KARLBench
Agents answer questions using vector search over fixed document collections. The suite measures six kinds of evidence gathering and reasoning.
The report compares seven named GPT and Claude models with KARL on six tasks.
- Made by
- Databricks AI Research
- Released
- March 2026
- Kind
- Questions that need search
- Level
- University
- Scored
- Graded by a model
- Answers
- Short answer
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- APIs and tools
- Refresh
- One fixed set
- Run by
- A host
- In views
- Agents
- Strongest model tested
- Claude Opus 4.6 #39 on the capability index
- Last model added
- February 2026
- Scores refreshed
- November 21st, 2026
Claude Opus 4.6 leads with 67.5 among models released by February 2026 Claude Opus 4.6 leads with 67.5 among models released by February 2026
Best score per entry from the benchmark's paper, read on November 21st, 2026. A paper’s table doesn’t add newer models. 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. This view shows only models we track, each at its best.
The newest model with a score here came out in February 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 Claude Opus 4.6 67.5
- 2 KARL parallel N = 10 67.5
- 3 Claude Sonnet 4.6 62.3
- 4 Claude Opus 4.5 61.5
- 5 GPT-5 60.1
- 6 Claude Sonnet 4.5 58.6
- 7 KARL-TREC single-task RL 56.8
- 8 KARL-BCP single-task RL 55.5
- 9
Qwen3.5 397B-A17B 55.5
- 10
MiniMax-M2.5 55.1
- 11 Claude Haiku 4.5 53.9
- 12 GPT-5.2 52.8
- 13 GLM-4.5-Air 52.6
Source: The KARLBench paper
- 1 Claude Opus 4.6 67.5
- 2 Claude Sonnet 4.6 62.3
- 3 Claude Opus 4.5 61.5
- 4 GPT-5 60.1
- 5 Claude Sonnet 4.5 58.6
- 6
Qwen3.5 397B-A17B 55.5
- 7
MiniMax-M2.5 55.1
- 8 Claude Haiku 4.5 53.9
- 9 GPT-5.2 52.8
- 10 GLM-4.5-Air 52.6
Source: The KARLBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.6 | 67.5 | Feb 2026 | Official leaderboard | |
| 2 | KARL | 67.5 | parallel N = 10 | Official leaderboard | |
| 3 | Claude Sonnet 4.6 | 62.3 | Feb 2026 | Official leaderboard | |
| 4 | Claude Opus 4.5 | 61.5 | Nov 2025 | Official leaderboard | |
| 5 | GPT-5 | 60.1 | Aug 2025 | Official leaderboard | |
| 6 | Claude Sonnet 4.5 | 58.6 | Sep 2025 | Official leaderboard | |
| 7 | KARL-TREC | 56.8 | single-task RL | Official leaderboard | |
| 8 | KARL-BCP | 55.5 | single-task RL | Official leaderboard | |
| 9 | | 55.5 | Feb 2026 | Official leaderboard | |
| 10 | | 55.1 | Feb 2026 | Official leaderboard | |
| 11 | Claude Haiku 4.5 | 53.9 | Oct 2025 | Official leaderboard | |
| 12 | GPT-5.2 | 52.8 | Dec 2025 | Official leaderboard | |
| 13 | GLM-4.5-Air | 52.6 | Jul 2025 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.6 | 67.5 | Feb 2026 | Official leaderboard | |
| 2 | Claude Sonnet 4.6 | 62.3 | Feb 2026 | Official leaderboard | |
| 3 | Claude Opus 4.5 | 61.5 | Nov 2025 | Official leaderboard | |
| 4 | GPT-5 | 60.1 | Aug 2025 | Official leaderboard | |
| 5 | Claude Sonnet 4.5 | 58.6 | Sep 2025 | Official leaderboard | |
| 6 | | 55.5 | Feb 2026 | Official leaderboard | |
| 7 | | 55.1 | Feb 2026 | Official leaderboard | |
| 8 | Claude Haiku 4.5 | 53.9 | Oct 2025 | Official leaderboard | |
| 9 | GPT-5.2 | 52.8 | Dec 2025 | Official leaderboard | |
| 10 | GLM-4.5-Air | 52.6 | Jul 2025 | Official leaderboard |
Rankings by part
We read 6 parts from the benchmark’s paper, each ranking models on one part of the test, such as browseCompPlus, trecBiogen and freshStack. Claude Opus 4.6 leads 4 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
browseCompPlus
10 models we track have a score here.
