Opus 5.5 leads the index and burns 260M tokens doing it

Claude Opus 5.5, which Anthropic released on September 22, scores 58 on the Artificial Analysis Intelligence Index at max effort, first among the 206 models it is compared with. At high effort, Artificial Analysis puts it at 54, third of 212, and that run cost about a quarter as much to evaluate.

Both runs use adaptive reasoning with default fallback. The only difference between them is the effort setting. Index v4.3.2 combines 10 evaluations, among them Terminal-Bench 4.0 for agentic coding and terminal work, SciCode, Humanity’s Last Exam, the physics test CritPt, the long-context reasoning test AA-LCR v1.1 and AA-Omniscience, which rewards correct answers and penalizes hallucinated ones. The median score among comparable models is 25 in both rankings. For a proprietary model, comparable means the same price band, measured on a blended 3:1 input-to-output price.

What the last four points cost

Running the full index at max effort cost $8,708.20. At high effort it cost $2,172.43. Artificial Analysis puts the weighted average cost per task at $5.98 for max and $1.82 for high, so one max-effort task costs more than three times as much. Per-token prices are the same in both runs. What differs is how much the model writes: 260M output tokens across the index at max effort, 53M at high.

Those token counts land on opposite sides of the median. Artificial Analysis calls the max run “very verbose” against a class median of 92M tokens, and the high run “fairly concise” against a median of 88M. Max ranks 89th on verbosity and high ranks 33rd. Max effort buys only four more points on the score, from 54 to 58.

A reasoning model writes out its working before it answers. When Artificial Analysis builds the cost per task, it prices reasoning tokens alongside answer tokens, input, cache hits and cache writes.

At high effort, Opus 5.5 produces 90.7 output tokens per second. That is faster than the class average of 71 and ranks 59th of 212.

Price against Opus 5

On Anthropic’s API, Opus 5.5 costs $4.00 per million input tokens and $20.00 per million output tokens. Artificial Analysis rates both as “somewhat expensive” against class medians of $2.00 and $10.00. On cost per task, max ranks 87th of 206 and high ranks 65th of 212. The model lists a 95% cache discount. That matches the $0.20 per million cache-read price Anthropic gives on its Opus page, which is 60% below Opus 5. With a 7:2:1 mix of cache hits, input and output, Artificial Analysis gets a blended rate of $2.94 per million tokens.

Anthropic puts the per-token price 20% below Opus 5, which shipped on July 24. The company also says Opus 5.5 uses fewer tokens per task, so work billed by the token costs about 40% less than on Opus 5. It describes that figure as an estimate for typical workloads. The effort setting moves the bill as well: on the same model, the high and max runs cost $1.82 and $5.98 per task.

Fast mode is up to 2.5x faster and costs $8 per million input tokens and $40 per million output tokens. US-only inference costs 1.1x the standard price. Artificial Analysis lists one API provider for the model, Anthropic. Anthropic also offers it on AWS, Google Cloud and Microsoft Foundry under the API name claude-opus-5-5.

Max or high?

The rest of the spec is the same in both configurations: text and image input, text output, a 1M-token context window and proprietary weights. Anthropic has not disclosed the parameter count. Artificial Analysis also places both runs on its Finance & Accounting index, which uses seven of the same evaluations.

The index is a single test bed, and max effort doesn’t come out ahead everywhere. In CodeRabbit’s review pipeline in September, Opus 5.5 at Max caught 50 of 80 known bug patterns. The lower-effort Standard setup caught 51, and Max left more comments to read.