Benchmark Almanac

Verifying the Verifiers (Project APE CRED benchmark)

GPT-6 Astra 99.5% leads since September 2026
Not in model cards

Verifier LLMs must find planted errors (prose-table, paper-code, execution, data) in AI-generated economics papers. Credit needs a verbatim quote of the edit.

Mutation testing on 100 reproducible host papers with deterministic scoring and no LLM judge. Best single config 99% recall (Codex 5.6 Sol high). Instances and answer key withheld against contamination. Up to 28% of per-error outcomes flip between identical runs. Repo holds the APE papers, not the CRED code.

Released
July 2026
Kind
Research science
Level
Expert
Scored
Checked automatically
Answers
Short answer
Tasks
Kept private
Strongest model tested
GPT-6 Astra #3 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026
24
Entries on its own leaderboard, 23 models we track

GPT-6 Astra leads with 99.5% among models released by September 2026

Best score per entry 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. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

24 entries. Scroll sideways for the rest.

  1. 99.5 GPT-6 Astra
  2. 99.0 GPT-5.6 Sol
  3. 97.5 GPT-5.6 Luna
  4. 97.0 GPT-5.6 Terra
  5. 96.5 GPT-5.5
  6. 96.5 Claude Fable 5
  7. 91.0 openai/gpt-5.6-luna-pro
  8. 90.0 GLM-5.3-Flash
  9. 89.5 GLM-5.2
  10. 87.5 GPT-6 Luna
  11. 85.0 Claude Opus 4.8
  12. 83.5 Claude Sonnet 4.6
  13. 83.5 DeepSeek-V4-Pro
  14. 80.5 DeepSeek-V4-Flash
  15. 79.5 Gemini 3.8 Flash
  16. 77.5 Claude Sonnet 5
  17. 74.0 Gemini 3.7 Flash
  18. 68.5 Gemini 3.5 Flash
  19. 50.5 Tencent Hy3 preview
  20. 47.0 Gemini 3 Flash
  21. 32.0 Mistral Large 3
  22. 19.0 Mistral Small 3.2
  23. 14.0 GPT-4.1
  24. 11.5 GPT-4o
  1. 1 GPT-6 Astra low 99.5%
  2. 2 GPT-5.6 Sol high 99%
  3. 3 GPT-5.6 Luna xhigh 97.5%
  4. 4 GPT-5.6 Terra high 97%
  5. 5 GPT-5.5 high 96.5%
  6. 6 Claude Fable 5 high 96.5%
  7. 7 openai/gpt-5.6-luna-pro 91%
  8. 8 GLM-5.3-Flash xhigh 90%
  9. 9 GLM-5.2 high 89.5%
  10. 10 GPT-6 Luna high 87.5%
  11. 11 Claude Opus 4.8 medium 85%
  12. 12 Claude Sonnet 4.6 high 83.5%
  13. 13 DeepSeek-V4-Pro xhigh 83.5%
  14. 14 DeepSeek-V4-Flash xhigh 80.5%
  15. 15 Gemini 3.8 Flash high 79.5%
  16. 16 Claude Sonnet 5 high 77.5%
  17. 17 Gemini 3.7 Flash high 74%
  18. 18 Gemini 3.5 Flash high 68.5%
  19. 19 Tencent Hy3 preview 50.5%
  20. 20 Gemini 3 Flash low 47%
  21. 21 Mistral Large 3 32%
  22. 22 Mistral Small 3.2 19%
  23. 23 GPT-4.1 14%
  24. 24 GPT-4o Mar 2025 11.5%

Source: Official Verifying the Verifiers (Project APE CRED benchmark) leaderboard

#EntryScoreSettingDateSource
1 GPT-6 Astra 99.5% low, Codex CLI Sep 2026 Official leaderboard
2 GPT-5.6 Sol 99% high, Codex CLI Jul 2026 Official leaderboard
3 GPT-5.6 Luna 97.5% xhigh Jul 2026 Official leaderboard
4 GPT-5.6 Terra 97% high, Codex CLI Jul 2026 Official leaderboard
5 GPT-5.5 96.5% high, Codex CLI Apr 2026 Official leaderboard
6 Claude Fable 5 96.5% high, Claude Code Jun 2026 Official leaderboard
7 openai/gpt-5.6-luna-pro 91% Aug 2026 Official leaderboard
8 GLM-5.3-Flash 90% xhigh Aug 2026 Official leaderboard
9 GLM-5.2 89.5% high Jun 2026 Official leaderboard
10 GPT-6 Luna 87.5% high, Codex CLI Sep 2026 Official leaderboard
11 Claude Opus 4.8 85% medium, Claude Code May 2026 Official leaderboard
12 Claude Sonnet 4.6 83.5% high, Claude Code Feb 2026 Official leaderboard
13 DeepSeek-V4-Pro 83.5% xhigh Apr 2026 Official leaderboard
14 DeepSeek-V4-Flash 80.5% xhigh Apr 2026 Official leaderboard
15 Gemini 3.8 Flash 79.5% high Sep 2026 Official leaderboard
16 Claude Sonnet 5 77.5% high, Claude Code Jun 2026 Official leaderboard
17 Gemini 3.7 Flash 74% high Aug 2026 Official leaderboard
18 Gemini 3.5 Flash 68.5% high May 2026 Official leaderboard
19 Tencent Hy3 preview 50.5% Apr 2026 Official leaderboard
20 Gemini 3 Flash 47% low Dec 2025 Official leaderboard
21 Mistral Large 3 32% Dec 2025 Official leaderboard
22 Mistral Small 3.2 19% Jun 2025 Official leaderboard
23 GPT-4.1 14% Apr 2025 Official leaderboard
24 GPT-4o Mar 2025 11.5% Mar 2025 Official leaderboard

