Benchmark Almanac

Public Benefits Bench

Claude Opus 5 76.9% leads since July 2026
Not in model cards

Models answer SNAP benefit questions across seven lifecycle phases and 25 topics. A judge grades each answer against rubrics validated by SNAP policy experts.

459 scenarios with a 230-question test set. v1 tested four conditions with and without web search and multi-turn. The live board is v1.1, with v1 archived at /benchmarks/public-benefits-bench-v1.

Made by
Vals AI
Released
June 2026
Kind
Other professions
Level
Expert
Scored
Graded by a model
Answers
Open-ended
Tasks
Kept private
Strongest model tested
Claude Opus 5.5 #4 on the capability index
Last model added
October 2026
Scores refreshed
November 12th, 2026
50
Models scored on its own leaderboard

Claude Opus 5 leads with 76.9%

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

50 models. Scroll sideways for the rest.

  1. 76.9 Claude Opus 5
  2. 74.9 Claude Fable 5.1
  3. 73.8 Muse Wonder Full
  4. 70.6 Claude Opus 5.5
  5. 70.4 Claude Fable 5
  6. 70.2 Muse Wonder
  7. 69.8 Gemini 4 Argon
  8. 68.9 MiMo-V2.6 Pro
  9. 68.6 Tencent Hy4 preview
  10. 68.5 GLM-5.3
  11. 68.5 Muse Spark 1.2
  12. 68.2 Kimi K3
  13. 68.1 Claude Opus 4.8
  14. 67.6 MiMo-V2.6 Flash
  15. 67.2 Claude Sonnet 5.5
  16. 67.1 Qwen 3.8 Max
  17. 66.8 Grok 4.6
  18. 66.5 GPT-5.6 Sol
  19. 66.2 Ember-1
  20. 66.0 Claude Sonnet 5
  21. 65.6 Grok 4.7
  22. 65.3 Gemini 3.8 Flash
  23. 64.3 DeepSeek V4.1 Flash
  24. 64.1 MiniMax-M3
  25. 63.2 Claude Haiku 5.5
  26. 62.9 DeepSeek-V4-Pro
  27. 62.5 Claude Sonnet 4.6
  28. 62.4 GPT-5.6 Terra
  29. 61.8 GLM-5.1
  30. 61.3 Muse Spark 1.1
  31. 61.2 GPT-5.6 Luna
  32. 60.9 GPT-5.5
  33. 59.5 Gemini 3.5 Flash
  34. 59.4 Inkling-Small
  35. 59.3 GPT-6.1 Sol
  36. 58.5 Inkling
  37. 58.1 Step 5 Preview
  38. 57.6 GPT-6 Luna
  39. 56.6 Kimi K2.6
  40. 56.6 GPT-6 Sol
  41. 56.6 Gemini 3.6 Flash
  42. 54.3 Claude Haiku 4.5
  43. 53.8 Gemini 3.1 Pro
  44. 51.8 Gemini 3.5 Flash-Lite
  45. 51.7 Grok 4.3
  46. 50.7 Ling-3.0-flash
  47. 45.3 Mercury 2.5
  48. 44.8 Grok 4.1 Fast
  49. 43.8 Laguna M.1
  50. 41.4 Laguna XS.2
  1. 1 Claude Opus 5 76.9%
  2. 2 Claude Fable 5.1 74.9%
  3. 3 Muse Wonder Full 73.8%
  4. 4 Claude Opus 5.5 70.6%
  5. 5 Claude Fable 5 70.4%
  6. 6 Muse Wonder 70.2%
  7. 7 Gemini 4 Argon 69.8%
  8. 8 MiMo-V2.6 Pro 68.9%
  9. 9 Tencent Hy4 preview 68.6%
  10. 10 GLM-5.3 68.5%
  11. 11 Muse Spark 1.2 68.5%
  12. 12 Kimi K3 68.2%
  13. 13 Claude Opus 4.8 68.1%
  14. 14 MiMo-V2.6 Flash 67.6%
  15. 15 Claude Sonnet 5.5 67.2%
  16. 16 Qwen 3.8 Max 67.1%
  17. 17 Grok 4.6 66.8%
  18. 18 GPT-5.6 Sol 66.5%
  19. 19 Ember-1 66.2%
  20. 20 Claude Sonnet 5 66%
  21. 21 Grok 4.7 65.6%
  22. 22 Gemini 3.8 Flash 65.3%
  23. 23 DeepSeek V4.1 Flash 64.3%
  24. 24 MiniMax-M3 64.1%
  25. 25 Claude Haiku 5.5 63.2%
  26. 26 DeepSeek-V4-Pro 62.9%
  27. 27 Claude Sonnet 4.6 62.5%
  28. 28 GPT-5.6 Terra 62.4%
  29. 29 GLM-5.1 61.8%
  30. 30 Muse Spark 1.1 61.3%
  31. 31 GPT-5.6 Luna 61.2%
  32. 32 GPT-5.5 60.9%
  33. 33 Gemini 3.5 Flash 59.5%
  34. 34 Inkling-Small 59.4%
  35. 35 GPT-6.1 Sol 59.3%
  36. 36 Inkling 58.5%
  37. 37 Step 5 Preview 58.1%
  38. 38 GPT-6 Luna 57.6%
  39. 39 Kimi K2.6 56.6%
  40. 40 GPT-6 Sol 56.6%
  41. 41 Gemini 3.6 Flash 56.6%
  42. 42 Claude Haiku 4.5 non-thinking 54.3%
  43. 43 Gemini 3.1 Pro 53.8%
  44. 44 Gemini 3.5 Flash-Lite 51.8%
  45. 45 Grok 4.3 51.7%
  46. 46 Ling-3.0-flash 50.7%
  47. 47 Mercury 2.5 45.3%
  48. 48 Grok 4.1 Fast reasoning 44.8%
  49. 49 Laguna M.1 43.8%
  50. 50 Laguna XS.2 41.4%

