Benchmark Almanac

3SAT-Search

o3-mini 96.2% leads since March 2025
o3-mini 96.2% leads since March 2025
Only in system cards

Models find satisfying assignments for 960 3-SAT formulas across difficulty levels. A checker verifies each assignment.

Microsoft released the dataset on Hugging Face as microsoft/sat in June 2025. The early paper listed 800 formulas, and current author logs use 960.

Released
March 2025
Kind
Abstract puzzles and logic
Level
University
Scored
Checked automatically
Answers
Short answer
Strongest model tested
o1 #117 on the capability index
Last in a system card
April 2025
Last model added
April 2025
Scores refreshed
November 12th, 2026
14
Entries on its own leaderboard, 11 models we track
2
System cards report it, from 1 lab

o3-mini leads with 96.2% among models released by February 2025

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

o3-mini leads with 96.2%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.

The newest model with a score here came out in February 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

14 entries. Scroll sideways for the rest.

  1. 96.2 o3-mini
  2. 89.0 o1
  3. 80.7 DeepSeek-R1
  4. 78.0 PHI_4_REASONING_phi-16b (archive research variant)
  5. 76.7 Phi-4_reasoning (archive research variant)
  6. 72.2 DeepSeek-R1-Distill-Llama-70B
  7. 69.6 PHI_4_REASONING_phi-15b-rl (archive research variant)
  8. 62.6 Gemini 2.0 Flash Thinking
  9. 61.8 Claude 3.7 Sonnet
  10. 25.1 Gemini 2.0 Pro
  11. 17.1 Phi-4
  12. 12.9 Claude 3.5 Sonnet
  13. 12.7 GPT-4o
  14. 8.25 Llama 3.1-405B
  1. 1 o3-mini reasoning high 96.2%
  2. 2 o1 89%
  3. 3 DeepSeek-R1 max tokens 65536 80.7%
  4. 4 PHI_4_REASONING_phi-16b (archive research variant) 78%
  5. 5 Phi-4_reasoning (archive research variant) 76.7%
  6. 6 DeepSeek-R1-Distill-Llama-70B max tokens 65536 72.2%
  7. 7 PHI_4_REASONING_phi-15b-rl (archive research variant) 69.6%
  8. 8 Gemini 2.0 Flash Thinking Jan 2025, max tokens 32768 62.6%
  9. 9 Claude 3.7 Sonnet thinking budget 60000; max tokens 64000 61.8%
  10. 10 Gemini 2.0 Pro max tokens 4096 25.1%
  11. 11 Phi-4 max tokens 4096 17.1%
  12. 12 Claude 3.5 Sonnet max tokens 4096 12.9%
  13. 13 GPT-4o Aug 2024, max tokens 4096 12.7%
  14. 14 Llama 3.1-405B max tokens 4096 8.25%

Source: Official 3SAT-Search leaderboard

13 models. Scroll sideways for the rest.

  1. 96.2 o3-mini
  2. 89.0 o1
  3. 80.7 DeepSeek-R1
  4. 78.0 Phi-4-reasoning
  5. 72.2 DeepSeek-R1-Distill-Llama-70B
  6. 70.9 Phi-4-Reasoning-plus
  7. 62.6 Gemini 2.0 Flash Thinking
  8. 61.8 Claude 3.7 Sonnet
  9. 25.1 Gemini 2.0 Pro
  10. 17.1 Phi-4
  11. 12.9 Claude 3.5 Sonnet
  12. 12.7 GPT-4o
  13. 8.25 Llama 3.1-405B
  1. 1 o3-mini reasoning high 96.2%
  2. 2 o1 89%
  3. 3 DeepSeek-R1 max tokens 65536 80.7%
  4. 4 Phi-4-reasoning 78%
  5. 5 DeepSeek-R1-Distill-Llama-70B max tokens 65536 72.2%
  6. 6 Phi-4-Reasoning-plus 70.9%
  7. 7 Gemini 2.0 Flash Thinking Jan 2025, max tokens 32768 62.6%
  8. 8 Claude 3.7 Sonnet thinking budget 60000; max tokens 64000 61.8%
  9. 9 Gemini 2.0 Pro max tokens 4096 25.1%
  10. 10 Phi-4 max tokens 4096 17.1%
  11. 11 Claude 3.5 Sonnet max tokens 4096 12.9%
  12. 12 GPT-4o Aug 2024, max tokens 4096 12.7%
  13. 13 Llama 3.1-405B max tokens 4096 8.25%