- 1 Claude Opus 4.6 75.9%
- 2 GPT-5 ↑2 68.3%
- 3 Claude Opus 4.5 62.5%
- 4 Claude Sonnet 4.6 ↓2 57.9%
- 5
MiniMax-M2.5 ↑2 56.5%
- 6
Qwen3.5 397B-A17B 55.8%
- 7 Claude Sonnet 4.5 ↓2 54.6%
- 8 GPT-5.2 ↑1 47.8%
- 9 Claude Haiku 4.5 ↓1 45.8%
- 10 GLM-4.5-Air 44.7%
trecBiogen
10 models we track have a score here.
- 1 Claude Opus 4.6 79.9%
- 2 Claude Sonnet 4.6 77.7%
- 3 Claude Sonnet 4.5 ↑2 75.2%
- 4 Claude Opus 4.5 ↓1 74.7%
- 5 Claude Haiku 4.5 ↑3 72.4%
- 6
MiniMax-M2.5 ↑1 69.3%
- 7 GPT-5 ↓3 68.2%
- 8
Qwen3.5 397B-A17B ↓2 68.2%
- 9 GLM-4.5-Air ↑1 66%
- 10 GPT-5.2 ↓1 62%
freshStack
10 models we track have a score here.
- 1 Claude Sonnet 4.6 ↑1 62.6%
- 2 Claude Opus 4.6 ↓1 61.4%
- 3 Claude Opus 4.5 57.4%
- 4 GPT-5 55.6%
- 5 Claude Sonnet 4.5 55%
- 6
MiniMax-M2.5 ↑1 53.3%
- 7 GLM-4.5-Air ↑3 52.9%
- 8
Qwen3.5 397B-A17B ↓2 51.9%
- 9 Claude Haiku 4.5 ↓1 48.7%
- 10 GPT-5.2 ↓1 47.9%
financeBench
10 models we track have a score here.
- 1 GPT-5 ↑3 86.7%
- 2 Claude Opus 4.6 ↓1 83%
- 3 Claude Sonnet 4.6 ↓1 81.3%
- 4 Claude Opus 4.5 ↓1 80.7%
- 5 GPT-5.2 ↑4 80.3%
- 6 Claude Sonnet 4.5 ↓1 79.3%
- 7
Qwen3.5 397B-A17B ↓1 79.3%
- 8
MiniMax-M2.5 ↓1 78%
- 9 Claude Haiku 4.5 ↓1 73.7%
- 10 GLM-4.5-Air 72.7%
qampari
10 models we track have a score here.
- 1 Claude Opus 4.6 58.6%
- 2 Claude Opus 4.5 ↑1 54.9%
- 3 Claude Sonnet 4.5 ↑2 54.8%
- 4 Claude Sonnet 4.6 ↓2 50.2%
- 5 Claude Haiku 4.5 ↑3 48%
- 6 GLM-4.5-Air ↑4 45.9%
- 7 GPT-5 ↓3 44.4%
- 8
Qwen3.5 397B-A17B ↓2 42.8%
- 9 GPT-5.2 41.1%
- 10
MiniMax-M2.5 ↓3 39.3%
pmBench
10 models we track have a score here.
- 1 Claude Opus 4.6 46.1%
- 2 Claude Sonnet 4.6 43.8%
- 3 Claude Opus 4.5 39.1%
- 4 GPT-5.2 ↑5 37.9%
- 5 GPT-5 ↓1 37.5%
- 6 Claude Haiku 4.5 ↑2 35%
- 7
Qwen3.5 397B-A17B ↓1 34.7%
- 8
MiniMax-M2.5 ↓1 34.5%
- 9 GLM-4.5-Air ↑1 33.4%
- 10 Claude Sonnet 4.5 ↓5 32.6%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The KARLBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official KARLBench 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 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
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 9 months ago, in February 2026, for Claude Sonnet 4.6. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (13 entries).
- Pass Models scored 10 models we track, plus 3 other entries on its leaderboard.
- Warning Newest score The newest came 9 months ago, in February 2026, for Claude Sonnet 4.6.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude Opus 4.6 (#39).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on databricks.com: 10 rows for models we track and 3 other entries at the last read.
- Pass Last read November 21st, 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.