Source: Official Public Benefits Bench leaderboard

#ModelScoreSettingDateSource
1 Claude Opus 5 76.9% Jul 2026 Official leaderboard
2 Claude Fable 5.1 74.9% Sep 2026 Official leaderboard
3 Muse Wonder Full 73.8% Oct 2026 Official leaderboard
4 Claude Opus 5.5 70.6% Sep 2026 Official leaderboard
5 Claude Fable 5 70.4% Jun 2026 Official leaderboard
6 Muse Wonder 70.2% Oct 2026 Official leaderboard
7 Gemini 4 Argon 69.8% Sep 2026 Official leaderboard
8 MiMo-V2.6 Pro 68.9% Sep 2026 Official leaderboard
9 Tencent Hy4 preview 68.6% Aug 2026 Official leaderboard
10 GLM-5.3 68.5% Aug 2026 Official leaderboard
11 Muse Spark 1.2 68.5% Aug 2026 Official leaderboard
12 Kimi K3 68.2% Jul 2026 Official leaderboard
13 Claude Opus 4.8 68.1% May 2026 Official leaderboard
14 MiMo-V2.6 Flash 67.6% Sep 2026 Official leaderboard
15 Claude Sonnet 5.5 67.2% Sep 2026 Official leaderboard
16 Qwen 3.8 Max 67.1% Aug 2026 Official leaderboard
17 Grok 4.6 66.8% Aug 2026 Official leaderboard
18 GPT-5.6 Sol 66.5% Jul 2026 Official leaderboard
19 Ember-1 66.2% Sep 2026 Official leaderboard
20 Claude Sonnet 5 66% Jun 2026 Official leaderboard
21 Grok 4.7 65.6% Sep 2026 Official leaderboard
22 Gemini 3.8 Flash 65.3% Sep 2026 Official leaderboard
23 DeepSeek V4.1 Flash 64.3% Sep 2026 Official leaderboard
24 MiniMax-M3 64.1% Jun 2026 Official leaderboard
25 Claude Haiku 5.5 63.2% Oct 2026 Official leaderboard
26 DeepSeek-V4-Pro 62.9% Apr 2026 Official leaderboard
27 Claude Sonnet 4.6 62.5% Feb 2026 Official leaderboard
28 GPT-5.6 Terra 62.4% Jul 2026 Official leaderboard
29 GLM-5.1 61.8% Apr 2026 Official leaderboard
30 Muse Spark 1.1 61.3% Jul 2026 Official leaderboard
31 GPT-5.6 Luna 61.2% Jul 2026 Official leaderboard
32 GPT-5.5 60.9% Apr 2026 Official leaderboard
33 Gemini 3.5 Flash 59.5% May 2026 Official leaderboard
34 Inkling-Small 59.4% Jul 2026 Official leaderboard
35 GPT-6.1 Sol 59.3% Sep 2026 Official leaderboard
36 Inkling 58.5% Jul 2026 Official leaderboard
37 Step 5 Preview 58.1% Sep 2026 Official leaderboard
38 GPT-6 Luna 57.6% Sep 2026 Official leaderboard
39 Kimi K2.6 56.6% Apr 2026 Official leaderboard
40 GPT-6 Sol 56.6% Sep 2026 Official leaderboard
41 Gemini 3.6 Flash 56.6% Jul 2026 Official leaderboard
42 Claude Haiku 4.5 54.3% non-thinking Oct 2025 Official leaderboard
43 Gemini 3.1 Pro 53.8% Feb 2026 Official leaderboard
44 Gemini 3.5 Flash-Lite 51.8% Jul 2026 Official leaderboard
45 Grok 4.3 51.7% Apr 2026 Official leaderboard
46 Ling-3.0-flash 50.7% Aug 2026 Official leaderboard
47 Mercury 2.5 45.3% Sep 2026 Official leaderboard
48 Grok 4.1 Fast 44.8% reasoning Nov 2025 Official leaderboard
49 Laguna M.1 43.8% Apr 2026 Official leaderboard
50 Laguna XS.2 41.4% Apr 2026 Official leaderboard