Sources: Official 3SAT-Search leaderboard and labs’ model cards

#EntryScoreSettingDateSource
1 o3-mini 96.2% reasoning high Jan 2025 Official leaderboard
2 o1 89% Dec 2024 Official leaderboard
3 DeepSeek-R1 80.7% max tokens 65536 Jan 2025 Official leaderboard
4 PHI_4_REASONING_phi-16b (archive research variant) 78% Official leaderboard
5 Phi-4_reasoning (archive research variant) 76.7% Official leaderboard
6 DeepSeek-R1-Distill-Llama-70B 72.2% max tokens 65536 Jan 2025 Official leaderboard
7 PHI_4_REASONING_phi-15b-rl (archive research variant) 69.6% Official leaderboard
8 Gemini 2.0 Flash Thinking Jan 2025 62.6% max tokens 32768 Jan 2025 Official leaderboard
9 Claude 3.7 Sonnet 61.8% thinking budget 60000; max tokens 64000 Feb 2025 Official leaderboard
10 Gemini 2.0 Pro 25.1% max tokens 4096 Feb 2025 Official leaderboard
11 Phi-4 17.1% max tokens 4096 Dec 2024 Official leaderboard
12 Claude 3.5 Sonnet 12.9% max tokens 4096 Jun 2024 Official leaderboard
13 GPT-4o Aug 2024 12.7% max tokens 4096 Aug 2024 Official leaderboard
14 Llama 3.1-405B 8.25% max tokens 4096 Jul 2024 Official leaderboard
#ModelScoreSettingDateSource
1 o3-mini 96.2% reasoning high Jan 2025 Official leaderboard
2 o1 89% Dec 2024 Official leaderboard
3 DeepSeek-R1 80.7% max tokens 65536 Jan 2025 Official leaderboard
4 Phi-4-reasoning 78% Apr 2025 Lab model card
5 DeepSeek-R1-Distill-Llama-70B 72.2% max tokens 65536 Jan 2025 Official leaderboard
6 Phi-4-Reasoning-plus 70.9% Apr 2025 Lab model card
7 Gemini 2.0 Flash Thinking Jan 2025 62.6% max tokens 32768 Jan 2025 Official leaderboard
8 Claude 3.7 Sonnet 61.8% thinking budget 60000; max tokens 64000 Feb 2025 Official leaderboard
9 Gemini 2.0 Pro 25.1% max tokens 4096 Feb 2025 Official leaderboard
10 Phi-4 17.1% max tokens 4096 Dec 2024 Official leaderboard
11 Claude 3.5 Sonnet 12.9% max tokens 4096 Jun 2024 Official leaderboard
12 GPT-4o Aug 2024 12.7% max tokens 4096 Aug 2024 Official leaderboard
13 Llama 3.1-405B 8.25% max tokens 4096 Jul 2024 Official leaderboard

Labs report scores for 2 models in their cards. Switch the scores above to rank them together.

