HoloBench
Models answer questions by combining facts across database rows written as sentences. A judge gives full, partial, or zero credit for each expected result.
A later Megagon Labs paper tested five language models on 90 HoloBench examples with 100,000 tokens of context.
- Made by
- Megagon Labs
- Released
- October 2024
- Kind
- Reasoning over long documents
- Level
- School
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Generated
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- o3-mini #124 on the capability index
- Last model added
- February 2025
- Scores refreshed
- November 20th, 2026
Gemini 2.0 Flash leads with 59.6% among models released by February 2025
Best score per model 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.
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:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1
Gemini 2.0 Flash Feb 2025, Constant information amount (2k) 59.6%
- 2
Gemini 2.0 Flash Thinking Jan 2025, Constant information amount (2k) 57.2%
- 3
Llama 3.1-405B Constant information amount (2k) 56.5%
- 4 GPT-4o Aug 2024, Constant information amount (2k) 55.9%
- 5 o3-mini Constant information amount (2k) 54.5%
- 6
DeepSeek-V3 Constant information amount (2k) 52.6%
- 7 o1-mini Constant information amount (2k) 51.2%
- 8 Claude 3.5 Sonnet Constant information amount (2k) 46.4%
- 9
Llama 3.1-70B Constant information amount (2k) 45.7%
- 10 GPT-4o mini Constant information amount (2k) 45.3%
- 11
Gemini 1.5 Pro May 2024, Constant information amount (2k) 40.6%
- 12
DeepSeek-R1 Constant information amount (2k) 37.5%
- 13
Gemini 1.5 Flash May 2024, Constant information amount (2k) 31.4%
- 14
Llama 3.1-8B Constant information amount (2k) 23%
Source: The HoloBench paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 59.6% | Constant information amount (2k) | Feb 2025 | Official leaderboard |
| 2 | | 57.2% | Constant information amount (2k) | Jan 2025 | Official leaderboard |
| 3 | | 56.5% | Constant information amount (2k) | Jul 2024 | Official leaderboard |
| 4 | GPT-4o Aug 2024 | 55.9% | Constant information amount (2k) | Aug 2024 | Official leaderboard |
| 5 | o3-mini | 54.5% | Constant information amount (2k) | Jan 2025 | Official leaderboard |
| 6 | | 52.6% | Constant information amount (2k) | Dec 2024 | Official leaderboard |
| 7 | o1-mini | 51.2% | Constant information amount (2k) | Sep 2024 | Official leaderboard |
| 8 | Claude 3.5 Sonnet | 46.4% | Constant information amount (2k) | Jun 2024 | Official leaderboard |
| 9 | | 45.7% | Constant information amount (2k) | Jul 2024 | Official leaderboard |
| 10 | GPT-4o mini | 45.3% | Constant information amount (2k) | Jul 2024 | Official leaderboard |
| 11 | | 40.6% | Constant information amount (2k) | May 2024 | Official leaderboard |
| 12 | | 37.5% | Constant information amount (2k) | Jan 2025 | Official leaderboard |
| 13 | | 31.4% | Constant information amount (2k) | May 2024 | Official leaderboard |
| 14 | | 23% | Constant information amount (2k) | Jul 2024 | 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 4k tokens, 8k tokens and 16k tokens. Gemini 2.0 Flash Thinking leads 3 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.
4k tokens
14 models we track have a score here.
- 1
Gemini 2.0 Flash Thinking Jan 2025 ↑1 69.5%
- 2 o3-mini ↑3 68.6%
- 3
Gemini 2.0 Flash Feb 2025 ↓2 66.2%
- 4
Llama 3.1-405B ↓1 64%
- 5 o1-mini ↑2 63.5%
- 6 GPT-4o Aug 2024 ↓2 62.5%
- 7
Llama 3.1-70B ↑2 60.5%
- 8
DeepSeek-V3 ↓2 59.5%
- 9 GPT-4o mini ↑1 51.5%
- 10
DeepSeek-R1 ↑2 51.1%
8k tokens
14 models we track have a score here.
- 1 o3-mini ↑4 62.6%
- 2
Gemini 2.0 Flash Feb 2025 ↓1 62.3%
- 3
Llama 3.1-405B 61%
- 4
DeepSeek-R1 ↑8 60.4%
- 5 GPT-4o Aug 2024 ↓1 57.5%
- 6
DeepSeek-V3 56.9%
- 7 o1-mini 55.8%
- 8
Gemini 2.0 Flash Thinking Jan 2025 ↓6 55.3%
- 9
Llama 3.1-70B 53.7%
- 10 Claude 3.5 Sonnet ↓2 52.2%
16k tokens
14 models we track have a score here.