Accuracy against latency

The Vals scatter chart with its x axis switched from Cost to Latency: mean time per test against accuracy, on a log scale. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

40% 60% 80% 100% 101001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Claude Haiku 4.5 (non-thinking): Accuracy 54.3%, latency (mean per test) 178 s Grok 4.1 Fast (reasoning): Accuracy 44.8%, latency (mean per test) 117 s Claude Sonnet 4.6: Accuracy 62.5%, latency (mean per test) 883 s Gemini 3.1 Pro: Accuracy 53.8%, latency (mean per test) 186 s GLM-5.1: Accuracy 61.8%, latency (mean per test) 580 s Kimi K2.6: Accuracy 56.6%, latency (mean per test) 595 s GPT-5.5: Accuracy 60.9%, latency (mean per test) 2,377 s DeepSeek-V4-Pro: Accuracy 62.9%, latency (mean per test) 922 s Laguna M.1: Accuracy 43.8%, latency (mean per test) 335 s Laguna XS.2: Accuracy 41.4%, latency (mean per test) 305 s Grok 4.3: Accuracy 51.7%, latency (mean per test) 205 s Gemini 3.5 Flash: Accuracy 59.5%, latency (mean per test) 339 s Claude Opus 4.8: Accuracy 68.1%, latency (mean per test) 768 s MiniMax-M3: Accuracy 64.1%, latency (mean per test) 691 s Claude Sonnet 5: Accuracy 66%, latency (mean per test) 1,269 s GPT-5.6 Luna: Accuracy 61.2%, latency (mean per test) 2,320 s GPT-5.6 Sol: Accuracy 66.5%, latency (mean per test) 3,678 s GPT-5.6 Terra: Accuracy 62.4%, latency (mean per test) 1,172 s Muse Spark 1.1: Accuracy 61.3%, latency (mean per test) 395 s Inkling: Accuracy 58.5%, latency (mean per test) 811 s Inkling-Small: Accuracy 59.4%, latency (mean per test) 521 s Qwen 3.8 Max: Accuracy 67.1%, latency (mean per test) 2,444 s Ling-3.0-flash: Accuracy 50.7%, latency (mean per test) 275 s Grok 4.6: Accuracy 66.8%, latency (mean per test) 1,553 s GLM-5.3: Accuracy 68.5%, latency (mean per test) 2,151 s Tencent Hy4 preview: Accuracy 68.6%, latency (mean per test) 3,755 s Claude Fable 5.1: Accuracy 74.9%, latency (mean per test) 2,882 s DeepSeek V4.1 Flash: Accuracy 64.3%, latency (mean per test) 599 s Step 5 Preview: Accuracy 58.1%, latency (mean per test) 1,639 s Grok 4.7: Accuracy 65.6%, latency (mean per test) 1,017 s MiMo-V2.6 Flash: Accuracy 67.6%, latency (mean per test) 1,256 s MiMo-V2.6 Pro: Accuracy 68.9%, latency (mean per test) 2,234 s Claude Opus 5.5: Accuracy 70.6%, latency (mean per test) 4,772 s GPT-6 Luna: Accuracy 57.6%, latency (mean per test) 3,179 s GPT-6 Sol: Accuracy 56.6%, latency (mean per test) 1,863 s Ember-1: Accuracy 66.2%, latency (mean per test) 801 s Claude Sonnet 5.5: Accuracy 67.2%, latency (mean per test) 4,528 s GPT-6.1 Sol: Accuracy 59.3%, latency (mean per test) 3,025 s Gemini 4 Argon: Accuracy 69.8%, latency (mean per test) 1,567 s Claude Haiku 5.5: Accuracy 63.2%, latency (mean per test) 2,193 s Muse