OpenAI has officially released GPT-5.5 and GPT-5.5 Pro, now live in ChatGPT and Codex, though the API isn't open yet. The new model keeps the same per-token latency as GPT-5.4 but uses significantly fewer tokens to complete the same Codex tasks. OpenAI says users no longer need to micromanage every step — just give it a vague multi-step task, and the model will plan, call tools, check results, and push through ambiguity on its own.
Here are the key benchmark numbers: Terminal-Bench 2.0 scored 82.7%, a 7.6-point jump over GPT-5.4's 75.1%, while Claude Opus 4.7 sits at 69.4%. On ARC-AGI-2, GPT-5.5 hit 85.0% versus GPT-5.4's 73.3% and Gemini 3.1 Pro's 77.1%. In long-context MRCR (512K–1M tokens), GPT-5.5 scored 74.0%, more than double GPT-5.4's 36.6% and far ahead of Opus 4.6's 32.2%. On SWE-Bench Pro, GPT-5.5 managed 58.6%, though Opus 4.7 leads there with 64.3%. According to Artificial Analysis' Intelligence Index, GPT-5.5's API cost at equivalent intelligence is half that of competing frontier coding models.
On the research front, an internal version of GPT-5.5 discovered a new proof for a long-standing open asymptotic fact about off-diagonal Ramsey numbers — a core object in combinatorics that studies how large a network must be to guarantee a specific ordered structure. The proof has been formally verified in Lean.
API pricing: GPT-5.5 costs $5 per million input tokens and $30 per million output tokens, with a 1M-token context window. GPT-5.5 Pro costs $30 input and $180 output. On ChatGPT, GPT-5.5 is available to Plus, Pro, Business, and Enterprise users; GPT-5.5 Pro is for Pro, Business, and Enterprise. On Codex, GPT-5.5 supports a 400K context window, and there's a Fast mode that runs 1.5x faster at 2.5x the price. This launch comes as OpenAI enlists Accenture and PwC to sell Codex to businesses, signaling a major enterprise push.