Benchmark Almanac

OpenAI MRCR v2

Gemini 3.7 Flash 89% leads since August 2026
GPT-6 Astra 100% leads since September 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 December 5th, 2025 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
December 2025
Kind
Finding facts in long inputs
Level
Everyday
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Strongest model tested
GPT-6 Astra #3 on the capability index
Last in a launch post
October 2026
Last model added
October 2026
Scores refreshed
November 12th, 2026
48
Entries on Context Arena’s leaderboard, 47 models we track
12
Launch posts list it, from 3 labs
5
System cards report it, from 3 labs

Gemini 3.7 Flash leads with 89% among models released by August 2026

Best score per entry from Context Arena’s leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

GPT-6 Astra leads with 100%, 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 August 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

48 entries. Scroll sideways for the rest.

  1. 89.0 Gemini 3.7 Flash
  2. 83.5 GPT-5.6 Sol
  3. 75.3 GPT-5.5
  4. 73.9 Gemini 3.6 Flash
  5. 72.7 Claude Sonnet 4.6
  6. 71.8 Claude Opus 4.6
  7. 71.2 Qwen 3.8 Max
  8. 67.9 GPT-5.6 Terra
  9. 65.6 Claude Opus 5
  10. 61.8 Claude Opus 4.8
  11. 60.1 GLM-5.3-Flash
  12. 59.0 Muse Spark 1.2
  13. 57.8 GLM-5.3
  14. 54.6 Gemini 3.5 Flash
  15. 52.7 GPT-5.4
  16. 52.7 Muse Spark 1.1
  17. 52.2 Claude Sonnet 5
  18. 47.7 Grok 4.6
  19. 47.4 Gemini 3.1 Pro
  20. 46.5 GLM-5.2
  21. 45.1 GPT-5.6 Luna
  22. 44.9 Gemini 3 Flash
  23. 44.5 Qwen3.7-Max
  24. 44.5 DeepSeek V4 Flash 0731
  25. 43.1 Qwen3.7 Flash
  26. 41.0 Qwen 3.6 Plus
  27. 40.3 Gemini 3.5 Flash-Lite
  28. 38.9 DeepSeek V4 Pro 0813
  29. 38.1 Qwen 3.6 Flash
  30. 34.8 Kimi K3
  31. 33.2 DeepSeek-V4-Pro
  32. 33.0 GPT-5.4 Nano
  33. 31.3 DeepSeek-V4-Flash
  34. 29.7 Qwen3.5-Plus
  35. 28.3 GPT-5.4 Mini
  36. 28.0 MiMo-V2.5
  37. 23.5 MiMo-V2-Pro
  38. 22.5 Gemini 3.1 Flash-Lite
  39. 20.3 Qwen3.5-Flash
  40. 18.7 pokee/pokee-isaac
  41. 18.6 MiMo-V2.5-Pro
  42. 18.6 MiniMax-M3
  43. 15.9 Grok 4.20
  44. 14.4 Solar Pro 4
  45. 10.0 Inkling-Small
  46. 6.64 Grok 4.3
  47. 4.49 Laguna S 2.1
  48. 2.74 Claude Opus 4.7
  1. 1 Gemini 3.7 Flash 8-needle, 128K to 256K bin, high reasoning 89%
  2. 2 GPT-5.6 Sol 8-needle, 128K to 256K bin, max reasoning 83.5%
  3. 3 GPT-5.5 8-needle, 128K to 256K bin, high reasoning 75.3%
  4. 4 Gemini 3.6 Flash 8-needle, 128K to 256K bin, high reasoning 73.9%
  5. 5 Claude Sonnet 4.6 8-needle, 128K to 256K bin, low reasoning 72.7%
  6. 6 Claude Opus 4.6 8-needle, 128K to 256K bin, medium reasoning 71.8%
  7. 7 Qwen 3.8 Max 8-needle, 128K to 256K bin, low reasoning 71.2%
  8. 8 GPT-5.6 Terra 8-needle, 128K to 256K bin, max reasoning 67.9%
  9. 9 Claude Opus 5 8-needle, 128K to 256K bin, max reasoning 65.6%