Wonder Full: Accuracy 73.8%, latency (mean per test) 4,365 s Claude Fable 5: Accuracy 70.4%, latency (mean per test) 1,277 s Kimi K3: Accuracy 68.2%, latency (mean per test) 439 s Gemini 3.5 Flash-Lite: Accuracy 51.8%, latency (mean per test) 69.3 s Gemini 3.6 Flash: Accuracy 56.6%, latency (mean per test) 122 s Claude Opus 5: Accuracy 76.9%, latency (mean per test) 1,809 s Muse Spark 1.2: Accuracy 68.5%, latency (mean per test) 535 s Gemini 3.8 Flash: Accuracy 65.3%, latency (mean per test) 201 s Mercury 2.5: Accuracy 45.3%, latency (mean per test) 68.3 s Muse Wonder: Accuracy 70.2%, latency (mean per test) 1,185 s Claude Opus 5Claude Fable 5Muse WonderMuse Spark 1.2Kimi K3Gemini 3.8 FlashGemini 3.6 FlashGemini 3.5 Flash-LiteMercury 2.5 40% 60% 80% 100% 101001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Claude Haiku 4.5 (non-thinking): Accuracy 54.3%, latency (mean per test) 178 s Grok 4.1 Fast (reasoning): Accuracy 44.8%, latency (mean per test) 117 s Claude Sonnet 4.6: Accuracy 62.5%, latency (mean per test) 883 s Gemini 3.1 Pro: Accuracy 53.8%, latency (mean per test) 186 s GLM-5.1: Accuracy 61.8%, latency (mean per test) 580 s Kimi K2.6: Accuracy 56.6%, latency (mean per test) 595 s GPT-5.5: Accuracy 60.9%, latency (mean per test) 2,377 s DeepSeek-V4-Pro: Accuracy 62.9%, latency (mean per test) 922 s Laguna M.1: Accuracy 43.8%, latency (mean per test) 335 s Laguna XS.2: Accuracy 41.4%, latency (mean per test) 305 s Grok 4.3: Accuracy 51.7%, latency (mean per test) 205 s Gemini 3.5 Flash: Accuracy 59.5%, latency (mean per test) 339 s Claude Opus 4.8: Accuracy 68.1%, latency (mean per test) 768 s MiniMax-M3: Accuracy 64.1%, latency (mean per test) 691 s Claude Sonnet 5: Accuracy 66%, latency (mean per test) 1,269 s GPT-5.6 Luna: Accuracy 61.2%, latency (mean per test) 2,320 s GPT-5.6 Sol: Accuracy 66.5%, latency (mean per test) 3,678 s GPT-5.6 Terra: Accuracy 62.4%, latency (mean per test) 1,172 s Muse Spark 1.1: Accuracy 61.3%, latency (mean per test) 395 s Inkling: Accuracy 58.5%, latency (mean per test) 811 s Inkling-Small: Accuracy 59.4%, latency (mean per test) 521 s Qwen 3.8 Max: Accuracy 67.1%, latency (mean per test) 2,444 s Ling-3.0-flash: Accuracy 50.7%, latency (mean per test) 275 s Grok 4.6: Accuracy 66.8%, latency (mean per test) 1,553 s GLM-5.3: Accuracy 68.5%, latency (mean per test) 2,151 s Tencent Hy4 preview: Accuracy 68.6%, latency (mean per test) 3,755 s Claude Fable 5.1: Accuracy 74.9%, latency (mean per test) 2,882 s DeepSeek V4.1 Flash: Accuracy 64.3%, latency (mean per test) 599 s Step 5 Preview: Accuracy 58.1%, latency (mean per test) 1,639 s Grok 4.7: Accuracy 65.6%, latency (mean per test) 1,017 s MiMo-V2.6 Flash: Accuracy 67.6%, latency (mean per test) 1,256 s MiMo-V2.6 Pro: