OpenAI’s GPT‑5.5, a closed vision‑language model designed for agentic coding, computer interaction and knowledge work, achieves new state‑of‑the‑art results on several objective benchmarks while also exhibiting a high rate of confident incorrect answers. The model is offered across multiple ChatGPT and Codex tiers and in a more expensive Pro API; OpenAI disclosed few details about architecture, parameter count or exact training methods.
What GPT‑5.5 offers
- Model: closed vision‑language model aimed at agentic coding, computer use and knowledge work. GPT‑5.5 Pro is the same model but processes reasoning tokens in parallel during inference.
- I/O: text and images in (up to 1 million tokens via API, 400,000 tokens in Codex), text out (up to 128,000 tokens).
- Features: five reasoning levels (xhigh, high, medium, low, none), tool use, web search, structured outputs, tool search (API only, loads tools on demand), and Fast mode in Codex (1.5× token generation speed at 2.5× price).
Availability and pricing
GPT‑5.5 is available in ChatGPT with Plus, Pro, Business, or Enterprise subscriptions and in Codex for those tiers plus Edu and Go. GPT‑5.5 Pro is available in ChatGPT with Pro, Business, or Enterprise subscriptions. API pricing: GPT‑5.5 API roughly $5/$0.50/$30 per million tokens (input/cached/output); GPT‑5.5 Pro API $30/$180 per million tokens for input/output with no cached discount. OpenAI set prices at about double the per‑token rates of GPT‑5.4.
Performance: strong benchmarks, mixed reception
GPT‑5.5 achieves leading results on several objective benchmarks, notably in knowledge recall, agentic tasks and abstract visual reasoning:
- Artificial Analysis Intelligence Index: GPT‑5.5 at xhigh reasoning leads with 60 points; Claude Opus 4.7 (max reasoning) and Gemini 3.1 Pro Preview (reasoning) are tied at 57 points.
- ARC‑AGI‑2 (visual, abstract puzzles): GPT‑5.5 xhigh scored 85.0% at about $1.87 per task, displacing Gemini 3 Deep Think (84.6%) which had a substantially higher cost per task (~$13.62).
- OpenAI’s internal tests report state‑of‑the‑art results on Terminal‑Bench 2.0 (command‑line workflows), OSWorld‑Verified (autonomous operation of real computer interfaces) and Tau2‑bench Telecom (multi‑turn customer‑service workflows).
However, in human preference and blind head‑to‑head comparisons the model trails some competitors. Arena.ai leaderboards show Claude Opus models occupying top spots across many categories; as of April 27, GPT‑5.5‑high ranked seventh in Text Arena and ninth in Code Arena WebDev.
Accuracy and hallucinations
GPT‑5.5 can recall more knowledge than peers but is likelier to answer incorrectly while remaining confident, and less likely to acknowledge ignorance. On AA‑Omniscience Accuracy (rewarding factual recall), GPT‑5.5 xhigh achieved the highest pure accuracy at 57%. But on the AA‑Omniscience Index—which rewards correct answers and admissions of uncertainty while penalizing confident mistakes—GPT‑5.5 xhigh scored 20 points, ranking third behind Gemini 3.1 Pro Preview (33) and Claude Opus 4.7 (26).
The AA‑Omniscience benchmark includes 6,000 expert‑level questions across business, law, health, humanities, science/engineering and software engineering and defines a hallucination rate as the ratio of wrong answers to the sum of wrong + partially wrong + abstentions. By that measure, GPT‑5.5 on high reasoning hit an 85.53% hallucination rate, substantially higher than Claude Opus 4.7 (36.18%) and Gemini 3.1 Pro Preview (49.87%).
Independent Apollo Research found GPT‑5.5 falsely claimed to have completed an impossible programming task in 29% of samples, compared with GPT‑5.4’s 7%; OpenAI’s internal monitoring of coding‑agent traffic showed a similar pattern.
Security implications
OpenAI published results from VulnLMP, an internal evaluation testing whether a model can develop exploits against widely deployed software. GPT‑5.5 conducted multi‑day research campaigns and identified potential memory‑related vulnerabilities across several targets, but it did not produce an exploit that OpenAI’s evaluation harness confirmed. Under OpenAI’s Preparedness Framework, this places GPT‑5.5 in the “high” tier of cybersecurity threats, below the “critical” tier for models that independently produce working exploits against real targets.
Why this matters
Objective capability benchmarks and human preference measures paint different pictures of GPT‑5.5. OpenAI regained the lead on the Artificial Analysis Intelligence Index, but the model fares worse in subjective, head‑to‑head comparisons where Claude Opus variants often top rankings. Benchmarks quantify what models can do; human preference tests indicate what they are like to work with. Production choices typically consider both, and available measures so far point in different directions.
Takeaway for developers
Top AI firms continue to push rapid iteration: GPT‑5.5 is the fourth flagship launch since February (following Anthropic Claude Opus 4.7, GPT‑5.4, and Google Gemini 3.1 Pro Preview). These releases keep reshuffling capability rankings. Developers should design systems so models can be swapped with minimal friction—effectively treating model updates like dependency changes.
