AI BriefingArtificial Analysis Intelligence Index v5 Reveals 13x Cost Variance Across Frontier Models
·2026.10.29 16:22
Key point
The new index requires 67% more output tokens than v4, with median task length rising from 35K to 59K tokens.
Details
- Artificial Analysis launched Intelligence Index v5, expanding coverage to agentic coding, science, and knowledge work while keeping 60% of evaluation datasets private.
- v5 tasks demand 67% more output tokens than v4, increasing the median from 35K to 59K tokens per task.
- Evaluation costs vary by more than 13x across frontier models, ranging from $0.96 to $13.04 per task.
- For long agentic coding sessions, cost is dominated by cache reads, typically comprising 50-60% of total expenses, followed by cache writes (20-30%), output (15-25%), and input (1-10%).
- A weighted leaderboard prioritizes cache read cost with 70% weight and intelligence score with 30% weight to identify Pareto efficiency.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Kimi K3.1 Pro Scores 55 on Artificial Analysis Intelligence Index
SlopCodeBench October 2026: GPT-6 Astra leads strict pass rate, XHigh effort yields diminishing returnsStarSkirmish Bench v0.2: GPT-6 Astra Leads LLM StarCraft Bot Coding, Human Bots Still DominateOpenAI releases GPT-6.1 Terra with 50% lower API costs and improved computer use