Artificial Analysis puts 60 percent of its index on private tests, and Opus 5.5 leads at 58

What happened

Artificial Analysis published Intelligence Index v5 on October 27, 2026, and said it is the biggest change since v4 in January. Sixty percent of the weight is now on evaluations with private test sets or held-out answers. Two evaluations are new: FrontierCode 2.0, 132 private coding tasks run three times, and Terminal-Bench-Science, 70 research tasks in a terminal. Three are upgraded, including a 1,000-question HLE-Diamond in place of Humanity's Last Exam. GDPval-AA, SciCode, and AA-LCR left the index. On the composite, Claude Opus 5.5 at max effort scores 58, GPT-6 Astra 57, and GPT-6.1 Sol 55. Gemini 4 Argon at high effort is next at 50. Artificial Analysis says the open-weights gap widened from 3 points on v4.3 to 12 on v5. Kimi K3.1 Pro scores 46, and Artificial Analysis notes the weights are not released. DeepSeek V4.1 Flash 1017 and GLM-5.3 score 32, which is 26 points behind the leader and below Claude Haiku 5.5 at 36. Of the ten highest scores, GPT-6.1 Terra is the cheapest at $0.96 per task and Claude Fable 5.1 the most expensive at $13.04. Claude Opus 5.5 scores 35 at low effort and 58 at max, and the cost per task goes from $0.98 to $10.04. No model leads every evaluation. The same post introduces a separate Progress Index for change over time, and says the Intelligence Index remains the measure for current use.

Why it matters

People quote one Artificial Analysis number as if it were the same test as last month. v5 changed the tests and moved most of the weight onto questions the public cannot study. A model that drops on this index did not necessarily get worse at the old one. The open-weights gap Artificial Analysis reports is wider on the new mix, and the cheap model near the top is GPT-6.1 Terra, not the leader.

On FrontierCode 2.0, Opus 5.5 scores 46 percent. Astra, Terra, and Sonnet 5.5 each score 42 percent. An attempt that fails a blocking check scores zero.

On Terminal-Bench-Science, Astra leads at 63 percent, Opus at 59, and Sol at 58. The best released open weights, DeepSeek's October 17 checkpoint, score 13 percent.

The median model writes about 59,000 output tokens per index task, up from 35,000 on v4.3. Four agent evaluations are 55 percent of the weight and about 90 percent of those tokens.