The best score, 99.5%, costs $0.78 a task

Runs on the cost frontier score higher than any run that's cheaper. GPT-5.6 Luna at xhigh comes within 2 points of the best for $0.05 a task, 16 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Mistral AI OpenAI Z.ai Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra, low GPT-5.6 Luna, xhigh GLM-5.3-Flash, xhigh GLM-5.3-Flash, low Mistral Small 3.2 0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra, low GPT-5.6 Luna, high

Source: Official Verifying the Verifiers (Project APE CRED benchmark) leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra low Codex CLI 99.5% $0.78
GPT-6 Astra medium Codex CLI 99.5% $0.82
GPT-5.6 Sol high Codex CLI 99% $0.86
GPT-5.6 Sol medium Codex CLI 98.5% $0.62
GPT-5.6 Luna xhigh 97.5% $0.05
GPT-5.6 Terra high Codex CLI 97% $0.37
GPT-5.5 high Codex CLI 96.5% $0.82
Claude Fable 5 high Claude Code 96.5% $2.04
Claude Fable 5 medium Claude Code 96% $1.66
GPT-5.6 Luna high Codex CLI 95% $0.23
GPT-5.6 Sol low Codex CLI 94.5% $0.44
GPT-5.6 Luna high 94% $0.02
GPT-5.5 medium Codex CLI 93.5% $0.58
Claude Fable 5 low Claude Code 92.5% $1.16
GLM-5.3-Flash xhigh 90% $0.003
GLM-5.2 high 89.5% $0.12
GPT-5.5 low Codex CLI 88.5% $0.44
GLM-5.3-Flash high 88.5% $0.003
GPT-6 Luna high Codex CLI 87.5% $0.009
GPT-5.6 Luna low 86.5% $0.008
GPT-5.6 Luna medium 86.5% $0.01
Claude Opus 4.8 medium Claude Code 85% $0.78
Claude Opus 4.8 high Claude Code 85% $0.93
Claude Opus 4.8 low Claude Code 84% $0.92
Claude Sonnet 4.6 high Claude Code 83.5% $0.72
DeepSeek-V4-Pro xhigh 83.5% $0.03
DeepSeek-V4-Flash xhigh 80.5% $0.007
Gemini 3.8 Flash high 79.5% $0.14
Claude Sonnet 4.6 medium Claude Code 78% $0.50
Claude Sonnet 5 high Claude Code 77.5% $0.49
GLM-5.3-Flash low 75% $0.002
Gemini 3.7 Flash high 74% $0.03
GPT-6 Luna low Codex CLI 73% $0.006
Claude Sonnet 4.6 low Claude Code 72% $0.27
Gemini 3.5 Flash high 68.5% $0.21
Gemini 3.7 Flash low 64.5% $0.01
Gemini 3.8 Flash low 64.5% $0.03
Gemini 3.5 Flash low 51.5% $0.07
Tencent Hy3 preview 50.5% $0.003
Gemini 3 Flash low 47% $0.02
Mistral Large 3 32% $0.02
Mistral Small 3.2 19% $0.002
GPT-4.1 14% $0.05
GPT-4o Mar 2025 11.5% $0.07

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

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

OpenAI
Own leaderboard Passed 85% in Jul 2026, by GPT-5.5
GPT-6 Astra 99.5%

Source: Official Verifying the Verifiers (Project APE CRED benchmark) leaderboard

GPT-5.5 on the official leaderboard passed 85% in April 2026

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% 2025 2026 Near the ceiling (85%) GPT-6 Astra GPT-5.5 Claude Sonnet 4.6 Gemini 3 Flash Mistral Large 3 Mistral Small 3.2 GPT-4.1 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-6 Astra Claude Sonnet 4.6

Source: Official Verifying the Verifiers (Project APE CRED benchmark) 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 Science and reasoning

All 92 in Science and reasoning
  • GPQA Diamond In use 198 graduate-level 'Google-proof' multiple-choice questions in biology, chemistry and physics. Gemini 4 Carbon 96.5%
  • Terminal-Bench-Science In use 70 scientific research-workflow tasks across five science areas, run in Harbor containers. Hidden tests give the resolution rate. GPT-6 Astra 68.1%
  • FrontierScience In use Models solve physics, chemistry, and biology problems in an Olympiad track graded by short answer and a Research track graded by 10-point rubrics. Intern-S2-397B 87%
  • BioMysteryBench In use Bioinformatics problems on anonymized real datasets that require real analysis with tools. Rubric-graded, with human-solvable and human-difficult splits. Claude Fable 5.2 83.7%
  • HiPhO In use Models solve theory problems from 2024 and 2025 physics olympiad exams, some with diagrams. Official marking schemes grade answers and steps, mapped to medal levels. Gemini 3 Pro 78.3%
  • LABBench2 In use 11 biology research sub-tasks (literature, figures, tables, patents, trials and more) with tool access. Macro-averaged accuracy. GPT-5.2 60.9%