  10. 10 Claude Opus 4.8 8-needle, 128K to 256K bin, max reasoning 61.8%
  11. 11 GLM-5.3-Flash 8-needle, 128K to 256K bin, low reasoning 60.1%
  12. 12 Muse Spark 1.2 8-needle, 128K to 256K bin, high reasoning 59%
  13. 13 GLM-5.3 8-needle, 128K to 256K bin, max reasoning 57.8%
  14. 14 Gemini 3.5 Flash 8-needle, 128K to 256K bin, high reasoning 54.6%
  15. 15 GPT-5.4 8-needle, 128K to 256K bin, medium reasoning 52.7%
  16. 16 Muse Spark 1.1 8-needle, 128K to 256K bin, minimal reasoning 52.7%
  17. 17 Claude Sonnet 5 8-needle, 128K to 256K bin, max reasoning 52.2%
  18. 18 Grok 4.6 8-needle, 128K to 256K bin, xhigh reasoning 47.7%
  19. 19 Gemini 3.1 Pro 8-needle, 128K to 256K bin, high reasoning 47.4%
  20. 20 GLM-5.2 8-needle, 128K to 256K bin, max reasoning 46.5%
  21. 21 GPT-5.6 Luna 8-needle, 128K to 256K bin, max reasoning 45.1%
  22. 22 Gemini 3 Flash 8-needle, 128K to 256K bin, high reasoning 44.9%
  23. 23 Qwen3.7-Max 8-needle, 128K to 256K bin 44.5%
  24. 24 DeepSeek V4 Flash 0731 8-needle, 128K to 256K bin, max reasoning 44.5%
  25. 25 Qwen3.7 Flash 8-needle, 128K to 256K bin, enabled reasoning 43.1%
  26. 26 Qwen 3.6 Plus 8-needle, 128K to 256K bin 41%
  27. 27 Gemini 3.5 Flash-Lite 8-needle, 128K to 256K bin, medium reasoning 40.3%
  28. 28 DeepSeek V4 Pro 0813 8-needle, 128K to 256K bin, high reasoning 38.9%
  29. 29 Qwen 3.6 Flash 8-needle, 128K to 256K bin 38.1%
  30. 30 Kimi K3 8-needle, 128K to 256K bin, max reasoning 34.8%
  31. 31 DeepSeek-V4-Pro 8-needle, 128K to 256K bin, enabled reasoning 33.2%
  32. 32 GPT-5.4 Nano 8-needle, 128K to 256K bin, xhigh reasoning 33%
  33. 33 DeepSeek-V4-Flash 8-needle, 128K to 256K bin, enabled reasoning 31.3%
  34. 34 Qwen3.5-Plus 8-needle, 128K to 256K bin, enabled reasoning 29.7%
  35. 35 GPT-5.4 Mini 8-needle, 128K to 256K bin, high reasoning 28.3%
  36. 36 MiMo-V2.5 8-needle, 128K to 256K bin, enabled reasoning 28%
  37. 37 MiMo-V2-Pro 8-needle, 128K to 256K bin 23.5%
  38. 38 Gemini 3.1 Flash-Lite 8-needle, 128K to 256K bin, high reasoning 22.5%
  39. 39 Qwen3.5-Flash 8-needle, 128K to 256K bin 20.3%
  40. 40 pokee/pokee-isaac 8-needle, 128K to 256K bin, enabled reasoning 18.7%
  41. 41 MiMo-V2.5-Pro 8-needle, 128K to 256K bin 18.6%
  42. 42 MiniMax-M3 8-needle, 128K to 256K bin, enabled reasoning 18.6%
  43. 43 Grok 4.20 8-needle, 128K to 256K bin, enabled reasoning 15.9%
  44. 44 Solar Pro 4 8-needle, 128K to 256K bin, enabled reasoning 14.4%
  45. 45 Inkling-Small 8-needle, 128K to 256K bin, minimal reasoning 10%
  46. 46 Grok 4.3 8-needle, 128K to 256K bin 6.64%
  47. 47 Laguna S 2.1 8-needle, 128K to 256K bin, enabled reasoning 4.49%
  48. 48 Claude Opus 4.7 8-needle, 128K to 256K bin, medium reasoning 2.74%

Source: Context Arena’s leaderboard

53 models. Scroll sideways for the rest.

