Skip to content

Benchmark

HealthBench Hard

Scores are percentages. Higher is better.

Models measured
19
Scores recorded
43
Sources
3
IQ weight
1.87

Score distribution

28.4%
Mean
31.4%
Median
9.5%
Std dev
10.8%
10th percentile
36.6%
90th percentile
1.6%
Min
1.6%42.8%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.87 for chat models, measured over 19 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Muse Spark· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim42.8%
2Muse Sparkraw score (no length adjustment), GPT-4.1 grader via the OpenAI simple-evals implementation, Muse Spark Thinking, Meta run, launch-post benchmark table.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.42.8%
3GPT 5.2GPT-5.2 Thinking · system-card implementation · raw score· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.42%
4GPT 5.2Raw/unadjusted HealthBench Hard score as evaluated by Baichuan in its M3 report, distinct from OpenAI’s length-adjusted protocol.· HealthBench Hard (Arcophos) data docs42%
5GPT-6 Astralength-adjusted, max reasoning effort (37.8 unadjusted, 2,192 mean response chars), GPT-6 Astra system card Table 6, column 'gpt-6 Astra'.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.36.6%
6GPT-6.1 SolGPT-6.1 Sol system card · length-adjusted· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.36.2%
7GPT-6.1 SolOpenAI, length-adjusted, 36.2 (33.4, 1646) (adjusted, raw, mean response characters), addendum Table 7 refers to the GPT-6 Astra card for method· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.36.2%
8GPT-6.1 Sol· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim36.2%
9GPT 5length-adjusted, max reasoning effort (41.6 unadjusted, 2,880 mean response chars), GPT-5.6 system card Table 6, column GPT-5. The GPT-5 launch system card printed 46.2% raw for gpt-5-thinking.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.34.7%
10GPT 5.2length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.2 (38.9 unadjusted, 2585 chars)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.34.3%
11GPT-6.1 SolGPT-6.1 Sol system card · unadjusted· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.33.4%
12GPT-5.6 Sol· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim33.1%
13GPT-5.6 Sollength-adjusted, max reasoning effort (31.1 unadjusted, 1,751 mean response chars), GPT-5.6 system card Table 6, column GPT-5.6-SOL. API max-effort setting, not the August ChatGPT production setting measured in row 490.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.33.1%
14GPT-5.6 Terra· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim32.7%
15GPT-5.6 Terralength-adjusted, max reasoning effort (34.3 unadjusted, 2,199 mean response chars), GPT-5.6 system card Table 6, column GPT-5.6-TERRA.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.32.7%
16GPT-5.6 Luna· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim32%
17GPT-5.6 Lunalength-adjusted, max reasoning effort (31.4 unadjusted, 1,923 mean response chars), GPT-5.6 system card Table 6, column GPT-5.6-LUNA. API max-effort setting, not the August ChatGPT production setting measured in row 491.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.32%
18GPT 5.5length-adjusted · no tools · GPT-5.5 system-card implementation· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.31.5%
19GPT 5.5length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.5 (33.8 unadjusted, 2289 chars)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.31.5%
20GPT-5.6 SolChatGPT production (Instant) deployment setting, length-adjusted, GPT-5.6 August Updates PDF column GPT-5.6 Sol (August) (27.1 unadjusted, 1,450 chars)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.31.4%
21GPT-6 LunaOpenAI, length-adjusted, maximum reasoning effort, 31.4 (25.4, 1241) (adjusted, raw, mean response characters)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.31.4%
22GPT-6.1 Terra· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim31.4%
23GPT-6.1 TerraOpenAI, length-adjusted, 31.4 (26.0, 1306) (adjusted, raw, mean response characters), addendum Table 5 refers to the GPT-6 Astra card for method· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.31.4%
24GPT-6 SolOpenAI, length-adjusted, maximum reasoning effort, 30.1 (22.1, 974) (adjusted, raw, mean response characters)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.30.1%
25GPT OSS 120B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim30%
26GPT OSS 120Bofficial model card Table 3 · reasoning high · published setting; Table 3 does not further specify tool access · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.30%
27GPT OSS 120Braw score (%), reasoning level high, gpt-oss model card Table 3. Not length-adjusted, unlike the GPT-5.x rows on this board.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.30%
28GPT 5.4length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.4 (30.3 unadjusted, 2161 chars)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.29.1%
29GPT-5.6 LunaChatGPT production (Instant) deployment setting, length-adjusted, GPT-5.6 August Updates PDF column GPT-5.6 Luna (August) (24.9 unadjusted, 1,523 chars)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.28.7%
30GPT OSS 120Bofficial model card Table 3 · reasoning medium · published setting; Table 3 does not further specify tool access · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.9%
31GPT-5.3 Chatlatest· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim25.9%
32GPT-5.3 Chatraw score (no length adjustment), GPT-5.3 Instant system card Table 3, column GPT-5.3-INSTANT, OpenAI later cards print 20.2 length-adjusted (17.8 unadjusted) for the re-run model.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.25.9%
33GPT 5.1length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.1 (41.4 unadjusted, 4049 chars)· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.25.4%
34GPT-5.5 Instant· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim22.9%
35GPT-5.5 Instantlength-adjusted (21.3 unadjusted, 1,794 mean response chars), GPT-5.5 Instant system card Table 5, column GPT-5.5 INSTANT.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.22.9%
36GPT OSS 120Bofficial model card Table 3 · reasoning low · published setting; Table 3 does not further specify tool access · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.22.8%
37Seed 2.0 Mini· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim15.3%
38GPT OSS 20bofficial model card Table 3 · reasoning medium · published setting; Table 3 does not further specify tool access · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.12.9%
39GPT OSS 20bofficial model card Table 3 · reasoning high · published setting; Table 3 does not further specify tool access · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.10.8%
40GPT OSS 20braw score (%), reasoning level high, gpt-oss model card Table 3. Not length-adjusted, unlike the GPT-5.x rows on this board.· HealthBench Hard (Arcophos) data docs· Reported by the model's developer. The source relaying it did not verify it.10.8%
41GPT OSS 20b· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim10.8%
42GPT OSS 20bofficial model card Table 3 · reasoning low · published setting; Table 3 does not further specify tool access · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.9%
43GPT 5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim1.6%

Data credits: Arcophos. HealthBench Hard Leaderboard. https://healthbenchhard.ai. CC BY 4.0. Source document linked on each result. (licence) · Data by LLM Stats · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.