Accuracy 68.9%, latency (mean per test) 2,234 s Claude Opus 5.5: Accuracy 70.6%, latency (mean per test) 4,772 s GPT-6 Luna: Accuracy 57.6%, latency (mean per test) 3,179 s GPT-6 Sol: Accuracy 56.6%, latency (mean per test) 1,863 s Ember-1: Accuracy 66.2%, latency (mean per test) 801 s Claude Sonnet 5.5: Accuracy 67.2%, latency (mean per test) 4,528 s GPT-6.1 Sol: Accuracy 59.3%, latency (mean per test) 3,025 s Gemini 4 Argon: Accuracy 69.8%, latency (mean per test) 1,567 s Claude Haiku 5.5: Accuracy 63.2%, latency (mean per test) 2,193 s Muse Wonder Full: Accuracy 73.8%, latency (mean per test) 4,365 s Claude Fable 5: Accuracy 70.4%, latency (mean per test) 1,277 s Kimi K3: Accuracy 68.2%, latency (mean per test) 439 s Gemini 3.5 Flash-Lite: Accuracy 51.8%, latency (mean per test) 69.3 s Gemini 3.6 Flash: Accuracy 56.6%, latency (mean per test) 122 s Claude Opus 5: Accuracy 76.9%, latency (mean per test) 1,809 s Muse Spark 1.2: Accuracy 68.5%, latency (mean per test) 535 s Gemini 3.8 Flash: Accuracy 65.3%, latency (mean per test) 201 s Mercury 2.5: Accuracy 45.3%, latency (mean per test) 68.3 s Muse Wonder: Accuracy 70.2%, latency (mean per test) 1,185 s Claude Opus 5Claude Fable 5Muse WonderMuse Spark 1.2Kimi K3Gemini 3.8 FlashGemini 3.6 FlashGemini 3.5 Flash-LiteMercury 2.5
ModelSettingAccuracyLatency (mean per test)
Claude Opus 5 76.9% 1,809 s
Claude Fable 5.1 74.9% 2,882 s
Muse Wonder Full 73.8% 4,365 s
Claude Opus 5.5 70.6% 4,772 s
Claude Fable 5 70.4% 1,277 s
Muse Wonder 70.2% 1,185 s
Gemini 4 Argon 69.8% 1,567 s
MiMo-V2.6 Pro 68.9% 2,234 s
Tencent Hy4 preview 68.6% 3,755 s
GLM-5.3 68.5% 2,151 s
Muse Spark 1.2 68.5% 535.16 s
Kimi K3 68.2% 438.67 s
Claude Opus 4.8 68.1% 767.85 s
MiMo-V2.6 Flash 67.6% 1,256 s
Claude Sonnet 5.5 67.2% 4,528 s
Qwen 3.8 Max 67.1% 2,444 s
Grok 4.6 66.9% 1,553 s
GPT-5.6 Sol 66.5% 3,678 s
Ember-1 66.2% 801.41 s
Claude Sonnet 5 66% 1,269 s
Grok 4.7 65.6% 1,017 s
Gemini 3.8 Flash 65.3% 201.42 s
DeepSeek V4.1 Flash 64.3% 598.96 s
MiniMax-M3 64.1% 691.44 s
Claude Haiku 5.5 63.2% 2,193 s
DeepSeek-V4-Pro 62.9% 922.39 s
Claude Sonnet 4.6 62.5% 883 s
GPT-5.6 Terra 62.4% 1,172 s
GLM-5.1 61.8% 580.38 s
Muse Spark 1.1 61.3% 395.05 s
GPT-5.6 Luna 61.2% 2,320 s
GPT-5.5 60.9% 2,377 s
Gemini 3.5 Flash 59.5% 338.87 s
Inkling-Small 59.4% 521.34 s
GPT-6.1 Sol 59.3% 3,025 s
Inkling 58.5% 810.83 s
Step 5 Preview 58.1% 1,639 s
GPT-6 Luna 57.7% 3,179 s
Kimi K2.6 56.6% 595.3 s
GPT-6 Sol 56.6% 1,863 s
Gemini 3.6 Flash 56.6% 121.52 s
Claude Haiku 4.5 non-thinking 54.3% 177.66 s
Gemini 3.1 Pro 53.8% 185.55 s
Gemini 3.5 Flash-Lite 51.8% 69.26 s
Grok 4.3 51.7% 205.01 s
Ling-3.0-flash 50.7% 275.25 s
Mercury 2.5 45.3% 68.33 s
Grok 4.1 Fast reasoning 44.8% 116.82 s
Laguna M.1 43.8% 334.83 s
Laguna XS.2 41.4% 305.41 s

