On the Terminal-Bench 2.0 benchmark, which tests a model's ability to navigate and complete tasks in an isolated terminal environment, GPT-5.5 achieved an accuracy of 82.7%. This significantly outperforms Claude Opus 4.7 (69.4%) and even narrowly beats the heavily limited Claude Mythos Preview (82.0%). On the internal Expert-SWE test, which involves long-term coding tasks with a median human completion time of 20 hours, GPT-5.5 scored 73.1% with significantly lower token consumption.

Overall, GPT-5.5 leads in 14 benchmark disciplines, while Claude Opus 4.7 dominates in 4 and Google Gemini 3.1 Pro in 2. OpenAI thus claims primacy in agentic computer control, economic analysis (GDPval), specialized cybersecurity (CyberGym), and complex mathematics (FrontierMath).

But it must be added that in pure academic reasoning without tools, the competition remains strong. On the Humanity's Last Exam test without external tools, GPT-5.5 Pro achieved 43.1%, while Claude Opus 4.7 scored 46.9% and the inaccessible Mythos Preview even 56.8%. In software engineering according to the public SWE-bench Pro test, Claude Opus 4.7 leads (64.3%) ahead of GPT-5.5 (58.6%).

Safety: strictest guardrails and special license for cybernetics

With greater capability comes greater responsibility. OpenAI states that GPT-5.5 has "the most robust set of guardrails against misuse we have ever released." The model was evaluated across a range of safety frameworks, tested by internal and external red teams, and subjected to targeted testing of advanced cybernetic and biological capabilities. Nearly 200 trusted partners with early access got their hands on it before launch.

According to OpenAI's Preparedness Framework, GPT-5.5 is classified as a model with "high" risk in biology and cybersecurity. Therefore, the company introduced stricter "cyber risk classifiers" for regular users. On the other hand, for legitimate security professionals, it offers a "cyber-permissive" license, which allows verified defenders of critical infrastructure to use the model with fewer refusals on security questions.

This dual approach reflects the reality that advanced AI can accelerate defense, but also be misused as a weapon. OpenAI therefore works with government partners to ensure that tools strengthen digital resilience rather than undermine it.

What this means for Czech users and companies

For Czech users, the key message is that ChatGPT with GPT-5.5 is available in Czech — just like previous versions, it supports the Czech language, though with the caveat that it achieves the best results primarily in English. Companies and developers in the Czech Republic who use ChatGPT Plus or Enterprise plans will automatically get access to the model within their subscription.

For local startups and development teams, however, the double API price hike is a significant signal. While large corporations will likely absorb the cost, smaller Czech companies may hesitate whether to switch to the new model or stay with the cheaper GPT-5.4. OpenAI argues that higher intelligence and lower token consumption may ultimately balance out, but that will depend on the specific use case.

From a European perspective, it is also worth watching how the EU AI Act will approach the model. The classification of GPT-5.5 as "high risk" in cybernetic and biological areas could lead to stricter regulatory requirements when deployed in critical infrastructure or healthcare. Czech companies should therefore monitor not only technical parameters but also the development of the regulatory framework.

Among the first users of the model, remarkable reactions appeared. Dan Shipper, CEO of Every, called GPT-5.5 "the first coding model with true conceptual clarity." An anonymous NVIDIA engineer who had early access stated that "losing access to GPT-5.5 is like losing a limb." The scientific community, in turn, appreciates the model's ability to analyze datasets that would take humans months.

Conclusion: AI is leaving the chat window and moving into the operating system

GPT-5.5 is not another minor iteration. It is a tool designed for a world where people delegate entire workflows, not just individual queries. By doing so, OpenAI is pushing the boundaries of what publicly available artificial intelligence can do, and simultaneously signaling that the best models will be increasingly expensive and more strictly regulated.

For Czech readers and companies, this is an invitation to start actively testing the model — whether directly in ChatGPT or in the programming tool Codex. With API access approaching and the competitive race with Anthropic and Google, the pace of innovation is expected to accelerate further. The question remains who can best leverage the new capabilities, not just who creates them.