Clinical Benchmarks

DeepSeek V4 Pro

Released 24 Apr 2026$1.32 input, $3.96 output per million tokens1M context1.6T total / 49B activemitAlso written as DS-V4-Pro Max, DeepSeek-V4-Pro, deepseek/deepseek-v4-pro-0813, deepseek/deepseek-v4-pro, DeepSeek V4 Pro 0813, deepseek-v4, deepseek-v4-pro, DeepSeek-V4-pro, deepseek-v4-pro-0813, DeepSeek V4, DeepSeek-V4-Pro-Max

11 results on 6 of 14 boards, latest measured Oct 2026. Compare with other models

Clinical Benchmarks Index41.8Rank 89 of 1285 of 12 boardsThis site’s own calculation. Show the working How it is computed

On the Clinical Benchmarks Index

The top 15 on the index, then DeepSeek V4 Pro at rank 89. Select a column to open that model.

Board by board

DeepSeek V4 Pro against the top 10 on each board it has a result on, on that board's own scale. Its own column is in colour.

Bedside Benchrank 15 of 15

  1. 94.6Claude Opus 5
  2. 94.3Muse Spark Open
  3. 93.4GPT-6 Astra
  4. 92.7Gemini 4 Argon
  5. 91.1GPT-5.6 Sol
  6. 89.7Kimi K3
  7. 89.1Ember-1
  8. 88.1GPT-6 Sol
  9. 87.2GPT-6 Luna
  10. 80.6DeepSeek V4 Pro 0813

15 modelsFull board

APEX-1 GPrank 11 of 12

  1. 73.04Claude Opus 5.5
  2. 72.12Claude Opus 5
  3. 68.36Gemini 3.8 Flash
  4. 64.35Kimi K3
  5. 63.75GPT-5.6 Sol
  6. 63.2GLM 5.3
  7. 61.03GPT-6 Astra
  8. 60.61GPT-6.1 Sol
  9. 59.38Grok 4.6
  10. 56.53DeepSeek V4 Pro 0813

12 modelsFull board

MedScribe (Vals AI)rank 56 of 106

  1. 91.43Claude Opus 5.5
  2. 91.29Claude Fable 5.1
  3. 91.10Claude Sonnet 5.5
  4. 90.98Claude Opus 5
  5. 90.06Muse Spark 1.2
  6. 89.38Grok 4.7
  7. 88.94GLM 5.3 Flash
  8. 88.89Muse Spark 1.1
  9. 88.81GLM 5.3
  10. 80.17DeepSeek V4 Pro 0813

95 modelsFull board

MedCode (Vals AI)rank 48 of 104

  1. 63.57Claude Opus 5
  2. 59.06Gemini 3.1 Pro Preview (02/26)
  3. 58.80Gemini 4 Argon
  4. 56.07Claude Fable 5
  5. 55.92Gemini 3 Flash (12/25)
  6. 55.83Gemini 3.5 Flash
  7. 54.86Claude Opus 4.7
  8. 53.51Claude Fable 5.1
  9. 53.39Gemini 3.7 Flash
  10. 42.47DeepSeek V4 Pro 0813

95 modelsFull board

PhysicianBenchrank 20 of 28

  1. 68.4Claude Opus 5.5 (max)
  2. 63.2Claude Sonnet 5.5 (max)
  3. 61.0Claude Fable 5.1 (max)
  4. 57.6Claude Opus 5 (max)
  5. 46.3GPT-5.5
  6. 43.0Claude Haiku 5.5 (max)
  7. 37.4Claude Sonnet 5 (max)
  8. 33.6GPT-6 Luna (Anthropic run)
  9. 31.7Claude Opus 4.6
  10. 18.7DeepSeek V4-Pro

20 modelsFull board

CHI-Benchrank 24 of 43

  1. 54.7erius + claude-opus-5
  2. 37.3erius + claude-opus-4-8
  3. 28.0claude-code + claude-opus-4-6
  4. 26.2claude-code + claude-sonnet-4-6
  5. 25.3codex + gpt-5.6-sol
  6. 25.3openai-agents + kimi-k3
  7. 24.4claude-code + claude-opus-4-7
  8. 24.0claude-code + claude-fable-5
  9. 20.9codex + gpt-5.5
  10. 14.2openai-agents + deepseek-v4-pro

24 modelsFull board

Every result, with its source

Each line sits on its own board's scale. Tall tick: the board leader.