The best score, 76.9%, costs $3.95 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Anthropic Xiaomi Corp Meta Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 Cost per task, log scale Claude Opus 5 Muse Wonder MiMo-V2.6 Pro MiMo-V2.6 Flash 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 Cost per task, log scale Claude Opus 5 Muse Wonder

Source: Official Public Benefits Bench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5 76.9% $3.95 54,388
Claude Fable 5.1 74.9% $7.48 49,079
Muse Wonder Full 73.8% $4.64 41,617
Claude Opus 5.5 70.6% $5.29 79,648
Claude Fable 5 70.4% $4.59 42,384
Muse Wonder 70.2% $2.46 31,883
Gemini 4 Argon 69.8% $2.79 29,464
MiMo-V2.6 Pro 68.9% $0.08 44,452
Tencent Hy4 preview 68.6% $0.25 35,390
GLM-5.3 68.5% $0.95 54,089
Muse Spark 1.2 68.5% $0.59 24,957
Kimi K3 68.2% $0.91 26,118
Claude Opus 4.8 68.1% $1.89 40,305
MiMo-V2.6 Flash 67.6% $0.03 43,258
Claude Sonnet 5.5 67.2% $3.76 126,103
Qwen 3.8 Max 67.1% $0.99 36,046
Grok 4.6 66.8% $0.92 13,846
GPT-5.6 Sol 66.5% $8.85 34,790
Ember-1 66.2% $1.02 20,552
Claude Sonnet 5 66% $1.29 55,215
Grok 4.7 65.6% $2.25 30,109
Gemini 3.8 Flash 65.3% $1.01 22,990
DeepSeek V4.1 Flash 64.3% $0.07 34,045
MiniMax-M3 64.1% $0.25 19,754
Claude Haiku 5.5 63.2% $0.53 161,566
DeepSeek-V4-Pro 62.9% $0.46 15,647
Claude Sonnet 4.6 62.5% $1.11 24,904
GPT-5.6 Terra 62.4% $1.20 23,017
GLM-5.1 61.8% $0.34 9,522
Muse Spark 1.1 61.3% $0.28 17,500
GPT-5.6 Luna 61.2% $0.27 37,326
GPT-5.5 60.9% $3.97 25,217
Gemini 3.5 Flash 59.5% $1.11 26,838
Inkling-Small 59.4% $0.10 19,832
GPT-6.1 Sol 59.3% $0.69 33,413
Inkling 58.5% $0.37 19,727
Step 5 Preview 58.1% $0.50 52,211
GPT-6 Luna 57.6% $0.22 77,346
Kimi K2.6 56.6% $0.91 9,804
GPT-6 Sol 56.6% $2.49 30,497
Gemini 3.6 Flash 56.6% $0.39 14,273
Claude Haiku 4.5 non-thinking 54.3% $0.22 11,164
Gemini 3.1 Pro 53.8% $0.38 10,421
Gemini 3.5 Flash-Lite 51.8% $0.10 8,509
Grok 4.3 51.7% $0.26 14,935
Ling-3.0-flash 50.7% $0.05 20,246
Mercury 2.5 45.3% $0.07 10,293
Grok 4.1 Fast reasoning 44.8% $0.03 10,321
Laguna M.1 43.8% 8,832
Laguna XS.2 41.4% 8,411

3 models have held number one on the official leaderboard since its release in June 2026

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Own leaderboard 3 leaders
Claude Opus 5 76.9%

Source: Official Public Benefits Bench leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Opus 5 Claude Fable 5 DeepSeek-V4-Pro Claude Haiku 4.5 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Opus 5 Claude Fable 5

Source: Official Public Benefits Bench leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Law, finance and medicine

All 179 in Law, finance and medicine
  • MedXpertQA-MM In use Models answer expert-level medical exam questions that combine images with clinical records and test results. Scoring is accuracy. Qwen 3.8 Max 80.4%
  • HealthBench Professional In use 525 physician-authored tasks from real clinician chats (care consult, documentation, research) with physician-adjudicated rubrics. Claude Sonnet 5.5 69.2%
  • HealthBench In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
  • FinSearchComp In use Agents search the web to fetch live financial data, look up history, and run multi-step investigations, across 635 expert-written questions. Scored by answer accuracy. Ling-3.1-flash 81.7%
  • KBL In use Models answer Korean legal knowledge and reasoning tasks and Korean bar exam questions. It has 7 knowledge tasks, 4 reasoning tasks and 2,510 bar exam items. Solar Pro 4 77.5%
  • KorMedMCQA In use Models answer 7,469 multiple-choice questions from Korean licensing exams for doctors, nurses, pharmacists and dentists, held from 2012 to 2024. Solar Pro 4 93.2%