  1. 100.0 GPT-6 Astra
  2. 94.4 K-EXAONE 2.0
  3. 91.5 GPT-5.6 Sol
  4. 89.6 GPT-5.6 Terra
  5. 89.0 Gemini 3.7 Flash
  6. 87.5 GPT-5.5
  7. 83.5 DeepSeek-V4-Pro
  8. 79.3 GPT-5.4
  9. 77.0 GPT-5.2
  10. 73.9 Gemini 3.6 Flash
  11. 72.7 Claude Sonnet 4.6
  12. 71.8 Claude Opus 4.6
  13. 71.2 Qwen 3.8 Max
  14. 65.6 Claude Opus 5
  15. 61.8 Claude Opus 4.8
  16. 60.1 GLM-5.3-Flash
  17. 60.0 MAI-Thinking-1
  18. 59.0 Muse Spark 1.2
  19. 57.8 GLM-5.3
  20. 54.6 Gemini 3.5 Flash
  21. 52.7 Muse Spark 1.1
  22. 52.3 K-EXAONE
  23. 52.2 Claude Sonnet 5
  24. 47.7 Grok 4.6
  25. 47.7 GPT-5.4 Mini
  26. 47.4 Gemini 3.1 Pro
  27. 46.5 GLM-5.2
  28. 45.8 Spark-X2.6-4B
  29. 45.1 GPT-5.6 Luna
  30. 44.9 Gemini 3 Flash
  31. 44.5 Qwen3.7-Max
  32. 44.5 DeepSeek V4 Flash 0731
  33. 44.2 GPT-5.4 Nano
  34. 43.1 Qwen3.7 Flash
  35. 41.0 Qwen 3.6 Plus
  36. 40.3 Gemini 3.5 Flash-Lite
  37. 38.9 DeepSeek V4 Pro 0813
  38. 38.1 Qwen 3.6 Flash
  39. 34.8 Kimi K3
  40. 31.3 DeepSeek-V4-Flash
  41. 29.7 Qwen3.5-Plus
  42. 28.0 MiMo-V2.5
  43. 23.5 MiMo-V2-Pro
  44. 22.5 Gemini 3.1 Flash-Lite
  45. 20.3 Qwen3.5-Flash
  46. 18.6 MiMo-V2.5-Pro
  47. 18.6 MiniMax-M3
  48. 15.9 Grok 4.20
  49. 14.4 Solar Pro 4
  50. 10.0 Inkling-Small
  51. 6.64 Grok 4.3
  52. 4.49 Laguna S 2.1
  53. 2.74 Claude Opus 4.7
  1. 1 GPT-6 Astra 100%
  2. 2 K-EXAONE 2.0 94.4%
  3. 3 GPT-5.6 Sol 91.5%
  4. 4 GPT-5.6 Terra 89.6%
  5. 5 Gemini 3.7 Flash 8-needle, 128K to 256K bin, high reasoning 89%
  6. 6 GPT-5.5 87.5%
  7. 7 DeepSeek-V4-Pro 83.5%
  8. 8 GPT-5.4 79.3%
  9. 9 GPT-5.2 77%
  10. 10 Gemini 3.6 Flash 8-needle, 128K to 256K bin, high reasoning 73.9%
  11. 11 Claude Sonnet 4.6 8-needle, 128K to 256K bin, low reasoning 72.7%
  12. 12 Claude Opus 4.6 8-needle, 128K to 256K bin, medium reasoning 71.8%
  13. 13 Qwen 3.8 Max 8-needle, 128K to 256K bin, low reasoning 71.2%
  14. 14 Claude Opus 5 8-needle, 128K to 256K bin, max reasoning 65.6%
  15. 15 Claude Opus 4.8 8-needle, 128K to 256K bin, max reasoning 61.8%
  16. 16 GLM-5.3-Flash 8-needle, 128K to 256K bin, low reasoning 60.1%
  17. 17 MAI-Thinking-1 60%
  18. 18 Muse Spark 1.2 8-needle, 128K to 256K bin, high reasoning 59%
  19. 19 GLM-5.3 8-needle, 128K to 256K bin, max reasoning 57.8%