Clinical reasoning and knowledge

  1. Bedside Bench
    Score (avg@3)
    80.6
    Rank 15 of 15 here, 16 models on the boardLeader Claude Opus 5 94.6
    Official leaderboard
    Measured Oct 2026
  2. APEX-1 GP
    Score (avg@4), reasoning max
    56.53
    Rank 11 of 12Leader Claude Opus 5.5 73.04
    Official leaderboard
    Measured Oct 2026

Documentation and coding

  1. MedScribe (Vals AI)
    model ID deepseek/deepseek-v4-pro-0813; max_output_tokens=30000; reasoning_effort=max
    80.17
    Rank 56 of 106 here, 105 models on the boardLeader Claude Opus 5.5 91.43
    Official leaderboard
    Measured Sep 2026
  2. MedScribe (Vals AI)
    model ID deepseek/deepseek-v4-pro; max_output_tokens=128000; reasoning_effort=max
    75.14
    Rank 78 of 106 here, 105 models on the boardLeader Claude Opus 5.5 91.43
    Official leaderboard
    Measured Sep 2026
  3. MedCode (Vals AI)
    model ID deepseek/deepseek-v4-pro-0813; reasoning_effort=max; max_output_tokens=30000
    42.47
    Rank 48 of 104 here, 103 models on the boardLeader Claude Opus 5 63.57
    Official leaderboard
    Measured Sep 2026
  4. MedCode (Vals AI)
    model ID deepseek/deepseek-v4-pro; reasoning_effort=max; max_output_tokens=128000
    40.45
    Rank 62 of 104 here, 103 models on the boardLeader Claude Opus 5 63.57
    Official leaderboard
    Measured Sep 2026

EHR and workflow agents

  1. PhysicianBench
    Pass@1 over 3 runs; Pass^3 6.0; shared FHIR tool harness; up to 100 turns; high reasoning when supported
    18.7
    Rank 20 of 28 here, 12 models on the boardLeader Claude Opus 5.5 (max) 68.4
    Official leaderboard
    Measured May 2026
  2. CHI-Bench
    Listed as openai-agents + deepseek-v4-pro
    All Domains pass@1; openai-agents harness
    14.2
    Rank 24 of 43 here, 45 models on the boardLeader erius + claude-opus-5 54.7
    Official leaderboard
    Measured May 2026
  3. CHI-Bench
    Listed as hermes + deepseek-v4-pro
    All Domains pass@1; hermes harness
    13.8
    Rank 25 of 43 here, 45 models on the boardLeader erius + claude-opus-5 54.7
    Official leaderboard
    Measured May 2026
  4. CHI-Bench
    Listed as openclaw + deepseek-v4-pro
    All Domains pass@1; openclaw harness
    11.1
    Rank 29 of 43 here, 45 models on the boardLeader erius + claude-opus-5 54.7
    Official leaderboard
    Measured May 2026
  5. CHI-Bench
    Listed as deepagents + deepseek-v4-pro
    All Domains pass@1; deepagents harness
    10.7
    Rank 31 of 43 here, 45 models on the boardLeader erius + claude-opus-5 54.7
    Official leaderboard
    Measured May 2026

Index working

Show the working: each board’s result, how hard the board runs and the field average

The Clinical Benchmarks Index is this site's own calculation from 12 boards. How it is computed

  • MedScribe (Vals AI)80.1780.2+33.8 easy board46.40.15 weight, near its ceilingNear its ceiling: leader 91.4, weight 0.15model ID deepseek/deepseek-v4-pro-0813; max_output_tokens=30000; reasoning_effort=max
  • MedCode (Vals AI)42.4742.5−3.3 hard board45.80.319 weightmodel ID deepseek/deepseek-v4-pro-0813; reasoning_effort=max; max_output_tokens=30000
  • PhysicianBench18.718.7−16.4 hard board35.10.79 weightPass@1 over 3 runs; Pass^3 6.0; shared FHIR tool harness; up to 100 turns; high reasoning when supported
  • Bedside Bench80.680.6+39.2 easy board41.40.25 weight, near its ceilingNear its ceiling: leader 94.6, weight 0.25Score (avg@3)
  • APEX-1 GP56.5356.5+11.9 easy board44.60.674 weightScore (avg@4), reasoning max
Adjusted average of 5 boards, by weight
41.1
Field average, counts for 23.1% with 5 of 12 boards
44.2
Index
= 41.8

Score: the result on the board’s own 0 to 100 scale (dot; tick: the board’s leader). Board runs: how far the board’s scores sit above or below the same models’ scores on the other boards. Adjusted: the score minus that offset. The index is the adjusted average with the field average mixed in, more of it the fewer boards a model has. Each result links to its quote and page; the line under it is the setting it was measured with.

