Benchmark Almanac

ATM-Bench

Claude Opus 4.5 86% leads since March 2026
Not in model cards

Agents answer questions about a person’s history using images, videos, and email records. The Harbor package contains 31 hard questions with schema-guided memory.

The README agent table available November 19, 2026 reports GPT-5.5 at extra-high effort first with 48.1%, followed by Claude Opus 4.7 at 46.6%.

Made by
Jingbiao Mei and colleagues
Released
March 2026
Kind
Finding facts in long inputs
Level
University
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images, Video → Text, Actions
Interaction
Agentic
Environment
APIs and tools
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
In views
Agents, Vision
Strongest model tested
GPT-5.4 #28 on the capability index
Last model added
June 2026
Scores refreshed
November 20th, 2026
29
Entries on its own leaderboard, 13 models we track

Claude Opus 4.5 leads with 86% among models released by June 2026 Claude Opus 4.5 leads with 86% among models released by June 2026

Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran on its own with several settings, we keep its best. Agents other companies built around a model get a pale shade of that model's lab color. Gray marks blends and entries not in our model list, such as fine-tunes and older models. None of these count toward leaders. Agents built around a model, such as Graphon V1 + GPT-6 Astra, are their own entries. We don't count them as the model's score. This view shows only models we track, each at its best.

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

  1. 1 Claude Opus 4.5 Oracle 86%
  2. 2 GPT-5 Oracle 85.3%
  3. 3 GPT-5.4 Oracle 85.1%
  4. 4 Qwen3.6 27B Oracle 84.5%
  5. 5 GPT-5.2 Oracle 84.1%
  6. 6 MiniMax-M3 Oracle 83.6%
  7. 7 Claude Sonnet 4.5 Oracle 83.3%
  8. 8 MiMo-V2.5 Oracle 82.6%
  9. 9 Graphon V1 + GPT-6 Astra Memory / Graphon V1 / Graphon V1 81.7%
  10. 10 Graphon V1 + DeepSeek V4.1 Flash Memory / Graphon V1 / Graphon V1 79.2%
  11. 11 Gemini 2.5 Pro Jun 2025, Oracle 78.6%
  12. 12 Qwen3-VL-8B-Instruct Oracle 78.2%
  13. 13 Gemini 2.5 Flash Jun 2025, Oracle 76.9%
  14. 14 DeepSeek V4.1 Memory / DeepSeek V4.1 / BGE-M3 + bge-reranker-v2-m3 76.7%
  15. 15 MiMo-V2.5-Pro Oracle 71.7%
  16. 16 DeepSeek V4 Pro (OpenRouter) RAG / Qwen3-VL-2B-Instruct / Voyage-4-large + Voyage rerank-2.5 68.4%
  17. 17 Memexa + DeepSeek-V4-Flash Memory / DeepSeek-V4-flash + Qwen3.6-27B captions / BGE-M3 + bge-reranker-v2-m3 68%
  18. 18 MemPalace + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 56.8%
  19. 19 Qwen3-VL-8B-Instruct (OpenRouter) RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 + BM25 56.1%
  20. 20 ScrapMem (No-Forget) + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 52.5%
  21. 21 ATM-RAG + Qwen3-VL-8B-Instruct RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 51%
  22. 22 Self-RAG + Qwen3-VL-8B-Instruct RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 50.3%
  23. 23 ScrapMem + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 48.4%
  24. 24 MemoryOS + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 47.2%
  25. 25 A-Mem + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 44.8%
  26. 26 Mem0 + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 43.5%
  27. 27 HippoRAG2 + Qwen3-VL-8B-Instruct RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 42.9%
  28. 28 Qwen3-VL-2B-Instruct Oracle 34.8%
  29. 29 SimpleMem + Qwen3-VL-8B-Instruct Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 27.3%