Accuracy versus token usage

The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0% 20% 40% 60% 80% 100% 1001,00010K100K Output tokens per task, log scale, lower is better Accuracy Claude 3.5 Sonnet (max tokens 4096): Accuracy 12.9%, output tokens per task 416 Llama 3.1-405B (max tokens 4096): Accuracy 8.25%, output tokens per task 1,219 GPT-4o (max tokens 4096): Accuracy 12.7%, output tokens per task 958 Phi-4 (max tokens 4096): Accuracy 17.1%, output tokens per task 1,521 DeepSeek-R1 (max tokens 65536): Accuracy 80.7%, output tokens per task 10K DeepSeek-R1-Distill-Llama-70B (max tokens 65536): Accuracy 72.2%, output tokens per task 9,716 Gemini 2.0 Pro (max tokens 4096): Accuracy 25.1%, output tokens per task 1,575 Claude 3.7 Sonnet (thinking budget 60000; max tokens 64000): Accuracy 61.8%, output tokens per task 21K PHI_4_REASONING_phi-16b (archive research variant): Accuracy 78%, output tokens per task 16K. Not in our model list Phi-4_reasoning (archive research variant): Accuracy 76.7%, output tokens per task 16K. Not in our model list PHI_4_REASONING_phi-15b-rl (archive research variant): Accuracy 69.6%, output tokens per task 18K. Not in our model list o1: Accuracy 89%, output tokens per task 8,302 Gemini 2.0 Flash Thinking (max tokens 32768): Accuracy 62.6%, output tokens per task 201 o3-mini (reasoning high): Accuracy 96.2%, output tokens per task 20K o3-mini (reasoning high)o1Gemini 2.0 Flash Thinking (max tokens 32768) 0% 20% 40% 60% 80% 100% 1001,00010K100K Output tokens per task, log scale, lower is better Accuracy Claude 3.5 Sonnet (max tokens 4096): Accuracy 12.9%, output tokens per task 416 Llama 3.1-405B (max tokens 4096): Accuracy 8.25%, output tokens per task 1,219 GPT-4o (max tokens 4096): Accuracy 12.7%, output tokens per task 958 Phi-4 (max tokens 4096): Accuracy 17.1%, output tokens per task 1,521 DeepSeek-R1 (max tokens 65536): Accuracy 80.7%, output tokens per task 10K DeepSeek-R1-Distill-Llama-70B (max tokens 65536): Accuracy 72.2%, output tokens per task 9,716 Gemini 2.0 Pro (max tokens 4096): Accuracy 25.1%, output tokens per task 1,575 Claude 3.7 Sonnet (thinking budget 60000; max tokens 64000): Accuracy 61.8%, output tokens per task 21K PHI_4_REASONING_phi-16b (archive research variant): Accuracy 78%, output tokens per task 16K. Not in our model list Phi-4_reasoning (archive research variant): Accuracy 76.7%, output tokens per task 16K. Not in our model list PHI_4_REASONING_phi-15b-rl (archive research variant): Accuracy 69.6%, output tokens per task 18K. Not in our model list o1: Accuracy 89%, output tokens per task 8,302 Gemini 2.0 Flash Thinking (max tokens 32768): Accuracy 62.6%, output tokens per task 201 o3-mini (reasoning high): Accuracy 96.2%, output tokens per task 20K o3-mini (reasoning high)o1Gemini 2.0 Flash Thinking (max tokens 32768)
ModelSettingAccuracyOutput tokens per task
o3-mini reasoning high 96.2% 20,471
o1 89% 8,302
DeepSeek-R1 max tokens 65536 80.7% 10,345
PHI_4_REASONING_phi-16b (archive research variant) 78% 16,313
Phi-4_reasoning (archive research variant) 76.7% 16,066
DeepSeek-R1-Distill-Llama-70B max tokens 65536 72.2% 9,716
PHI_4_REASONING_phi-15b-rl (archive research variant) 69.6% 18,248
Gemini 2.0 Flash Thinking Jan 2025 max tokens 32768 62.6% 201.45
Claude 3.7 Sonnet thinking budget 60000; max tokens 64000 61.8% 20,730
Gemini 2.0 Pro max tokens 4096 25.1% 1,575
Phi-4 max tokens 4096 17.2% 1,521
Claude 3.5 Sonnet max tokens 4096 12.9% 416.31
GPT-4o Aug 2024 max tokens 4096 12.7% 958.38
Llama 3.1-405B max tokens 4096 8.3% 1,219

Number one over time Number one over time

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

OpenAI
Own leaderboard Passed 85% in Mar 2025, by o3-mini
o3-mini 96.2%

Source: Official 3SAT-Search leaderboard

O1 on the official leaderboard passed 85% in December 2024

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 Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2024 2025 Near the ceiling (85%) o3-mini o1 Phi-4 Claude 3.5 Sonnet 0% 20% 40% 60% 80% 100% 2024 2025 Near the ceiling (85%) o3-mini o1

Sources: Official 3SAT-Search leaderboard and labs’ model cards

Labs report it only in system cards

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

2 system cards report it, from 1 lab

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
Apr 30, 2025 Phi-4-reasoning technical report 78 SAT (3SAT-Search), average pass@1 (Figure 8, printed bar label) Capabilities
Apr 30, 2025 Phi-4-Reasoning-plus technical report 70.9 SAT (3SAT-Search), average pass@1 (Figure 8, printed bar label) Capabilities

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%