  20. 20 Gemini 3.5 Flash 8-needle, 128K to 256K bin, high reasoning 54.6%
  21. 21 Muse Spark 1.1 8-needle, 128K to 256K bin, minimal reasoning 52.7%
  22. 22 K-EXAONE 52.3%
  23. 23 Claude Sonnet 5 8-needle, 128K to 256K bin, max reasoning 52.2%
  24. 24 Grok 4.6 8-needle, 128K to 256K bin, xhigh reasoning 47.7%
  25. 25 GPT-5.4 Mini 47.7%
  26. 26 Gemini 3.1 Pro 8-needle, 128K to 256K bin, high reasoning 47.4%
  27. 27 GLM-5.2 8-needle, 128K to 256K bin, max reasoning 46.5%
  28. 28 Spark-X2.6-4B 45.8%
  29. 29 GPT-5.6 Luna 8-needle, 128K to 256K bin, max reasoning 45.1%
  30. 30 Gemini 3 Flash 8-needle, 128K to 256K bin, high reasoning 44.9%
  31. 31 Qwen3.7-Max 8-needle, 128K to 256K bin 44.5%
  32. 32 DeepSeek V4 Flash 0731 8-needle, 128K to 256K bin, max reasoning 44.5%
  33. 33 GPT-5.4 Nano 44.2%
  34. 34 Qwen3.7 Flash 8-needle, 128K to 256K bin, enabled reasoning 43.1%
  35. 35 Qwen 3.6 Plus 8-needle, 128K to 256K bin 41%
  36. 36 Gemini 3.5 Flash-Lite 8-needle, 128K to 256K bin, medium reasoning 40.3%
  37. 37 DeepSeek V4 Pro 0813 8-needle, 128K to 256K bin, high reasoning 38.9%
  38. 38 Qwen 3.6 Flash 8-needle, 128K to 256K bin 38.1%
  39. 39 Kimi K3 8-needle, 128K to 256K bin, max reasoning 34.8%
  40. 40 DeepSeek-V4-Flash 8-needle, 128K to 256K bin, enabled reasoning 31.3%
  41. 41 Qwen3.5-Plus 8-needle, 128K to 256K bin, enabled reasoning 29.7%
  42. 42 MiMo-V2.5 8-needle, 128K to 256K bin, enabled reasoning 28%
  43. 43 MiMo-V2-Pro 8-needle, 128K to 256K bin 23.5%
  44. 44 Gemini 3.1 Flash-Lite 8-needle, 128K to 256K bin, high reasoning 22.5%
  45. 45 Qwen3.5-Flash 8-needle, 128K to 256K bin 20.3%
  46. 46 MiMo-V2.5-Pro 8-needle, 128K to 256K bin 18.6%
  47. 47 MiniMax-M3 8-needle, 128K to 256K bin, enabled reasoning 18.6%
  48. 48 Grok 4.20 8-needle, 128K to 256K bin, enabled reasoning 15.9%
  49. 49 Solar Pro 4 8-needle, 128K to 256K bin, enabled reasoning 14.4%
  50. 50 Inkling-Small 8-needle, 128K to 256K bin, minimal reasoning 10%
  51. 51 Grok 4.3 8-needle, 128K to 256K bin 6.64%
  52. 52 Laguna S 2.1 8-needle, 128K to 256K bin, enabled reasoning 4.49%
  53. 53 Claude Opus 4.7 8-needle, 128K to 256K bin, medium reasoning 2.74%