Sources

The quote, page and document for each of the 11 results
  1. 56MedScribe (Vals AI) deepseek-v4-pro-0813 80.17
    Printed as 80.17%Official leaderboard, measured Sep 2026Configuration: model ID deepseek/deepseek-v4-pro-0813; max_output_tokens=30000; reasoning_effort=max
    Vals AI MedScribe leaderboard official leaderboard, Vals AI, 3 Sep 2026. Vals AI MedScribe leaderboard, View: All Models, Task: Overall, row 55 of 105 (DeepSeek V4 Pro 0813), Accuracy column; Updated 9/29/2026. Rendered BenchmarkView table; configuration from embedded astro-island BenchmarkView props, benchmarkView.default.tasks.overall["deepseek/deepseek-v4-pro-0813"].
    55 | DeepSeek V4 Pro 0813 | 80.17%±2.00 | $1.32/$3.96 | 2m36s
    Every result from this document
  2. 78MedScribe (Vals AI) deepseek-v4-pro 75.14
    Printed as 75.14%Official leaderboard, measured Sep 2026Configuration: model ID deepseek/deepseek-v4-pro; max_output_tokens=128000; reasoning_effort=max
    Vals AI MedScribe leaderboard official leaderboard, Vals AI, 3 Sep 2026. Vals AI MedScribe leaderboard, View: All Models, Task: Overall, row 77 of 105 (DeepSeek V4), Accuracy column; Updated 9/29/2026. Rendered BenchmarkView table; configuration from embedded astro-island BenchmarkView props, benchmarkView.default.tasks.overall["deepseek/deepseek-v4-pro"].
    77 | DeepSeek V4 | 75.14%±2.00 | $1.32/$3.96 | 5m46s
    Every result from this document
  3. 48MedCode (Vals AI) deepseek-v4-pro-0813 42.47
    Printed as 42.47%Official leaderboard, measured Sep 2026Configuration: model ID deepseek/deepseek-v4-pro-0813; reasoning_effort=max; max_output_tokens=30000
    Vals AI MedCode leaderboard official leaderboard, Vals AI, 29 Sep 2026. MedCode leaderboard, Overall task, All Models expanded, rank 47 of 103; Accuracy column; board Updated 9/29/2026. Rendered table row; configuration from embedded BenchmarkView props benchmarkView.default.tasks.overall["deepseek/deepseek-v4-pro-0813"].
    47 | DeepSeek V4 Pro 0813 | 42.47%±2.16 | $1.32/$3.96 | 3m01s
    Every result from this document
  4. 62MedCode (Vals AI) deepseek-v4-pro 40.45
    Printed as 40.45%Official leaderboard, measured Sep 2026Configuration: model ID deepseek/deepseek-v4-pro; reasoning_effort=max; max_output_tokens=128000
    Vals AI MedCode leaderboard official leaderboard, Vals AI, 29 Sep 2026. MedCode leaderboard, Overall task, All Models expanded, rank 61 of 103; Accuracy column; board Updated 9/29/2026. Rendered table row; configuration from embedded BenchmarkView props benchmarkView.default.tasks.overall["deepseek/deepseek-v4-pro"].
    61 | DeepSeek V4 | 40.45%±2.12 | $1.32/$3.96 | 6m23s
    Every result from this document
  5. 20PhysicianBench Pass@1 over 3 runs 18.7
    Printed as 18.7 ± 2.9Official leaderboard, measured May 2026Configuration: Pass@1 over 3 runs; Pass^3 6.0; shared FHIR tool harness; up to 100 turns; high reasoning when supported
    PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments (arXiv 2605.02240v1 PDF) paper, Stanford University (HealthRex; Liu, Chen et al.), 4 May 2026. arXiv HTML 2605.02240v1, Section 5.2, Table 2; DeepSeek V4-Pro row, Pass@1 column (PDF p. 8).
    Model | Pass@1 | Pass@3 | Pass^3 | #Turns DeepSeek V4-Pro | 18.7 ± 2.9 | 27.9 | 6.0 | 35.3
    Every result from this document
  6. 24CHI-Bench All Domains pass@1 14.2
    Printed as 14.2%Official leaderboard, measured May 2026Configuration: All Domains pass@1; openai-agents harness
