TL;DR:

• OpenAI has released GPT-5.5 as what it calls a step-change in handling multi-part agentic tasks, with state-of-the-art scores on Terminal-Bench 2.0 (82.7%), Expert-SWE (73.1%), GDPval (84.9%) and OSWorld-Verified (78.7%).
• The model rolls out today to Plus, Pro, Business and Enterprise ChatGPT and Codex users, with GPT-5.5 Pro limited to the upper tiers and an API release promised within weeks.
• Per-token latency matches the previous GPT-5.4 version despite the capability jump, and OpenAI says the model uses fewer tokens to complete the same Codex tasks.

OpenAI has released GPT-5.5, which it positions as a model designed to handle messy, multi-part tasks end-to-end rather than requiring users to manage each step. The launch places a renewed emphasis on agentic coding, computer use and long-horizon knowledge work — areas where OpenAI claims the step up from GPT-5.4 is largest.

What is different about GPT-5.5

OpenAI reports state-of-the-art scores on the agentic coding benchmarks most often used to compare frontier models. GPT-5.5 hits 82.7% on Terminal-Bench 2.0 (against 69.4% for Claude Opus 4.7), 58.6% on SWE-Bench Pro for GitHub issue resolution, and 78.7% on OSWorld-Verified for operating real computer environments. On GDPval, which measures agentic knowledge work across 44 occupations, it scores 84.9%. Early testers quoted in OpenAI’s announcement — including senior engineers at Cursor and NVIDIA — described the model as noticeably stronger than GPT-5.4 at holding context across large systems and catching issues without prompting.

The efficiency story may matter more than headline benchmarks for UK enterprise buyers. OpenAI says the model matches GPT-5.4 per-token latency in production serving and needs significantly fewer tokens for the same Codex tasks, which translates directly into lower per-seat cost for sustained use. OpenAI itself reports that more than 85% of its staff now use Codex every week across engineering, finance, communications and marketing.

Safety and rollout

OpenAI says GPT-5.5 shipped with its strongest safeguards to date, including targeted evaluations for advanced cybersecurity and biology capabilities and feedback from nearly 200 early-access partners. API deployments are being handled separately while OpenAI works on safeguards for at-scale serving, with a full API release promised within weeks. The release follows Anthropic’s Claude Opus 4.7 and comes as the frontier cyber-capability debate intensifies in the UK, where the AI Security Institute has been testing Anthropic’s Claude Mythos for vulnerability-discovery risk.

Looking Forward

For UK businesses, GPT-5.5’s agentic positioning sharpens a procurement question most have not yet answered: what governance applies when a coding or knowledge-work tool can plan, use tools and carry an outcome over hours rather than minutes? The efficiency claims — fewer tokens, matched latency — make the model harder to reject on cost grounds. The hard work now sits with risk and legal teams deciding which long-horizon workflows a frontier model should be trusted to complete on its own.