The premiere arrived just six weeks after the release of version 5.4, setting a precedent in the industry’s history: a new major model every month and a half. GPT-5.5 has been available since Thursday for users of paid Plus, Pro, Business, and Enterprise plans in ChatGPT and Codex applications. The Pro version – only in the three highest plans. API access is expected “very soon.” And yesterday, April 24, OpenAI confirmed: the API is already live.
What the new model can do
OpenAI claims that GPT-5.5 better understands complex, multi-step tasks, independently plans work, and checks results. This is another step towards “agentic” models – those that can be given an imprecisely formulated problem and allowed to determine what needs to be done to solve it. Significantly, GPT-5.5 is the first model built from scratch since GPT-4.5 – all intermediate versions (5.1, 5.2, 5.3, 5.4) were modifications of the same base. This one has a new architecture, new training data, and new design goals focused on agentic work. The internal codename OpenAI used during testing was “Spud” – hence the potato in the launch announcement posters.
Greg Brockman, president of OpenAI, said during a press conference that GPT-5.5 “can look at an ambiguous problem and figure out what to do next on its own.” This is language the company is using more and more frequently. In recent months, OpenAI has been talking not about assistants, but about systems that perform work instead of the user. Specifically: writing and debugging code, operating programs, online research, data analysis, creating documents and spreadsheets. Mark Chen, OpenAI’s director of research, added that the new model “shows significant progress in scientific and research work” and can help experts, among other things, in discovering new drugs – an area the company has been paying particular attention to in recent months.
GPT-5.5 is the first OpenAI model available via API to have a context window of one million tokens – this practically means the ability to process the content of an entire book or several hundred pages of financial documents in a single query. The increase in performance does not come with a slowdown. OpenAI declares that GPT-5.5 responds as quickly as 5.4, and also consumes significantly fewer tokens – about 40 percent less – for the same task. On the other hand, the API price is twice as high: five dollars per million input tokens and thirty per million output tokens, compared to two and a half and fifteen dollars for GPT-5.4. For developers who calculate costs per query, this is a concrete change, but with fewer tokens per task, the difference works out to be smaller in practice.
Three versions, three different uses
GPT-5.5 does not enter the market as a single model. It’s three distinct tools. The first – GPT-5.5 Thinking, available in ChatGPT for all paid users, designed for tasks requiring longer reasoning. The second – GPT-5.5 Pro, the most powerful version, available in Pro, Business, and Enterprise plans. The third – the default, automatic Instant mode, which decides whether a task requires a quick answer or more extensive thinking.
When you launch Instant in ChatGPT and type a question, the system itself assesses its complexity. Simple – it returns an answer immediately. Complex – it switches to Thinking, sometimes displaying a brief description to the user of what it intends to do. This is the first step towards what OpenAI clearly calls: a “super app.” Combining ChatGPT, Codex, and the browser into one integrated system.
Why so fast
Six weeks between versions 5.4 and 5.5 is a pace that would have been unthinkable two years ago. The industry had set a standard: one major version per year, two to three smaller updates in between. Today, OpenAI releases a full-fledged model every month and a half. And it has no intention of slowing down.
The reason – unconcealed – is competition. Anthropic, which recently received another 25 billion dollars from Amazon in exchange for 100 billion dollars in AWS infrastructure commitments, is hot on OpenAI’s heels in the enterprise market. According to internal OpenAI sources, this is called “red code” – an alarm situation. The company’s management has shifted its focus from consumers to business clients in recent months. The numbers from Thursday were meant to confirm this narrative: 4 million Codex users, 9 million paid ChatGPT business users, 900 million weekly active users, and 50 million subscribers in total.
For comparison – the tech industry hasn’t seen such a pace since the smartphone war between Apple and Samsung in 2012–2014. Only that war was about gadgets, and this one is about who will be the thought infrastructure of the next decade.
Safety and controversies
OpenAI announced that GPT-5.5 has the “strongest set of safeguards” in the company’s history. Biological, chemical, and cybernetic models have been classified as “High” in the internal Preparedness framework, meaning they could theoretically aid in the development of weapons of mass destruction. Hence the one-day delay in API availability – OpenAI had to refine additional safeguards.
In the background also hangs a recent incident related to competition. Anthropic introduced a cybersecurity tool called Mythos a few weeks ago. Soon after, reports of unauthorized access to the program emerged. When a journalist asked during an OpenAI briefing if GPT-5.5 would have similar functions, Mia Glaese, VP of Research at OpenAI, responded cautiously: “We have a strong and long-term strategy in the cyber area; we have refined our approach to safely deploying models.”
What this means for Polonia
Specifically – a lot. Polish IT companies in America, which have been building products based on ChatGPT for years, now get a more powerful engine without needing to change their infrastructure. Polish programmers in New York, Boston, and Silicon Valley – and there are thousands of them – work daily with Codex, whose new version has just entered the game. Polish small and medium-sized businesses, which were afraid of AI just a year ago, now have concrete options available for 20 dollars a month.
Bank of New York – one of the largest financial institutions in America, where many Poles work – tested GPT-5.5 over the past weeks alongside competitor models. Leigh-Ann Russell, the bank’s Chief Information Officer, emphasized in a statement to Fortune that the new model shows a “significant leap in terms of hallucination resilience” – meaning generating false answers with false certainty. According to OpenAI’s internal measurements, GPT-5.5 has sixty percent fewer hallucinations than version 5.4. For the regulated financial sector, this is a key feature – in a bank, one false answer can cost a client millions.
It should also be mentioned that the entire model runs on NVIDIA GB200 NVL72 chips – the same infrastructure used by the largest AI installations in America. According to Nvidia data, the new chips reduce the cost of running advanced AI by up to thirty-five times per million tokens. This translates into economics: companies that a year ago could not afford to massively use AI in production now can.
On the other hand – it’s worth remembering that language models are powerful, but they still invent facts, confuse dates, and provide incorrect numbers with full conviction. Every output of AI must pass through human hands. This is true for banks, for newsrooms, for anyone considering using these tools in professional work. GPT-5.5 may be the most intelligent model on the market. But it still doesn’t replace a human who knows the facts.
Arthur Skok, poland.us
OpenAI GPT-5.5 – April 23, 2026 | Launch in ChatGPT and Codex: April 23 | API Launch: April 24 | Previous version (GPT-5.4): early March 2026 | Interval: ~6 weeks | Internal codename: “Spud” | First fully-retrained base model since GPT-4.5 | Three variants: Instant, Thinking, Pro | Context: 1 million tokens | API Price: 5 USD/M input, 30 USD/M output (2x more than 5.4) | Tokens per task: approx. 40% less | Hallucinations: 60% less than in 5.4 | Infrastructure: NVIDIA GB200 NVL72 | OpenAI figures: 900 million weekly users, 50 million subscribers, 4 million Codex users, 9 million business users | Competition: Anthropic Claude Opus 4.7 and Mythos | Based on OpenAI press releases and reports from CNBC, TechCrunch, Fortune, Axios, MacRumors, 9to5Mac, Cryptopolitan, ofox.ai, apidog, theplanettools (April 23–24, 2026)
Read also: Amazon × Anthropic – how OpenAI’s competitor secured 25 billion dollars | Tim Cook leaves Apple in the shadow of the AI race
Read more on Poland.us.
