Benchmark Almanac

OpenAI MRCR v2

K-EXAONE 2.0 94.4% leads since Jul 2026 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.

The 2025-12-05 fix replaced about 10% of samples that held too many needles and about 5% with wrong ground truth. OpenAI calls the result version 2 from GPT-5.2 on, and Anthropic system cards use the same name.

Made by OpenAI Released Dec 2025 Finding facts in long inputs All tasks public Last in a launch post Oct 2026 Last model added Oct 2026
Level
Everyday Most adults, without training
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Short answer A number, word or short string matched to a key.
Dataset openai.com
9
Models scored in lab cards
7
Launch posts list it, from 3 labs
5
System cards report it, from 3 labs

K-EXAONE 2.0 leads with 94.4%

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

  1. 94.4 K-EXAONE 2.0
  2. 89.6 GPT-5.6 Terra
  3. 83.5 DeepSeek-V4-Pro
  4. 77.0 GPT-5.2
  5. 60.0 MAI-Thinking-1
  6. 52.3 K-EXAONE
  7. 45.8 Spark-X2.6-4B
  8. 44.2 GPT-5.4 Nano
  9. 41.3 GPT-5.6 Luna
  1. 1 K-EXAONE 2.0 94.4%
  2. 2 GPT-5.6 Terra 89.6%
  3. 3 DeepSeek-V4-Pro 83.5%
  4. 4 GPT-5.2 77%
  5. 5 MAI-Thinking-1 60%
  6. 6 K-EXAONE 52.3%
  7. 7 Spark-X2.6-4B 45.8%
  8. 8 GPT-5.4 Nano 44.2%
  9. 9 GPT-5.6 Luna 41.3%
#ModelScoreSettingDateSource
1 K-EXAONE 2.0 94.4% Jul 2026 Model card
2 GPT-5.6 Terra 89.6% Jul 2026 Model card
3 DeepSeek-V4-Pro 83.5% Apr 2026 Model card
4 GPT-5.2 77% Dec 2025 Model card
5 MAI-Thinking-1 60% Jun 2026 Model card
6 K-EXAONE 52.3% Dec 2025 Model card
7 Spark-X2.6-4B 45.8% Oct 2026 Model card
8 GPT-5.4 Nano 44.2% Mar 2026 Model card
9 GPT-5.6 Luna 41.3% Jul 2026 Model card

4 models have held number one since its release in Dec 2025

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

OpenAI DeepSeek LG AI Research
Lab claims Passed 85% in Jul 2026, by GPT-5.6 Terra
K-EXAONE 2.0 94.4%

The best score passed 85% in Jul 2026

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) K-EXAONE 2.0 GPT-5.6 Terra DeepSeek-V4-Pro GPT-5.2 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) K-EXAONE 2.0 DeepSeek-V4-Pro

3 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

OpenAI
LG AI Research
iFlytek
DateLaunch postScoreAs the post writes it
22 Oct 2026 Spark-X2.6-4B 45.8 Mean of Openai_MRCR v2 256k-512k 65.6 / 512k-1M 26.0
31 Jul 2026 K-EXAONE 2.0 94.4 OpenAI-MRCR (bins up to 128K, macro-average)
9 Jul 2026 GPT-5.6 Luna 41.3 OpenAI MRCR v2 8-needle 256K-512K 41.3 (512K-1M 41.3)
9 Jul 2026 GPT-5.6 Terra 89.6 OpenAI MRCR v2 8-needle 256K-512K 89.6 (512K-1M 72.5)
17 Mar 2026 GPT-5.4 Nano 44.2 OpenAI MRCR v2 8-needle 64K-128K
31 Dec 2025 K-EXAONE 52.3 OpenAI-MRCR
11 Dec 2025 GPT-5.2 77 OpenAI MRCRv2, 8 needles (128k–256k bin)

5 system cards report it, from 3 labs

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
2 Jun 2026 MAI-Thinking-1 technical report 60 OpenAI MRCR, avg_similarity@256K (Appendix J.3 text) Capabilities
26 Apr 2026 DeepSeek-V4-Pro technical report 83.5 MRCR 1M (MMR), OpenAI MRCR, DS-V4-Pro Max (Table 6). Figure 9 prints 8-needle MMR by context length Capabilities
16 Apr 2026 Claude Opus 4.7 system card OpenAI MRCR v2 8-needle, Mean Match Ratio at max effort, no overall: 256k 59.2, 1M 32.2, printed on charts (8.7.2, Figures 8.7.2.A and B) Capabilities
17 Feb 2026 Claude Sonnet 4.6 system card OpenAI MRCR v2 8-needle, Mean Match Ratio, 256K 90.3 and 1M 65.8 at max effort (90.6 and 65.1 at 64k thinking), no overall (Table 2.16.A) Capabilities
5 Feb 2026 Claude Opus 4.6 system card OpenAI MRCR v2 8-needle, Mean Match Ratio, 256K 93.0 and 1M 76.0 at max effort (91.9 and 78.3 at 64k thinking), no overall (Table 2.18.A) Capabilities

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.

OpenAI MiniMax LG AI Research DeepSeek
OpenAI MRCR v2 Passed 85% in Jul 2026, by GPT-5.6 Terra
K-EXAONE 2.0 94.4%
OpenAI MRCR Passed 85% in Aug 2026, by K-EXAONE 2.0
K-EXAONE 2.0 94.4%
OpenAI MRCR v2 Passed 85% in Jul 2026, by GPT-5.6 Terra
K-EXAONE 2.0 94.4%
Replaces OpenAI MRCR

More in Long context

All 66 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. Beam 1.0 66.9%
  • 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 5 100%
  • 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%
  • LOCA-Bench In use Agentic tasks whose environment state is scaled to grow context length while task semantics stay fixed. Success rate versus context length. GPT-5.2 21.3%