Source: Official ATM-Bench leaderboard

#EntryScoreSettingDateLinkSource
1 Claude Opus 4.5 86% Oracle Nov 2025 Official leaderboard
2 GPT-5 85.3% Oracle Aug 2025 Official leaderboard
3 GPT-5.4 85.1% Oracle Mar 2026 Official leaderboard
4 Qwen3.6 27B 84.5% Oracle Apr 2026 Official leaderboard
5 GPT-5.2 84.1% Oracle Dec 2025 Official leaderboard
6 MiniMax-M3 83.6% Oracle Jun 2026 Official leaderboard
7 Claude Sonnet 4.5 83.3% Oracle Sep 2025 Official leaderboard
8 MiMo-V2.5 82.6% Oracle Apr 2026 Official leaderboard
9 Graphon V1 + GPT-6 Astra 81.7% Memory / Graphon V1 / Graphon V1 Sep 2026 graphon.ai Official leaderboard
10 Graphon V1 + DeepSeek V4.1 Flash 79.2% Memory / Graphon V1 / Graphon V1 Sep 2026 graphon.ai Official leaderboard
11 Gemini 2.5 Pro Jun 2025 78.6% Oracle Jun 2025 Official leaderboard
12 Qwen3-VL-8B-Instruct 78.2% Oracle Oct 2025 Official leaderboard
13 Gemini 2.5 Flash Jun 2025 76.9% Oracle Jun 2025 Official leaderboard
14 DeepSeek V4.1 76.7% Memory / DeepSeek V4.1 / BGE-M3 + bge-reranker-v2-m3, Memexa v2 github.com Official leaderboard
15 MiMo-V2.5-Pro 71.7% Oracle Apr 2026 Official leaderboard
16 DeepSeek V4 Pro (OpenRouter) 68.4% RAG / Qwen3-VL-2B-Instruct / Voyage-4-large + Voyage rerank-2.5, RE-call github.com Official leaderboard
17 Memexa + DeepSeek-V4-Flash 68% Memory / DeepSeek-V4-flash + Qwen3.6-27B captions / BGE-M3 + bge-reranker-v2-m3 Apr 2026 github.com Official leaderboard
18 MemPalace + Qwen3-VL-8B-Instruct 56.8% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
19 Qwen3-VL-8B-Instruct (OpenRouter) 56.1% RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 + BM25, MediaFind mediafind.io Official leaderboard
20 ScrapMem (No-Forget) + Qwen3-VL-8B-Instruct 52.5% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 arxiv.org Official leaderboard
21 ATM-RAG + Qwen3-VL-8B-Instruct 51% RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
22 Self-RAG + Qwen3-VL-8B-Instruct 50.3% RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
23 ScrapMem + Qwen3-VL-8B-Instruct 48.4% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 arxiv.org Official leaderboard
24 MemoryOS + Qwen3-VL-8B-Instruct 47.2% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
25 A-Mem + Qwen3-VL-8B-Instruct 44.8% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
26 Mem0 + Qwen3-VL-8B-Instruct 43.5% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
27 HippoRAG2 + Qwen3-VL-8B-Instruct 42.9% RAG / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
28 Qwen3-VL-2B-Instruct 34.8% Oracle Oct 2025 Official leaderboard
29 SimpleMem + Qwen3-VL-8B-Instruct 27.3% Memory / Qwen3-VL-2B-Instruct / MiniLM-L6 Oct 2025 github.com Official leaderboard
#ModelScoreSettingDateLinkSource
1 Claude Opus 4.5 86% Oracle Nov 2025 Official leaderboard
2 GPT-5 85.3% Oracle Aug 2025 Official leaderboard
3 GPT-5.4 85.1% Oracle Mar 2026 Official leaderboard
4 Qwen3.6 27B 84.5% Oracle Apr 2026 Official leaderboard
5 GPT-5.2 84.1% Oracle Dec 2025 Official leaderboard
6 MiniMax-M3 83.6% Oracle Jun 2026 Official leaderboard
7 Claude Sonnet 4.5 83.3% Oracle Sep 2025 Official leaderboard
8 MiMo-V2.5 82.6% Oracle Apr 2026 Official leaderboard
9 Gemini 2.5 Pro Jun 2025 78.6% Oracle Jun 2025 Official leaderboard
10 Qwen3-VL-8B-Instruct 78.2% Oracle Oct 2025 Official leaderboard
11 Gemini 2.5 Flash Jun 2025 76.9% Oracle Jun 2025 Official leaderboard
12 MiMo-V2.5-Pro 71.7% Oracle Apr 2026 Official leaderboard
13 Qwen3-VL-2B-Instruct 34.8% Oracle Oct 2025 Official leaderboard

Number one over time

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

Anthropic
Own leaderboard Passed 85% in Mar 2026, by Claude Opus 4.5

Source: Official ATM-Bench leaderboard

GPT-5 on the official leaderboard passed 85% in August 2025

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
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Opus 4.5 GPT-5 Gemini 2.5 Pro 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Opus 4.5 Gemini 2.5 Pro

Source: Official ATM-Bench 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

All 115 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

All 9 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
  • Pass Independent runs Scores from the official leaderboard (29 entries).
  • Pass Models scored 13 models we track, plus 16 other entries on its leaderboard.
  • Pass Newest score The newest is from June 2026, for MiniMax-M3.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is GPT-5.4 (#28).
  • Note Headroom Saturated. The best is 86%, past 85% since August 2025.
Leaderboard import 3 pass
  • Pass Importer We read the board on atmbench.github.io: 27 rows for models we track and 3 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