On October 8, 2026, a benchmark update reported that Claude Haiku 5.5 at maximum effort scored 43 on the Intelligence Index, with a weighted task cost of $0.21. That’s a benchmark average—not a flat API charge. Anthropic bills API use by input and output tokens, so the price of a particular request depends on its token counts and prompt-length tier.
What the benchmark task-cost figure measures
The $0.21 figure is the weighted average cost of a task across Intelligence Index evaluations. The calculation includes input, cached input, cache writes, reasoning and answer tokens, with the evaluations weighted in the average. The reported score for Claude Haiku 5.5 at maximum effort was 43.
That figure describes work in a specific benchmark, not a universal price for summarizing a document, writing code or handling a customer request. Those jobs can use different numbers of input and output tokens, and different amounts of reasoning.
Anthropic listed Claude Haiku 5.5 as released on October 7, 2026. For the original launch context and tiered pricing, see our earlier coverage of Claude Haiku 5.5.
Anthropic’s API rates by prompt length
Anthropic’s published API rates are in U.S. dollars per million tokens. The lower rates apply to prompts up to 100,000 tokens; longer prompts use the higher tier.
| Prompt length | Input per 1 million tokens | Output per 1 million tokens |
| Up to 100,000 tokens | $0.10 | $0.50 |
| Over 100,000 tokens | $0.50 | $2.50 |
For a simple request with 10,000 input tokens and 1,000 output tokens below the threshold, the uncached charge works out to $0.0015: $0.001 for input plus $0.0005 for output. At the higher tier, 120,000 input tokens and 2,000 output tokens cost $0.065 without caching: $0.06 for input plus $0.005 for output.
The prompt-length threshold and context window are separate limits. Anthropic lists a one-million-token context window, but the API pricing tier changes at 100,000 prompt tokens.
How the reported GPT-6 Luna comparison fits
The reported maximum-effort comparison put GPT-6 Luna at a similar Intelligence Index score for about one-third of Claude Haiku 5.5’s cost. It also put Haiku 5.5 at roughly 162,000 output tokens per benchmark task, compared with about 50,000 for Luna. Those figures belong to that benchmark and effort setting; they don’t set the cost of other workloads.
A separate high-effort comparison reported a score of 38 for each model, with about 55,000 output tokens per Haiku 5.5 task and 50,000 per GPT-6 Luna task. These token-use figures describe that evaluation; they are not an API price comparison.
What changes an actual API bill
Four things matter in day-to-day use: prompt length, input volume, generated output and any cached tokens. Anthropic lists separate cache rates: cache reads cost $0.01 per million tokens in the lower tier and $0.05 in the higher tier. Cache writes cost $0.125 per million tokens for a five-minute cache and $0.20 for a one-hour cache in the lower tier; above 100,000 prompt tokens, those rates are $0.625 and $1.00. Anthropic also lists a 50% discount on input and output prices through its Batch API.
Reasoning effort matters, too. Claude Haiku 5.5 supports adaptive thinking, with medium effort as the default; more reasoning can change token use and the resulting charge. Anthropic also says its newer tokenizer counts about 30% more tokens for the same text than Haiku 4.5’s tokenizer. That is a comparison of token counts, not a fixed increase in every task’s bill.