    CHI-Bench leaderboard (actAVA) official leaderboard, actAVA, 12 Aug 2026. CHI-Bench v1.0.0, All Domains, rank 24, Agent/Model and Accuracy columns; PA/UM/CM follow; board last updated 2026-08-12; board Date column 2026-05-01; accessed 2026-09-30. Linked submission provenance: run 2026-05-03T08:35:38Z to 2026-05-04T06:59:15Z; submitted_at 2026-05-04T06:59:15Z.
    24 | openai-agents | deepseek-v4-pro | Open-source | 14.2% | 10.7% | 28.0% | 4.0% | 2026-05-01
    Every result from this document
  7. 25CHI-Bench All Domains pass@1 13.8
    Printed as 13.8%Official leaderboard, measured May 2026Configuration: All Domains pass@1; hermes harness
    CHI-Bench leaderboard (actAVA) official leaderboard, actAVA, 12 Aug 2026. CHI-Bench v1.0.0, All Domains, rank 25, Agent/Model and Accuracy columns; PA/UM/CM follow; board last updated 2026-08-12; board Date column 2026-05-01; accessed 2026-09-30. Linked submission provenance: run 2026-05-03T06:28:45Z to 2026-05-04T05:51:44Z; submitted_at 2026-05-04T05:51:44Z.
    25 | hermes | deepseek-v4-pro | Open-source | 13.8% | 8.0% | 25.3% | 8.0% | 2026-05-01
    Every result from this document
  8. 29CHI-Bench All Domains pass@1 11.1
    Printed as 11.1%Official leaderboard, measured May 2026Configuration: All Domains pass@1; openclaw harness
    CHI-Bench leaderboard (actAVA) official leaderboard, actAVA, 12 Aug 2026. CHI-Bench v1.0.0, All Domains, rank 29, Agent/Model and Accuracy columns; PA/UM/CM follow; board last updated 2026-08-12; board Date column 2026-05-01; accessed 2026-09-30. Linked submission provenance: run 2026-05-02T09:01:51Z to 2026-05-04T03:00:01Z; submitted_at 2026-05-04T03:00:01Z.
    29 | openclaw | deepseek-v4-pro | Open-source | 11.1% | 14.7% | 12.0% | 6.7% | 2026-05-01
    Every result from this document
  9. 31CHI-Bench All Domains pass@1 10.7
    Printed as 10.7%Official leaderboard, measured May 2026Configuration: All Domains pass@1; deepagents harness
    CHI-Bench leaderboard (actAVA) official leaderboard, actAVA, 12 Aug 2026. CHI-Bench v1.0.0, All Domains, rank 31, Agent/Model and Accuracy columns; PA/UM/CM follow; board last updated 2026-08-12; board Date column 2026-05-01; accessed 2026-09-30. Linked submission provenance: run 2026-05-03T07:35:56Z to 2026-05-05T01:34:20Z; submitted_at 2026-05-05T01:34:20Z.
    31 | deepagents | deepseek-v4-pro | Open-source | 10.7% | 14.7% | 10.7% | 6.7% | 2026-05-01
    Every result from this document
  10. 15Bedside Bench Score 80.6
    Printed as 80.6Official leaderboard, measured Oct 2026Configuration: Score (avg@3)
    Bedside Bench on the Fireworks Specialized Intelligence Index official leaderboard, Fireworks AI, 22 Oct 2026. https://fireworks.ai/specialized-intelligence-index/benchmarks/healthcare/bedside-bench (feed: benchmarks[id=healthcare/bedside-bench].results)
    DeepSeek V4 Pro 0813 80.6 (Score (avg@3)), rank 16
    Every result from this document
  11. 11APEX-1 GP Score 56.53
    Printed as 56.53Official leaderboard, measured Oct 2026Configuration: Score (avg@4), reasoning max
    APEX-1: General Practitioner on the Fireworks Specialized Intelligence Index official leaderboard, Fireworks AI, 26 Oct 2026. https://fireworks.ai/specialized-intelligence-index/benchmarks/healthcare/apex-1-general-practitioner (feed: benchmarks[id=healthcare/apex-1-general-practitioner].results)
    DeepSeek V4 Pro 0813 56.53 (Score (avg@4)), rank 11
    Every result from this document

Other DeepSeek models: DeepSeek R1, DeepSeek V4.1 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Flash, DeepSeek V4 Flash 0731