Agentic Performance and Speed Without Compromise

The most significant change with GPT-5.5 is not just higher test scores, but the way it works. OpenAI emphasizes agentic capabilities — the model better understands assignments, independently plans steps, uses tools, checks its own outputs, and can carry a task through to completion without constant user supervision. This means that instead of prompting step by step, you can entrust GPT-5.5 with a complex, multi-part task and rely on it to handle it.

Interestingly, despite its greater performance, GPT-5.5 maintains the same latency as GPT-5.4 — meaning the time between query and response has not increased. OpenAI achieved this through deep integration with inference infrastructure on NVIDIA GB200 and GB300 NVL72 systems, which were designed specifically for this model. At the same time, GPT-5.5 is significantly more efficient: for the same programming tasks in Codex, it consumes fewer tokens than its predecessor, which translates to lower costs during massive use.

How Does It Fare in a Direct Showdown with Claude Opus 4.7?

Anthropic released Claude Opus 4.7 on April 16, 2026, and it immediately became OpenAI's main competitor. However, GPT-5.5 shows in several key areas that OpenAI does not intend to cede ground. According to official OpenAI benchmarks, GPT-5.5 has the upper hand in the following disciplines:

On Terminal-Bench 2.0, which tests complex command lines requiring planning and tool coordination, GPT-5.5 achieved 82.7% compared to 68.5% for Claude Opus 4.7. In the field of cybersecurity (CyberGym), GPT-5.5 leads with 81.8% to 73.1%. In FrontierMath, a set of extremely demanding mathematical problems, GPT-5.5 handles 51.7% of level-one to level-three tasks, while Claude Opus 4.7 only manages 43.8%.

In the practical GDPval test, which simulates real knowledge work across 44 professions, GPT-5.5 achieved 84.9%, Claude Opus 4.7 80.3%. In computer control (OSWorld-Verified), GPT-5.5 scores 78.7%, Claude 78.0% — a tight contest. In web browsing and search (BrowseComp), GPT-5.5 leads with 84.4% to 79.3%.

Claude Opus 4.7, however, still holds several important points. In SWE-Bench Pro, a test of solving real problems on GitHub, Claude leads with 64.3% to 58.6% for GPT-5.5. In the demanding Humanity's Last Exam (without tools), Claude has 46.9% compared to 41.4% for OpenAI. In GPQA Diamond, a set of expert scientific questions, Claude achieves 94.2%, GPT-5.5 93.6%. Claude also has a slight advantage in some long-context tasks (Graphwalks).

The summary is clear: GPT-5.5 dominates in agentic workflows, programming, and tool use, while Claude Opus 4.7 maintains its advantage in pure abstract reasoning and some long-context tasks. For the average user and developer, this means the choice depends on the type of work — not on the absolute dominance of one model.

Price and Availability for the Czech Market

Starting April 23, 2026, GPT-5.5 is gradually being made available to subscribers of ChatGPT Plus, Pro, Business, and Enterprise. The GPT-5.5 Pro version is available to Pro, Business, and Enterprise users. For Czech users, this means full availability in Czech through the standard ChatGPT interface, which has supported the Czech language since earlier versions.

In the API, which is used by Czech companies and developers, GPT-5.5 will be available very soon at a price of USD 5 per 1 million input tokens and USD 30 per 1 million output tokens. The Pro version will be significantly more expensive: USD 30 per 1 million input and USD 180 per 1 million output tokens. Batch and Flex modes will offer half price, while Priority processing will be 2.5× higher. The ChatGPT Plus subscription remains at USD 20 per month (approximately CZK 460), Pro at USD 200 per month (about CZK 4,600).

For developers in the Czech Republic and the EU, it is key that OpenAI continues to operate in GDPR mode; however, the full application of the EU AI Act is approaching, which places stricter requirements on transparency and safety for high-impact models. OpenAI states that for GPT-5.5 it has deployed "the toughest safety measures in history," including external red-teaming and testing of biological and cyber risks.

Science, Mathematics, and the "Super App"

One of the most remarkable moments of the announcement is that an internal version of GPT-5.5 helped discover a new proof related to Ramsey numbers — a central problem in combinatorics. The proof was later verified in the Lean system. This is not a marketing trick, but a concrete example that the model can contribute to original scientific work, not just reproduce known facts.

OpenAI also again hinted at the direction toward a "super app" — a unified platform that combines ChatGPT, Codex, and a browser agent into one tool for enterprise customers. GPT-5.5 is to be a key building block of this vision. Greg Brockman, president of OpenAI, directly said that this model is "a big step toward agentic and intuitive computing."

Safety and Cyber Defense

With growing capabilities comes greater responsibility. OpenAI launched the Trusted Access for Cyber program with GPT-5.5, which allows verified security experts and organizations protecting critical infrastructure to use the model's advanced cyber capabilities with fewer restrictions. At the same time, the company introduced stricter filters for potentially harmful use. In the area of biological risks, OpenAI launched Bio Bug Bounty, a program rewarding the discovery of security weaknesses.

For European regulators and Czech companies working with sensitive data, it is important that OpenAI explicitly declares authenticated use and monitoring as part of its security strategy. This corresponds to the requirements of the approaching EU AI Act for high risk.