Benchmark Almanac

GradeBench

Claude Fable 5.2 57.3% leads since November 2026
Not in model cards

Models read a student's handwritten answer to a university STEM exam question, with the rubric and a reference solution, and pick the rubric items to award. Scored by exact match with the items the instructor chose, averaged over five runs.

464 tasks from 125 questions in 45 courses at more than 10 universities, taken from real exams graded on Pensive. The tasks are not released, and the site is the only source besides an X post. Two human graders agree on 92% of tasks. GradeBench-Diagram is a 258-task subset where students draw.

Made by
Pensive
Released
November 2026
Kind
Other professions
Level
University
Scored
Checked automatically
Answers
Short answer
Tasks
Kept private
Measures
Capability
Language
English
In and out
Text, Images → Text
Interaction
Single turn
Refresh
One fixed set
Run by
A host
Occupations
Educational instruction and library
In views
Vision
Strongest model tested
GPT-6.2 Astra #1 on the capability index
Last model added
November 2026
Scores refreshed
November 18th, 2026
18
Models scored on its own leaderboard

Claude Fable 5.2 leads with 57.3%

Best score per model from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

  1. 1 Claude Fable 5.2 single-shot user prompt; medium reasoning or nearest supported level 57.3%
  2. 2 Gemini 4 Carbon single-shot user prompt; medium reasoning or nearest supported level 56.6%
  3. 3 Gemini 4 Argon single-shot user prompt; medium reasoning or nearest supported level 56.5%
  4. 4 GPT-6.2 Astra single-shot user prompt; medium reasoning or nearest supported level 56.3%
  5. 5 GPT-6.2 Sol single-shot user prompt; medium reasoning or nearest supported level 55.5%
  6. 6 GPT-6 Astra single-shot user prompt; medium reasoning or nearest supported level 53.6%
  7. 7 Claude Opus 5.5 single-shot user prompt; medium reasoning or nearest supported level 53.5%
  8. 8 GPT-6.1 Sol single-shot user prompt; medium reasoning or nearest supported level 53.2%
  9. 9 Claude Fable 5.1 single-shot user prompt; medium reasoning or nearest supported level 53.1%
  10. 10 Gemini 4 Hydrogen single-shot user prompt; medium reasoning or nearest supported level 51.9%
  11. 11 Gemini 3.8 Flash single-shot user prompt; medium reasoning or nearest supported level 51.5%
  12. 12 Muse Wonder single-shot user prompt; medium reasoning or nearest supported level 48.1%
  13. 13 Claude Sonnet 5.5 single-shot user prompt; medium reasoning or nearest supported level 43.7%
  14. 14 Muse Spark 1.3 single-shot user prompt; medium reasoning or nearest supported level 41.4%
  15. 15 GLM-5.3-Flash single-shot user prompt; medium reasoning or nearest supported level 36.6%
  16. 16 GPT-6 Luna single-shot user prompt; medium reasoning or nearest supported level 36.1%
  17. 17 DeepSeek V4.1 Flash single-shot user prompt; medium reasoning or nearest supported level 35.3%
  18. 18 Grok 4.7 single-shot user prompt; medium reasoning or nearest supported level 35.3%

Source: Official GradeBench leaderboard

#ModelScoreSettingDateSource
1 Claude Fable 5.2 57.3% single-shot user prompt; medium reasoning or nearest supported level Nov 2026 Official leaderboard
2 Gemini 4 Carbon 56.6% single-shot user prompt; medium reasoning or nearest supported level Oct 2026 Official leaderboard
3 Gemini 4 Argon 56.5% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
4 GPT-6.2 Astra 56.3% single-shot user prompt; medium reasoning or nearest supported level Nov 2026 Official leaderboard
5 GPT-6.2 Sol 55.5% single-shot user prompt; medium reasoning or nearest supported level Nov 2026 Official leaderboard
6 GPT-6 Astra 53.6% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
7 Claude Opus 5.5 53.5% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
8 GPT-6.1 Sol 53.2% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
9 Claude Fable 5.1 53.1% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
10 Gemini 4 Hydrogen 51.9% single-shot user prompt; medium reasoning or nearest supported level Nov 2026 Official leaderboard
11 Gemini 3.8 Flash 51.5% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
12 Muse Wonder 48.1% single-shot user prompt; medium reasoning or nearest supported level Oct 2026 Official leaderboard
13 Claude Sonnet 5.5 43.7% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
14 Muse Spark 1.3 41.4% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
15 GLM-5.3-Flash 36.6% single-shot user prompt; medium reasoning or nearest supported level Aug 2026 Official leaderboard
16 GPT-6 Luna 36.1% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
17 DeepSeek V4.1 Flash 35.3% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard
18 Grok 4.7 35.3% single-shot user prompt; medium reasoning or nearest supported level Sep 2026 Official leaderboard

Rankings by part

We read 16 parts from the benchmark’s own leaderboard in 3 groups: domain, modality and category. Each ranks models on one part of the test. Claude Fable 5.2 leads 8 of them, and 5 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Engineering

16 models we track have a score here.

