Opus 5.5 tops Artificial Analysis index as Anthropic cuts prices

Claude Opus 5.5, the newest Opus model from Anthropic, has the highest score Artificial Analysis has recorded on its Intelligence Index, which combines ten evaluations the independent benchmarking firm runs itself (Artificial Analysis on X). At its maximum effort setting the model scores 58, which the firm says is several points above anything it has measured before. It leads on six of the ten tests, including Humanity's Last Exam at 61.4% against a previous best of 59.1% from Claude Fable 5.1. On Terminal-Bench 4.0, a test of agents working at a command line, it scores 59.6%, level with OpenAI's GPT-6 Astra and 11 points above Opus 5. It trails slightly on three tests: CritPt, AA-LCR and GDP.pdf.

The lower price per token does not make each task cheaper. Anthropic cut Opus list prices 20%, to $4 per million input tokens and $20 per million output tokens from $5 and $25 for Opus 5, and cut cache reads, the price charged when a request reuses input the model has already processed, from $0.50 to $0.20 per million. But at max effort Opus 5.5 writes about 119,000 output tokens per index task, against about 73,000 for Opus 5 and about 27,000 for GPT-6 Astra. Artificial Analysis says the two effects roughly cancel, leaving cost per task level with Opus 5.

In a follow-up post, the firm broke that down: Opus 5.5 at max costs $5.98 per Intelligence Index task against $5.86 for Opus 5. At Opus 5's prices, the extra tokens alone would have pushed it to about $10.51. The 20% base cut brings that to $8.41, and the cheaper cache reads take it the rest of the way, saving slightly more than the headline cut does. Cheaper cache reads help most in long agent sessions, which resend the same context on every step. For someone already paying for Opus, the upgrade buys the top index score at about the same cost per task rather than a lower bill.