With GPT-5.5, you can go further and for longer. π
OpenAI launches a model that thinks complex tasks through to the end on its own, from code to spreadsheets to research.
OpenAI has released GPT-5.5. The promise isn't another storm of marketing, but a model that noticeably shifts up a gear in coding agents, computer use, knowledge work, and early research. And it does so at the same latency as GPT-5.4.
The core: GPT-5.5 understands the intent behind a messy, multistep task faster and carries more of it out independently. Planning, using tools, reviewing its own work, navigating ambiguity, pushing ahead. Instead of directing every step, you hand over the problem and let the model get to work.
Its efficiency is remarkable. Larger models are usually slower, yet GPT-5.5 matches the per-token latency of GPT-5.4 in production and requires significantly fewer tokens for Codex tasks. According to the Artificial Analysis Coding Index, it delivers top-tier intelligence at half the price of competing frontier models.
Agentic Coding
82.7% on Terminal-Bench 2.0, 73.1% on Expert-SWE on tasks estimated at 20 hours of effort. In Codex, the model retains context across large systems, tests assumptions using tools, and carries changes through the codebase consistently.
Knowledge Work
84.9% on GDPval, 78.7% on OSWorld-Verified, 98.0% on Tau2-bench Telecom without prompt tuning. Strong on documents, spreadsheets, presentations, and when the model operates the screen itself.
Research
Clear progress on GeneBench (25.0%) and BixBench (80.5%). An internal variant discovered a new proof for Ramsey numbers, later verified in Lean. A co-scientist rather than an answer engine.
Cybersecurity
88.1% on internal CTF challenges, 81.8% on CyberGym. OpenAI rates its cyber and biology capabilities as High in its Preparedness Framework and is enforcing stricter classifiers.
Early testers report a qualitative leap, not just a quantitative one. Dan Shipper, CEO of Every, calls it the first coding model with serious conceptual clarity. Pietro Schirano from MagicPath had GPT-5.5 merge a branch containing hundreds of frontend and refactor changes into an equally heavily modified main branch, in about 20 minutes, in a single pass. An NVIDIA engineer noted that losing access felt like having a limb amputated.
At OpenAI itself, more than 85% of employees use Codex weekly. The communications department used it to build a scoring and risk framework for speaking requests complete with a Slack agent. Finance audited 24,771 K-1 tax forms spanning 71,637 pages, saving two weeks compared to the previous year. Go-to-Market automated weekly reporting, saving 5 to 10 hours per week.
A concrete research example: Derya Unutmaz, professor of immunology at the Jackson Laboratory, analyzed a gene expression dataset containing 62 samples and roughly 28,000 genes using GPT-5.5 Pro. The detailed research report would have taken his team months. Bartosz NaskrΔcki, mathematics professor in PoznaΕ, built an algebraic geometry app from a single prompt in eleven minutes using Codex, including a Weierstrass model of the intersection curve.
| Benchmark | GPT-5.5 | GPT-5.4 | Claude Opus 4.7 |
| Terminal-Bench 2.0 | 82.7% | 75.1% | 69.4% |
| OSWorld-Verified | 78.7% | 75.0% | 78.0% |
| FrontierMath Tier 4 | 35.4% | 27.1% | 22.9% |
| CyberGym | 81.8% | 79.0% | 73.1% |
The infrastructure is part of the story. GPT-5.5 was trained and served on NVIDIA GB200 and GB300 NVL72 systems. Codex helped draft and test optimizations faster. GPT-5.5 itself wrote heuristic algorithms for load balancing and partitioning, boosting token generation speed by more than 20%. The model improves the infrastructure that powers it.
Availability: GPT-5.5 is available immediately for Plus, Pro, Business, and Enterprise users in ChatGPT and Codex. GPT-5.5 Pro is available for Pro, Business, and Enterprise. In the API, GPT-5.5 is expected to cost $5 per 1M input tokens and $30 per 1M output tokens with a 1M context window. GPT-5.5 Pro is priced at $30 and $180, respectively.
The Takeaway
GPT-5.5 isn't a marketing leap, but a pragmatic one: sharper intellect, identical latency, fewer tokens. The interesting part isn't the next benchmark row, but what people build with it. Anyone serious about agentic workflows now has a tool that can sustain long chains. Anyone who just wants to chat will mainly notice that the answers become shorter and more to the point. Both represent progress.