Sources: Labs’ model cards and Context Arena’s leaderboard

#EntryScoreSettingDateSource
1 Gemini 3.7 Flash 89% 8-needle, 128K to 256K bin, high reasoning Aug 2026 Official leaderboard
2 GPT-5.6 Sol 83.5% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
3 GPT-5.5 75.3% 8-needle, 128K to 256K bin, high reasoning Apr 2026 Official leaderboard
4 Gemini 3.6 Flash 73.9% 8-needle, 128K to 256K bin, high reasoning Jul 2026 Official leaderboard
5 Claude Sonnet 4.6 72.7% 8-needle, 128K to 256K bin, low reasoning Feb 2026 Official leaderboard
6 Claude Opus 4.6 71.8% 8-needle, 128K to 256K bin, medium reasoning Feb 2026 Official leaderboard
7 Qwen 3.8 Max 71.2% 8-needle, 128K to 256K bin, low reasoning Aug 2026 Official leaderboard
8 GPT-5.6 Terra 67.9% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
9 Claude Opus 5 65.6% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
10 Claude Opus 4.8 61.8% 8-needle, 128K to 256K bin, max reasoning May 2026 Official leaderboard
11 GLM-5.3-Flash 60.1% 8-needle, 128K to 256K bin, low reasoning Aug 2026 Official leaderboard
12 Muse Spark 1.2 59% 8-needle, 128K to 256K bin, high reasoning Aug 2026 Official leaderboard
13 GLM-5.3 57.8% 8-needle, 128K to 256K bin, max reasoning Aug 2026 Official leaderboard
14 Gemini 3.5 Flash 54.6% 8-needle, 128K to 256K bin, high reasoning May 2026 Official leaderboard
15 GPT-5.4 52.7% 8-needle, 128K to 256K bin, medium reasoning Mar 2026 Official leaderboard
16 Muse Spark 1.1 52.7% 8-needle, 128K to 256K bin, minimal reasoning Jul 2026 Official leaderboard
17 Claude Sonnet 5 52.2% 8-needle, 128K to 256K bin, max reasoning Jun 2026 Official leaderboard
18 Grok 4.6 47.7% 8-needle, 128K to 256K bin, xhigh reasoning Aug 2026 Official leaderboard
19 Gemini 3.1 Pro 47.4% 8-needle, 128K to 256K bin, high reasoning Feb 2026 Official leaderboard
20 GLM-5.2 46.5% 8-needle, 128K to 256K bin, max reasoning Jun 2026 Official leaderboard
21 GPT-5.6 Luna 45.1% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
22 Gemini 3 Flash 44.9% 8-needle, 128K to 256K bin, high reasoning Dec 2025 Official leaderboard
23 Qwen3.7-Max 44.5% 8-needle, 128K to 256K bin May 2026 Official leaderboard
24 DeepSeek V4 Flash 0731 44.5% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
25 Qwen3.7 Flash 43.1% 8-needle, 128K to 256K bin, enabled reasoning Jul 2026 Official leaderboard
26 Qwen 3.6 Plus 41% 8-needle, 128K to 256K bin Mar 2026 Official leaderboard
27 Gemini 3.5 Flash-Lite 40.3% 8-needle, 128K to 256K bin, medium reasoning Jul 2026 Official leaderboard
28 DeepSeek V4 Pro 0813 38.9% 8-needle, 128K to 256K bin, high reasoning Aug 2026 Official leaderboard
29 Qwen 3.6 Flash 38.1% 8-needle, 128K to 256K bin Apr 2026 Official leaderboard
30 Kimi K3 34.8% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
31 DeepSeek-V4-Pro 33.2% 8-needle, 128K to 256K bin, enabled reasoning Apr 2026 Official leaderboard
32 GPT-5.4 Nano 33% 8-needle, 128K to 256K bin, xhigh reasoning Mar 2026 Official leaderboard
