Benchmark Almanac

CounterBench (1,200 questions)

DeepSeek-V3 91.8% leads since August 2025
Not in model cards
Version 1,200 questions

Models answer counterfactual questions using the stated causal rules. This release adds 200 backdoor questions to the original 1,000.

The CaDO team used both question files and reserved 230 questions for testing. Its split groups questions by their underlying causal mechanism.

Released
August 2025
Kind
Language reasoning
Level
University
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Subject
Mathematics
Language
English
In and out
Text → Text
Interaction
Single turn
Items from
Another benchmark
Run by
Anyone
Based on
CounterBench
Strongest model tested
DeepSeek-V3 #155 on the capability index
Last model added
December 2024
Scores refreshed
November 20th, 2026
10
Entries in its paper, 6 models we track

DeepSeek-V3 leads with 91.8% among models released by December 2024 DeepSeek-V3 leads with 91.8% among models released by December 2024

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. 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. This view shows only models we track, each at its best.

The newest model with a score here came out in December 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 DeepSeek-V3 91.8%
  2. 2 Gemini-1.5-flash (snapshot unspecified) 89.9%
  3. 3 Claude 3 Sonnet 89.8%
  4. 4 GPT-4o (snapshot unspecified) 89.4%
  5. 5 Gemini 1.5 Flash 8B 83.9%
  6. 6 GPT-4o mini 79.9%
  7. 7 Claude 3.5 Haiku 79.1%
  8. 8 GPT-3.5 turbo (snapshot unspecified) 58.9%
  9. 9 davinci-002 Standard 51.9%
  10. 10 GPT-3 (Babbage-002) Standard 49.6%

Source: The CounterBench (1,200 questions) paper

#EntryScoreSettingDateSource
1 DeepSeek-V3 91.8% CoIn Dec 2024 Official leaderboard
2 Gemini-1.5-flash (snapshot unspecified) 89.9% CoIn Official leaderboard
3 Claude 3 Sonnet 89.8% CoIn Feb 2024 Official leaderboard
4 GPT-4o (snapshot unspecified) 89.4% CoIn Official leaderboard
5 Gemini 1.5 Flash 8B 83.9% CoIn Oct 2024 Official leaderboard
6 GPT-4o mini 79.9% CoIn Jul 2024 Official leaderboard
7 Claude 3.5 Haiku 79.1% CoIn Oct 2024 Official leaderboard
8 GPT-3.5 turbo (snapshot unspecified) 58.9% CoIn Official leaderboard
9 davinci-002 51.9% Standard Mar 2022 Official leaderboard
10 GPT-3 (Babbage-002) 49.6% Standard Official leaderboard
#ModelScoreSettingDateSource
1 DeepSeek-V3 91.8% CoIn Dec 2024 Official leaderboard
2 Claude 3 Sonnet 89.8% CoIn Feb 2024 Official leaderboard
3 Gemini 1.5 Flash 8B 83.9% CoIn Oct 2024 Official leaderboard
4 GPT-4o mini 79.9% CoIn Jul 2024 Official leaderboard
5 Claude 3.5 Haiku 79.1% CoIn Oct 2024 Official leaderboard
6 davinci-002 51.9% Standard Mar 2022 Official leaderboard

Rankings by part

We read 5 parts from the benchmark’s paper, each ranking models on one part of the test, such as Basic, Conditional and Joint. DeepSeek-V3 leads 4 of them, and Claude 3 Sonnet leads the other one. Arrows show how far a model moved from its overall rank among the models we track.

Basic

6 models we track have a score here.

  1. 1 DeepSeek-V3 99.6%
  2. 2 Claude 3 Sonnet 97.6%
  3. 3 Gemini 1.5 Flash 8B 94.8%
  4. 4 Claude 3.5 Haiku ↑1 88%
  5. 5 GPT-4o mini ↓1 82.8%
  6. 6 davinci-002 56.8%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
DeepSeek-V3
Standard CausalCoT
74.2%
Gemini 1.5 Flash 8B
Standard CausalCoT
64.8%
Claude 3.5 Haiku
Standard CausalCoT
64.2%
GPT-4o mini
Standard CausalCoT
59.8%
Claude 3 Sonnet
Standard CausalCoT
59%
davinci-002
CausalCoT
51.9%

