AIEvals

Benchmarks

HealthBench (length-adjusted)

Quality of healthcare responses under a rubric score adjusted to account for response length rather than raw scoring alone.

HealthcareReported by AnthropicVendor-reported onlyReported by Anthropic

How it is run

Section 8.11 uses safety classifiers and Claude Opus 4.8 as the common judge. Haiku 5.5 max effort averages five runs; lower settings have one run each.

What it cannot tell you

Length-adjusted and raw scores are separate metrics. Other models were timed on different days and servers, so timing curves are not a shared speed test.

Why there is no independent table

No independent suite in this dataset runs HealthBench (length-adjusted). The numbers below are the ones model publishers report about their own models, which is a different kind of evidence: each publisher chooses its own harness and effort level, so a row is readable on its own and not against the other rows.

Read this with

Other versions of the same evaluation. Versions are not comparable: the task set changes between them.

  • HealthBench (raw)Same family. Raw healthcare response quality before the length adjustment used in Anthropic’s HealthBench reporting.
  • HealthBench Professional (length-adjusted)Same category. Professional healthcare response quality after Anthropic’s response-length adjustment, with deployed safety classifiers enabled.
  • HealthBench Professional (raw)Same category. Raw professional healthcare rubric scores before Anthropic’s response-length correction is applied.
  • PhysicianBenchSame category. Successfully carrying out 100 physician tasks inside an electronic health record, measured by the share of attempts passed.

As reported by each publisher

Kept separate per publisher, because the settings are not shared.

Anthropic

Claude Haiku 5.5: system card capability evaluations · 7 October 2026

RankModelScore
1Claude Haiku 5.561.6%
2GPT-6 Luna56.3%

Ranks compare reported scores in the same published row; ties share a rank.

HealthBench (length-adjusted). §8.11.3 / Figure 8.11.1.A: max mean of five runs; low/medium/high/xhigh 59.9/60.2/60.3/61.1%, one run each. Opus 4.8 judge, classifiers on.

Section 8, pages 111–139 of the Haiku 5.5 system card, read 8 October 2026. Values come from exact prose, tables and labelled figures. Unless specified, Haiku uses adaptive thinking at max effort and five trials with evaluation-specific context up to 1M. Each row retains its subset, metric, effort and source notes; no score is assigned to an unspecified SWE-Bench Pro version. Terminal runs use safeguards without fallback and no internet egress. ProgramBench uses filtered tasks and hidden-test pass rate without a time limit. OSWorld partial and strict results use the official 82-task offline subset. CAD uses 0–1 voxel IoU. Healthcare raw and length-adjusted scores stay separate. Life-science results disable Haiku biology safeguards. Independently run figures quoted by Anthropic remain publisher reports here. Missing comparison values are not inferred from rounded charts.

www.anthropic.com/claude-haiku-5-5-system-card

Sources

Every page used for this benchmark, read on 17 October 2026.