  1. 1 Claude Fable 5.2 52.8%
  2. 2 GPT-6 Astra ↑4 49.8%
  3. 3 Gemini 4 Carbon ↓1 49.7%
  4. 4 Gemini 4 Argon ↓1 49.1%
  5. 5 Claude Opus 5.5 ↑2 48.6%
  6. 6 GPT-6.1 Sol ↑2 48.1%
  7. 7 Claude Fable 5.1 ↑2 47.2%
  8. 8 Gemini 4 Hydrogen ↑2 44.7%
  9. 9 Gemini 3.8 Flash ↑2 44.6%
  10. 10 Muse Wonder ↑2 41%

The best score, 57.3%, costs $0.07 a task

Runs on the cost frontier score higher than any run that's cheaper. Gemini 4 Argon at single-shot user prompt; medium reasoning or nearest supported level comes within 0.8 points of the best for $0.009 a task, 7 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Z.ai Google DeepMind Anthropic Cost frontier One model at each effort level
0% 20% 40% 60% $0.001 $0.002 $0.005 $0.01 $0.02 $0.05 $0.10 Cost per task, log scale Claude Fable 5.2, single-shot user prompt; medium reasoning or nearest supported level GLM-5.3-Flash, single-shot user prompt; medium reasoning or nearest supported level 0% 20% 40% 60% $0.001 $0.01 $0.10 Cost per task, log scale Claude Fable 5.2, single-shot user prompt; medium reasoning or nearest supported level GLM-5.3-Flash, single-shot user prompt; medium reasoning or nearest supported level

Source: Official GradeBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Fable 5.2 single-shot user prompt; medium reasoning or nearest supported level 57.3% $0.07 252
Gemini 4 Carbon single-shot user prompt; medium reasoning or nearest supported level 56.6% $0.04 3,106
Gemini 4 Argon single-shot user prompt; medium reasoning or nearest supported level 56.5% $0.009 291
GPT-6.2 Astra single-shot user prompt; medium reasoning or nearest supported level 56.3% $0.04 34
GPT-6.2 Sol single-shot user prompt; medium reasoning or nearest supported level 55.5% $0.01 57
GPT-6 Astra single-shot user prompt; medium reasoning or nearest supported level 53.6% $0.06 96
Claude Opus 5.5 single-shot user prompt; medium reasoning or nearest supported level 53.5% $0.04 423
GPT-6.1 Sol single-shot user prompt; medium reasoning or nearest supported level 53.2% $0.01 113
Claude Fable 5.1 single-shot user prompt; medium reasoning or nearest supported level 53.1% $0.10 603
Gemini 4 Hydrogen single-shot user prompt; medium reasoning or nearest supported level 51.9% $0.02 1,793
Gemini 3.8 Flash single-shot user prompt; medium reasoning or nearest supported level 51.5% $0.01 1,966
Muse Wonder single-shot user prompt; medium reasoning or nearest supported level 48.1% $0.06 1,776
Claude Sonnet 5.5 single-shot user prompt; medium reasoning or nearest supported level 43.7% $0.02 392
Muse Spark 1.3 single-shot user prompt; medium reasoning or nearest supported level 41.4% $0.02 3,189
GLM-5.3-Flash single-shot user prompt; medium reasoning or nearest supported level 36.6% $0.001 2,190
GPT-6 Luna single-shot user prompt; medium reasoning or nearest supported level 36.1% $0.001 1,408
DeepSeek V4.1 Flash single-shot user prompt; medium reasoning or nearest supported level 35.3% $0.01 12,885
Grok 4.7 single-shot user prompt; medium reasoning or nearest supported level 35.3% $0.09 12,786

2 models have held number one on the official leaderboard since its release in November 2026

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

Google DeepMind Anthropic
Own leaderboard 2 leaders

Source: Official GradeBench leaderboard

Scores over time

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

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2027 Near the ceiling (85%) Claude Fable 5.2 Gemini 4 Argon GLM-5.3-Flash 0% 20% 40% 60% 80% 100% 2027 Near the ceiling (85%) Claude Fable 5.2 GLM-5.3-Flash

Source: Official GradeBench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Law, finance and medicine

All 637 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%
  • 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. o1-preview 92.7%
  • Harvey LAB-AA v1.1 In use Artificial Analysis's run of Harvey's Legal Agent Benchmark on Harvey's newer private set of 120 tasks across 24 practice areas. A deliverable with a material hallucination scores zero, and the headline is the share of judges who find every rubric criterion passed. GPT-6.2 Astra 15%

Data audit

All 11 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access The tasks are private.
  • Pass Confidence Checked against a primary source.
Scores 5 pass
  • Pass Independent runs Scores from the official leaderboard (18 entries).
  • Pass Models scored 18 models we track.
  • Pass Newest score The newest is from November 2026, for GPT-6.2 Astra.
  • Pass Frontier models 9 of the top 10 have a score, led by GPT-6.2 Astra (#1).
  • Pass Headroom The best is 57.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on gradebench.com: 18 rows for models we track at the last read.
  • Pass Last read November 18th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check