A mathematician gives one hint — receives a ready-made application in algebraic geometry in 11 minutes. This is precisely the example OpenAI President Greg Brockman cited during a press briefing while presenting GPT-5.5. But behind the impressive demonstration lies a more important signal: the model is the first to be truly designed to work without a human nearby.
What's New
GPT-5.5 doesn't simply answer queries better — it independently plans actions, switches between tools, and continues working even when the plan changes or an attempt fails. According to OpenAI, the model achieves higher quality results using fewer tokens and fewer attempts compared to GPT-5.4 — becoming cheaper to use while simultaneously becoming more powerful.
In agentic coding benchmarks: Terminal-Bench 2.0 — 82.7%, SWE-Bench Pro — 58.6%. On the internal Expert-SWE test, where tasks have a median human execution time of 20 hours, GPT-5.5 also surpassed its predecessor.
"It's a faster, more accurate thinker using fewer tokens compared to 5.4. This means more frontier-AI for businesses and consumers."
Greg Brockman, OpenAI President, at the press briefing
Surprising Fact: The Model is More Powerful — But It's Not in the API
GPT-5.5 is already available to ChatGPT users on paid plans (Plus, Pro, Business, Enterprise) and in Codex. However, developers who integrate OpenAI models through the API are still waiting. OpenAI explains: agentic capabilities require separate security measures for large-scale deployment, and the company is "actively working with partners" on their formalization. API access is promised "very soon" — without a specific date.
This is not a technical limitation. GPT-5.5 is already served on NVIDIA GB200 NVL72 infrastructure, which, according to NVIDIA, delivers 50 times higher performance per watt compared to the previous generation. More than 10,000 NVIDIA employees have been using the model for several weeks through Codex in a corporate environment with read-only access to production systems.
Context: Release Pace
GPT-5.5 launched six weeks after GPT-5.4 — one of the shortest update cycles in the company's history. Fortune describes this as evidence of intensifying competition for corporate clients. According to OpenAI, ChatGPT has more than 900 million weekly active users, 50 million subscribers, and 4 million active Codex users. Despite this, a narrative is growing in media circles that OpenAI is losing pace in competition with Anthropic — especially in the corporate segment.
On OpenAI's own comparison charts, GPT-5.5 surpasses Claude Opus 4.5 on Terminal-Bench 2.0, but Anthropic remains ahead on SWE-Bench Pro — and this is according to OpenAI's own data.
Where This Leads
OpenAI Chief Researcher Mark Chen noted that the model "shows tangible progress in scientific and technical research processes" and could help in drug discovery. Brockman called GPT-5.5 a step toward a "super-application" — a single AI interface for any computer work.
The real question is not whether GPT-5.5 will be useful. It already is — for those using it. The question is about conditions: if OpenAI opens the API without clearly defined mechanisms for controlling agentic sessions, what exactly will stop the model from taking actions the user didn't anticipate?