- 1
Gemini 2.0 Flash Thinking Jan 2025 ↑1 59.5%
- 2
Gemini 2.0 Flash Feb 2025 ↓1 59.1%
- 3
Llama 3.1-405B 57.5%
- 4 GPT-4o Aug 2024 53.6%
- 5
DeepSeek-V3 ↑1 52.5%
- 6 o3-mini ↓1 51.6%
- 7 o1-mini 48.8%
- 8 Claude 3.5 Sonnet 47.2%
- 9
Llama 3.1-70B 45.9%
- 10 GPT-4o mini 43.1%
32k tokens
14 models we track have a score here.
- 1
Gemini 2.0 Flash Thinking Jan 2025 ↑1 59.6%
- 2 GPT-4o Aug 2024 ↑2 59.4%
- 3
Gemini 2.0 Flash Feb 2025 ↓2 57.3%
- 4
Llama 3.1-405B ↓1 55.1%
- 5 o1-mini ↑2 47.6%
- 6
DeepSeek-V3 47.3%
- 7 o3-mini ↓2 46.8%
- 8 GPT-4o mini ↑2 43.2%
- 9 Claude 3.5 Sonnet ↓1 42.3%
- 10
Llama 3.1-70B ↓1 42.2%
64k tokens
14 models we track have a score here.
- 1
DeepSeek-V3 ↑5 53.6%
- 2
Gemini 2.0 Flash Feb 2025 ↓1 53%
- 3 GPT-4o Aug 2024 ↑1 52%
- 4
Llama 3.1-405B ↓1 44.8%
- 5
Gemini 2.0 Flash Thinking Jan 2025 ↓3 43%
- 6 o3-mini ↓1 42.7%
- 7
Gemini 1.5 Pro May 2024 ↑4 42.2%
- 8 Claude 3.5 Sonnet 41.9%
- 9 o1-mini ↓2 40.3%
- 10 GPT-4o mini 38.7%
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.
Source: The HoloBench paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Constant information amount (2k) | 59.6% | ||||
| Constant information amount (2k) | 57.2% | ||||
| Constant information density (50%) | 56.8% | ||||
| Constant information amount (2k) | 56.5% | ||||
| GPT-4o Aug 2024 | Constant information amount (2k) | 55.9% | |||
| o3-mini | Constant information amount (2k) | 54.5% | |||
| Constant information amount (2k) | 52.6% | ||||
| o1-mini | Constant information amount (2k) | 51.2% | |||
| Constant information density (50%) | 50.1% | ||||
| Constant information density (50%) | 50.1% | ||||
| Constant information density (50%) | 50% | ||||
| GPT-4o Aug 2024 | Constant information density (50%) | 49.2% | |||
| Claude 3.5 Sonnet | Constant information amount (2k) | 46.4% | |||
| o3-mini | Constant information density (50%) | 46% | |||
| Constant information amount (2k) | 45.7% | ||||
| Constant information density (50%) | 45.5% | ||||
| GPT-4o mini | Constant information amount (2k) | 45.3% | |||
| Claude 3.5 Sonnet | Constant information density (50%) | 44.8% | |||
| o1-mini | Constant information density (50%) | 44.8% | |||
| Constant information amount (2k) | 40.6% | ||||
| GPT-4o mini | Constant information density (50%) | 39.2% | |||
| Constant information amount (2k) | 37.5% | ||||
| Constant information density (50%) | 35.9% | ||||
| Constant information density (50%) | 34.8% | ||||
| Constant information amount (2k) | 31.4% | ||||
| Constant information density (50%) | 28.3% | ||||
| Constant information amount (2k) | 23% | ||||
| Constant information density (50%) | 20.5% |
3 models have held number one in the benchmark’s paper since its release in October 2024
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The HoloBench 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 HoloBench 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 Long context
- MRCR v2 In use Multi-round coreference: retrieve the n-th of several similar requests hidden in a long synthetic conversation, up to 1M tokens.
Muse Spark 1.3 98.5%
- LongBench v2 In use 503 multiple-choice questions over 8K-2M-word contexts (QA, dialogue, code repos, structured data). Accuracy.
Inkling 2 Small 70.6%
- GraphWalks In use A graph is given as an edge list in long context. The model answers BFS or parent queries, scored by F1. Claude Fable 5 99.96%
- OpenAI MRCR v2 In use Models read a long synthetic chat that hides 2, 4 or 8 identical writing requests and must reproduce the response to the i-th one. Scored by mean string match ratio per context bin.
Gemini 3.7 Flash 89%
- CL-bench In use Model must learn new knowledge from a supplied long context (rules, procedures, data) and solve tasks. Strict LLM-verified rubrics. GPT-5.4 27.9%
- AA-LCR 1.1 In use Models answer 100 questions that need reasoning across sets of real documents of about 100K tokens. An LLM equality checker with a fixed system prompt compares each answer with the key.
Step 5 Preview 88.3%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 21 months ago, in February 2025, for Gemini 2.0 Flash. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (14 entries).
- Pass Models scored 14 models we track.
- Warning Newest score The newest came 21 months ago, in February 2025, for Gemini 2.0 Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is o3-mini (#124).
- Pass Headroom The best is 59.6%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 28 rows for models we track 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.