Source: The CounterBench (1,200 questions) paper

ModelEffortHarnessScoreCost per taskOutput tokens
DeepSeek-V3 CoIn 91.8%
Claude 3 Sonnet CoIn 89.8%
Gemini 1.5 Flash 8B CoIn 83.9%
GPT-4o mini CoIn 79.9%
Claude 3.5 Haiku CoIn 79.1%
DeepSeek-V3 CausalCoT 74.2%
Gemini 1.5 Flash 8B CausalCoT 64.8%
Claude 3.5 Haiku CausalCoT 64.2%
GPT-4o mini CausalCoT 59.8%
Claude 3 Sonnet CausalCoT 59%
davinci-002 Standard 51.9%
DeepSeek-V3 Standard 51.9%
Claude 3 Sonnet Solver 51.8%
Claude 3 Sonnet Standard 51.6%
GPT-4o mini Standard 50.4%
Gemini 1.5 Flash 8B Standard 50.4%
Gemini 1.5 Flash 8B Solver 50.3%
davinci-002 Solver 50.1%
davinci-002 CoIn 49.6%
davinci-002 CausalCoT 49.3%
DeepSeek-V3 Solver 49.3%
Claude 3.5 Haiku Solver 48.3%
GPT-4o mini Solver 47.2%
Claude 3.5 Haiku Standard 39.8%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
DeepSeek-V3
Standard CausalCoT
74.2%
Gemini 1.5 Flash 8B
Standard CausalCoT
64.8%
Claude 3.5 Haiku
Standard CausalCoT
64.2%
GPT-4o mini
Standard CausalCoT
59.8%
Claude 3 Sonnet
Standard CausalCoT
59%
davinci-002
CausalCoT
51.9%

Source: The CounterBench (1,200 questions) paper

ModelEffortHarnessScoreCost per taskOutput tokens
DeepSeek-V3 CoIn 91.8%
Claude 3 Sonnet CoIn 89.8%
Gemini 1.5 Flash 8B CoIn 83.9%
GPT-4o mini CoIn 79.9%
Claude 3.5 Haiku CoIn 79.1%
DeepSeek-V3 CausalCoT 74.2%
Gemini 1.5 Flash 8B CausalCoT 64.8%
Claude 3.5 Haiku CausalCoT 64.2%
GPT-4o mini CausalCoT 59.8%
Claude 3 Sonnet CausalCoT 59%
davinci-002 Standard 51.9%
DeepSeek-V3 Standard 51.9%
Claude 3 Sonnet Solver 51.8%
Claude 3 Sonnet Standard 51.6%
GPT-4o mini Standard 50.4%
Gemini 1.5 Flash 8B Standard 50.4%
Gemini 1.5 Flash 8B Solver 50.3%
davinci-002 Solver 50.1%
davinci-002 CoIn 49.6%
davinci-002 CausalCoT 49.3%
DeepSeek-V3 Solver 49.3%
Claude 3.5 Haiku Solver 48.3%
GPT-4o mini Solver 47.2%
Claude 3.5 Haiku Standard 39.8%

Number one over time

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

DeepSeek
Benchmark paper Passed 85% in Aug 2025, by DeepSeek-V3
DeepSeek-V3 91.8%

Source: The CounterBench (1,200 questions) paper

Claude 3 Sonnet on the official leaderboard passed 85% in February 2024

Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2022 2023 2024 2025 Near the ceiling (85%) DeepSeek-V3 Claude 3 Sonnet davinci-002 0% 20% 40% 60% 80% 100% 2022 2023 2024 2025 Near the ceiling (85%) DeepSeek-V3 davinci-002

Source: Official CounterBench (1,200 questions) 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.

Versions

We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.

DeepSeek
CounterBench (1,200 questions) Passed 85% in Aug 2025, by DeepSeek-V3
DeepSeek-V3 91.8%
CounterBench Passed 85% in Feb 2025, by DeepSeek-V3
DeepSeek-V3 93.5%
CounterBench (1,200 questions) Passed 85% in Aug 2025, by DeepSeek-V3
DeepSeek-V3 91.8%

Source: Official leaderboards

More in Knowledge

All 332 in Knowledge
  • MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects. Intern-S2-397B 89.8%
  • MMLU In use 57-subject multiple-choice exam questions from elementary to professional level. Scored by accuracy. Seed1.8 92.3%
  • SuperGPQA In use 26,529 graduate-level multiple-choice questions across 285 disciplines. Accuracy. Qwen 3.6 Max (Preview) 73.9%
  • WideSearch In use 200 bilingual broad info-seeking tasks: agent collects many atomic facts into a table. Scored by success rate and row/item F1. Tencent Hy4 preview 83.9%
  • BBEH micro average In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is the share of all examples answered correctly. Gemma 4 31B IT 74.4%
  • HLE-Diamond In use Models answer 1,000 cleaned expert questions from Humanity's Last Exam, 500 on reasoning and 500 on knowledge, closed book. Graded on exact answers. Claude Fable 5.2 77.9%

Data audit

All 8 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
  • Pass Independent runs Scores from the benchmark’s paper (10 entries).
  • Pass Models scored 6 models we track, plus 4 other entries on its leaderboard.
  • Note Newest score The newest came 23 months ago, in December 2024, for DeepSeek-V3. That’s expected for a saturated test.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is DeepSeek-V3 (#155).
  • Note Headroom Saturated. The best is 91.8%, past 85% since February 2024.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 24 rows for models we track and 4 other entries at the last read.
  • Pass Last read November 20th, 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