GPT-5.5 Is Here: Does It Reclaim the AI Crown?

Agent Programming: From Execution to System Architecture Understanding

GPT-5.5 sets new records on several benchmarks: Terminal‑Bench 2.0 at 82.7%, OSWorld at 78.7%, and GDPval at 84.9%.

Mathematical and scientific performance also jumps, with FrontierMath Tier 4 rising from 27.1% (GPT‑5.4) to 35.4%.

Every founder Dan Shipper recounts a real‑world debugging scenario where GPT‑5.5, after being shown a broken system state, generated the same rewrite solution that top engineers eventually produced—a task GPT‑5.4 failed to accomplish.

MagicPath CEO Pietro Schirano reports that GPT‑5.5 merged hundreds of front‑end and refactor branches into a major main branch in roughly 20 minutes, solving the entire workload in one pass.

Knowledge Work: From Information Processing to Autonomous Decision‑Making

In document, spreadsheet, and slide generation, GPT‑5.5 outperforms its predecessors. Combined with Codex's computer‑use skills, the model can truly "work alongside you": it can view screen content, click, type, navigate interfaces, and move precisely between tools.

OpenAI says over 85% of its staff use Codex weekly across engineering, finance, communications, and marketing. Example workflows include:

Communications team : analyzed six months of speaking‑request data, built a scoring and risk framework, and validated an automated Slack agent.

Finance team : reviewed 24,771 K‑1 tax forms (71,637 pages total), finishing two weeks earlier than the prior year.

Marketing team : automated weekly business reports, saving 5–10 hours per week.

GPT‑5.5 "Thinking" provides faster help for harder problems, while GPT‑5.5 Pro excels in business, legal, education, and data‑science domains.

Scientific Research: From Answering Questions to Driving Discoveries

On the GeneBench suite (multi‑stage scientific data‑analysis evaluation), GPT‑5.5 shows a marked improvement over GPT‑5.4, tackling tasks equivalent to several days of expert work.

Remarkably, GPT‑5.5 helped uncover a new proof related to Ramsey numbers, which was later verified in Lean, demonstrating the model's ability to contribute substantive mathematical arguments.

Polish assistant professor Bartosz Naskręcki used GPT‑5.5 to construct an algebraic‑geometry application in 11 minutes from a single prompt, visualizing quadratic‑surface intersections and converting the resulting curve to a Weierstrass model.

Codex Browser Interaction: From Reading to Acting

With GPT‑5.5, Codex now interacts more effectively with browsers, files, documents, and desktop applications. It can click pages, capture screenshots, and iteratively refine actions based on visual feedback, matching developers' front‑end testing loops.

Performance and Efficiency Breakthroughs

In Artificial Analysis's ten core evaluations, GPT‑5.5 secured first place in five, achieving a composite score of 60—three points higher than Claude Opus 4.7 and Gemini 3.1 Pro Preview.

Maintaining the same per‑token latency as GPT‑5.4, GPT‑5.5 improves across almost all tests and dramatically reduces token consumption for identical Codex tasks, making it both more powerful and more efficient.

Next‑Generation Inference Efficiency: Model Self‑Optimization

GPT‑5.5 is built, trained, and served for NVIDIA GB200 and GB300 NVL72 systems, treating inference as an integrated pipeline rather than isolated optimizations.

Codex analyzed weeks of production traffic, wrote a custom scheduling algorithm to optimally partition and balance work, boosting token‑generation speed by over 20%.

Enhanced Cybersecurity Protections

To mitigate the model's growing ability to locate and patch vulnerabilities, OpenAI deployed multiple safeguards:

Industry‑leading security controls for higher‑risk activities and sensitive network requests.

Expanded network‑trust access starting with Codex, extending to GPT‑5.5’s advanced security capabilities.

Collaboration with governments to protect critical infrastructure and explore trusted‑official AI defenses.

OpenAI classifies GPT‑5.5’s bio‑chemical and cybersecurity abilities as high‑risk; while not yet at a critical security threshold, testing shows noticeable improvements over GPT‑5.4.

Availability and Pricing Strategy

GPT‑5.5 is now available to ChatGPT and Codex Plus, Pro, Business, and Enterprise users.

API pricing details:

gpt‑5.5: $5 per million input tokens, $30 per million output tokens.

gpt‑5.5‑pro: $30 per million input tokens, $180 per million output tokens.

Fast mode: 1.5× token‑generation speed at 2.5× cost.

Although priced higher than GPT‑5.4, GPT‑5.5 delivers smarter results with fewer tokens, and OpenAI has tuned the experience so most users achieve better outcomes with less token usage.

Community Reaction

Cursor co‑founder and CEO Michael Truell says, "GPT‑5.5 is noticeably smarter, more persistent, and delivers stronger coding performance with reliable tool use. It stays on tasks much longer without early stopping, which is crucial for our users' complex, long‑running workloads."

NVIDIA Enterprise AI VP Justin Boitano adds, "GPT‑5.5 provides sustained performance for compute‑intensive work. Built and served on NVIDIA GB200 NVL72, it lets our team launch end‑to‑end features from natural‑language prompts, cutting debugging time from days to hours and turning weeks of experiments into overnight progress."

From Ability Scores to Completion Rates

The launch signals a shift from single‑prompt ability metrics to agent‑workflow completion rates as the key measure of AI value. As one developer notes, "Every startup exec reading this will green‑light a project; now each task can call APIs ten‑fold more because it’s essentially free."

Model quality plateaus in narrow scopes, but completion rates compound across workflows. GPT‑5.5 marks a fundamental transition from asking "how smart is it?" to "how many end‑to‑end tasks can it finish without human intervention?"

References:

https://artificialanalysis.ai/models/gpt-5-5

https://openai.com/index/introducing-gpt-5-5/