GPT-5.5 Leap: Data Analysis of Artificial Analysis AI Index

OpenAI released its GPT-5.5 model family on April 23, 2026, and the results on the Artificial Analysis Intelligence Index v4.0 — a composite benchmark aggregating ten evaluations across mathematics, science, coding, and reasoning — mark a notable leap in frontier AI performance. GPT-5.5 (xhigh) achieved a score of 60, making it the first model to break out of the 50s range on the index and surpassing the previous state-of-the-art score of 57, which was shared by GPT-5.4 (xhigh), Claude Opus 4.7 (max), and Gemini 3.1 Pro Preview. GPT-5.5 (high) followed closely at 59, while GPT-5.5 (medium) matched the prior leaders at 57, effectively tying for the previous SOTA at a fraction of the cost.

The cost dynamics are particularly striking. Although per-token pricing for GPT-5.5 has doubled compared to GPT-5.4 — now $5 per 1M input tokens and $30 per 1M output tokens — a roughly 40% reduction in token usage means the total cost to run the full Intelligence Index on GPT-5.5 (xhigh) is only about 20% higher than its predecessor. More notably, GPT-5.5 (xhigh) is approximately 30% cheaper to run the full index than Claude Opus 4.7 (max), and GPT-5.5 (medium) delivers an equivalent benchmark score at roughly one-quarter the cost (~$1,200 vs. ~$4,800).

However, the headline-grabbing benchmark scores come with important caveats. GPT-5.5 (xhigh) exhibited a hallucination rate of 86% on the AA-Omniscience factual knowledge benchmark — far higher than Claude Opus 4.7 (max) at 36% — meaning it frequently provides confidently incorrect answers. Additionally, on SWE-Bench Pro, which tests real-world GitHub issue resolution, GPT-5.5 scored 58.6%, falling short of Claude Opus 4.7's 64.3%, though questions about data contamination in that benchmark complicate the comparison.