April 24th news, this morningOpenAI Official launch of the most powerful flagship model under the flag GPT-5.5 And GPT-5.5 Pro, whose core competencies are to accept complex, multi-step and complex tasks and to do them autonomously - to plan the path, call the tools, verify the results, continue to move without the need for gradual user intervention。
Terminal-Bench 2.0 score 82.71 TP3T (GPT-5.4 for 75.11 TP3T), SWE-Bench Pro up to 5861 TP3T, internal ministerial mission assessment Expert-SWE up to 7311 TP3T, with three fewer tests token Completed
GDPval Cross 44 Career rating won or evened 84.9%, Tu2-bench Telecom passenger service stream 98.0% (GPT-5.4 for 92.8%) and OSWorld computer operation 78.7%
A BixBench bioinformatics assessment 80.51 TP3T, first in the released fraction model; an internal version also confirms a long-term assumption of Ramsey numbers and completes formalization in Lean
Designed in conjunction with the GB200/GB300 NVL72 system, with each token delay equal to GPT-5.4 while increasing token generation speeds above 201 TP3T through load equilibrium optimizationSame Cordex task, GPT-5.5 consumes significantly less token.
The Cursor co-founder and CEO Michael Truell states that GPT-5.5 "could last longer in the face of complex, long-term tasks without stopping" and that Every's co-founder and CEO Dan Shipper validated its ability to independently provide the same system re-engineering as senior engineers
One of the engineers in the U.S. even said, "It feels like being amputated without access."。
GPT-5.5 has been opened to ChatGPT Plus, Pro, Business and Enterprise users, Codex supports up to 400K context windows; the API version is also on line, with standard prices of token 5 per million and output 30 per million; GPT-5.5 Pro API pricing $30 and $180 respectively。