Two days after ChatGPT Images 2.0, OpenAI continues with GPT-5.5, codenamed “Spud”. Presented as its most successful model, it is designed to handle long and vague tasks almost independently, in a context where Claude Opus 4.7 ofAnthropic And Gemini 3.1 Pro of Google seriously challenged him for the crown.
Shaken by the rise of the Claude ecosystem and the media coverage of Google, OpenAI speaks openly of a “remontada”. GPT-5.5 becomes the central piece of this counter-offensive, with significantly increased performance and a very productivity-oriented positioning… but also significantly higher prices and access reserved for paying users.
Can OpenAI’s GPT-5.5 really regain the advantage over Claude and Gemini?
With GPT-5.5, OpenAI promises an almost autonomous model. Where GPT-5.4 required very detailed instructions, GPT-5.5 accepts messy requests, plans the steps, carries out web searches, spreadsheets or code, checks its work and corrects its errors without asking the user each time they get stuck. The company highlights demanding uses such as agentic coding, full use of a PC, knowledge work or scientific research.
OpenAI claims, for example, that GPT-5.5 succeeded in demonstrating a new asymptotic fact about Ramsey numbers and that it shines on GeneBench in genetics. Above all, it does so with a latency close to GPT-5.4 while consuming fewer tokens to achieve the same result.
In terms of numbers, GPT-5.5 clearly comes back ahead on several benchmarks. On Terminal-Bench 2.0 (command line), it reaches 82.7%, compared to 69.4% for Claude Opus 4.7 and 68.5% for Gemini 3.1 Pro. On GDPval (professional tasks), it climbs to 84.9%, ahead of Claude at 80.3% and Gemini at 67.3%. On OSWorld (autonomous use of a PC), GPT-5.5 obtains 78.7% against 78.0% for Claude, while Google is not measured.
The situation is more subtle on SWE-Bench Pro, focused on fixing GitHub bugs: GPT-5.5 reached 58.6%, far behind the 64.3% of Claude Opus 4.7, which OpenAI nevertheless accuses of having potentially “memorized” part of the test game. Overall, GPT-5.5 dominates the command line, office automation and cybersecurity (CyberGym), but Claude and Gemini maintain advantages in certain code and abstract logic tests.
What does GPT-5.5 change for paid ChatGPT and Codex users?
GPT-5.5 is coming directly into ChatGPT And Codexbut only for Plus, Pro, Business and Enterprise subscribers. Millions of free users remain stuck on GPT-5.4, without access to new agentic capabilities.
OpenAI offers its model in two variants: GPT-5.5 Thinking and GPT-5.5 Pro. Thinking mode targets tough everyday problems, with faster, more concise answers to move complex work forward. The Pro version targets the hardest cases, where every detail counts, with a significant increase in precision according to the first testers.
On the developer side, GPT-5.5 is integrated into Codex with a context window of 400,000 tokens, convenient for handling huge code bases or massive documents. A Fast mode generates approximately 1.5 times faster, for a cost multiplied by 2.5, which clearly reserves it for uses where response time is critical.
OpenAI also emphasizes security. GPT-5.5 launches with “its strongest set of guardrails yet,” following extensive cybersecurity and biological evaluations and testing with approximately 200 trusted Early Access partners.
GPT-5.5: price, access and what OpenAI’s new strategy means for you
This move upmarket comes at a price. On the API, GPT-5.5 costs $5, or around €4.70, for 1 million input tokens and $30, around €28, for 1 million output tokens, which is twice as much as GPT-5.4. For GPT-5.5 Pro, the bill can go up to 180 dollars, almost 168 €, per million tokens output.
OpenAI defends this increase by explaining that GPT-5.5 is more “intelligent” and above all more economical in tokens, which can partially offset the bill on well-regulated workflows. The fact remains that for a studio, an SME or a French freelancer who processes large volumes of text or code, the budget line risks increasing significantly.
In ChatGPT, only the Plus subscription provides access to standard GPT-5.5, while GPT-5.5 Pro is reserved for Pro, Business and Enterprise plans. The message is clear: OpenAI targets high-value professional uses rather than the free general public, while Anthropic and Google are also pushing their premium offers.
For users, the question therefore becomes strategic: agree to pay more for the autonomy gains of GPT-5.5, stay on Claude Opus 4.7 or Gemini 3.1 Pro, or mix several suppliers depending on the tasks. One thing is certain, the battle between OpenAI, Anthropic and Google has just escalated.