Clinical Benchmarks

Boards by clinical job

Each board tests one clinical job. The strip shows every row on the board, with the leader marked.

Clinical reasoning and knowledge

Consults, diagnosis and management questions, image-based cases and broad clinical suites.

  • HealthBench Professional

    Physician-selected workplace tasks, from care consults to documentation and research, graded on physician-written rubrics.

    OpenAI, 525 tasks

    Leads this board
    43 rowsmeasured to Oct 2026
  • MedXpertQA (MM)

    Expert-level multiple-choice questions over clinical images across 17 specialties.

    Tsinghua University, 2,000 multimodal questions

    Leads this board
    22 rowsmeasured to Aug 2026
  • MAST (Medical AI Superintelligence Test)

    A composite of clinical benchmarks spanning diagnostic and management reasoning, safety, multimodal imaging, and agentic capability.

    ARISE AI Research Network, 6-benchmark composite

    Leads this board
    8 rows of 11 on the boardmeasured to Aug 2026
  • Bedside Bench

    Physician-written clinical cases on reasoning and safety, from differential diagnosis to medication safety.

    Doximity, 500 clinical cases

    Leads this board
    15 rows of 16 on the boardmeasured to Oct 2026
  • APEX-1 GP

    Primary-care consults graded by practising GPs on diagnosis, workup and safe escalation.

    Mercor, 100 GP consults

    Leads this board
    12 rowsmeasured to Oct 2026

Documentation and coding

Notes written from visits and diagnosis codes assigned to hospital stays.

  • MedScribe (Vals AI)

    Quality and compliance of SOAP notes generated from clinical visits, scored against documentation rubrics.

    Vals AI, 100 SOAP-note cases

    Leads this board
    106 rowsmeasured to Oct 2026
  • MedCode (Vals AI)

    ICD-10-CM coding of whole hospital stays from discharge summaries and notes, checked against professional coders.

    Vals AI, 2,755 patient records

    Leads this board
    104 rowsmeasured to Oct 2026
  • MLCR-AA

    Questions over long, fragmented medical records; a pass needs a complete answer with nothing that contradicts the record.

    Artificial Analysis, 60 record questions

    27 rowsmeasured to Nov 2026

EHR and workflow agents

Agents that query records, run multi-step physician workflows and complete administrative work.

  • PhysicianBench

    Agents carrying out long-horizon physician workflows inside real EHR systems, verified by execution against those systems.

    academic team, 100 clinical tasks

    Leads this board
    28 rowsmeasured to Oct 2026
  • HealthAgentBench

    Agent harnesses completing realistic terminal-based healthcare tasks built from real clinical artifacts.

    Microsoft Research, 54 agentic tasks

    Leads this board
    12 rowsmeasured to Jul 2026
  • CHI-Bench

    Long-horizon healthcare operations workflows for agents: prior authorization, utilization management, and care management.

    actAVA, 75 operations workflows

    Leads this board
    43 rows of 45 on the boardmeasured to Aug 2026
  • RCM-Bench

    Revenue cycle tasks: claims, billing, insurance updates and eligibility checks, checked on the resulting system state.

    Commure, 97 RCM tasks

    Leads this board
    20 rowsmeasured to Oct 2026
  • DAYJOB Health

    Long agent assignments from healthcare professionals, from clinical decision support to pharmacy and payer work.

    Surge AI, 50 assignments

    Leads this board
    15 rowsmeasured to Oct 2026

Safety

Whether recommendations could harm a patient, and whether safety rules hold when patient details change.

  • First, Do NOHARM (v2)

    How often, and how severely, model consultation recommendations contain potentially harmful errors.

    Stanford/Harvard consortium, 1,100 consultation cases

    Leads this board
    LiSA 2.586.2
    17 rows of 19 on the boardmeasured to Sep 2026

Benchmarks not tracked

16 considered and left out, each with the reason.
  • HealthBench ConsensusNear-saturated physician-consensus baseline; frontier runs stopped reporting it separately.
  • MedQA / MultiMedQAExam-style multiple choice, saturated above 95 percent since 2025; archived by its trackers.
  • AgentClinicNo public frontier-model results since 2025.
  • CRAFT-MDNo public frontier-model results since 2025.
  • MedAgentBenchV2 lives on inside the MAST composite; the standalone board has no current frontier rows.
  • SDBench / MAI-DxOMicrosoft's 2025 sequential-diagnosis study was not re-run on current models.
  • Open Medical-LLM Leaderboard (Hugging Face)Built on saturated exam sets; no frontier submissions in 2026.
  • MedArenaClinician preference arena; ratings pool too thin on current frontier models to quote.
  • AMIE evaluationsGoogle DeepMind research prototypes, never opened to cross-vendor comparison.
  • LiveClin, PrIME-LLM, MedMCP-CalcSingle studies with two or fewer current-frontier rows; tracked for a future qualifying update.
  • EHR-ComplexNot run on current models; newest result is from February 2026 and no Claude 5, GPT-6 or Gemini 4 model has a score.
  • MedHELMNot run on current models; the newest Claude on the board is Opus 4.6 and the task set adds little beyond our other clinical boards.
  • MedPIC-BenchA single study that has not been run on current models; newest Claude result is Sonnet 4.6.
  • HealthAdminBenchThe official board still lists only the five models from the April 2026 paper; later vendor runs used a different setup and are not comparable.
  • HealthBench HardOnly OpenAI and Meta publish scores; no Anthropic or Google model has an official result, so the board cannot compare labs.
  • WHBenchNot run on current models; newest results are from March 2026.