33 DeepSeek-V4-Flash 31.3% 8-needle, 128K to 256K bin, enabled reasoning Apr 2026 Official leaderboard
34 Qwen3.5-Plus 29.7% 8-needle, 128K to 256K bin, enabled reasoning Feb 2026 Official leaderboard
35 GPT-5.4 Mini 28.3% 8-needle, 128K to 256K bin, high reasoning Mar 2026 Official leaderboard
36 MiMo-V2.5 28% 8-needle, 128K to 256K bin, enabled reasoning Apr 2026 Official leaderboard
37 MiMo-V2-Pro 23.5% 8-needle, 128K to 256K bin Mar 2026 Official leaderboard
38 Gemini 3.1 Flash-Lite 22.5% 8-needle, 128K to 256K bin, high reasoning Mar 2026 Official leaderboard
39 Qwen3.5-Flash 20.3% 8-needle, 128K to 256K bin Feb 2026 Official leaderboard
40 pokee/pokee-isaac 18.7% 8-needle, 128K to 256K bin, enabled reasoning Official leaderboard
41 MiMo-V2.5-Pro 18.6% 8-needle, 128K to 256K bin Apr 2026 Official leaderboard
42 MiniMax-M3 18.6% 8-needle, 128K to 256K bin, enabled reasoning Jun 2026 Official leaderboard
43 Grok 4.20 15.9% 8-needle, 128K to 256K bin, enabled reasoning Feb 2026 Official leaderboard
44 Solar Pro 4 14.4% 8-needle, 128K to 256K bin, enabled reasoning Aug 2026 Official leaderboard
45 Inkling-Small 10% 8-needle, 128K to 256K bin, minimal reasoning Jul 2026 Official leaderboard
46 Grok 4.3 6.64% 8-needle, 128K to 256K bin Apr 2026 Official leaderboard
47 Laguna S 2.1 4.49% 8-needle, 128K to 256K bin, enabled reasoning Jul 2026 Official leaderboard
48 Claude Opus 4.7 2.74% 8-needle, 128K to 256K bin, medium reasoning Apr 2026 Official leaderboard
#ModelScoreSettingDateSource
1 GPT-6 Astra 100% Sep 2026 Lab model card
2 K-EXAONE 2.0 94.4% Jul 2026 Lab model card
3 GPT-5.6 Sol 91.5% Jul 2026 Lab model card
4 GPT-5.6 Terra 89.6% Jul 2026 Lab model card
5 Gemini 3.7 Flash 89% 8-needle, 128K to 256K bin, high reasoning Aug 2026 Official leaderboard
6 GPT-5.5 87.5% Apr 2026 Lab model card
7 DeepSeek-V4-Pro 83.5% Apr 2026 Lab model card
8 GPT-5.4 79.3% Mar 2026 Lab model card
9 GPT-5.2 77% Dec 2025 Lab model card
10 Gemini 3.6 Flash 73.9% 8-needle, 128K to 256K bin, high reasoning Jul 2026 Official leaderboard
11 Claude Sonnet 4.6 72.7% 8-needle, 128K to 256K bin, low reasoning Feb 2026 Official leaderboard
12 Claude Opus 4.6 71.8% 8-needle, 128K to 256K bin, medium reasoning Feb 2026 Official leaderboard
13 Qwen 3.8 Max 71.2% 8-needle, 128K to 256K bin, low reasoning Aug 2026 Official leaderboard
14 Claude Opus 5 65.6% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
15 Claude Opus 4.8 61.8% 8-needle, 128K to 256K bin, max reasoning May 2026 Official leaderboard
16 GLM-5.3-Flash 60.1% 8-needle, 128K to 256K bin, low reasoning Aug 2026 Official leaderboard
17 MAI-Thinking-1 60% Jun 2026 Lab model card
18 Muse Spark 1.2 59% 8-needle, 128K to 256K bin, high reasoning Aug 2026 Official leaderboard
19 GLM-5.3 57.8% 8-needle, 128K to 256K bin, max reasoning Aug 2026 Official leaderboard
20 Gemini 3.5 Flash 54.6% 8-needle, 128K to 256K bin, high reasoning May 2026 Official leaderboard
21 Muse Spark 1.1 52.7% 8-needle, 128K to 256K bin, minimal reasoning Jul 2026 Official leaderboard
22 K-EXAONE 52.3% Dec 2025 Lab model card
23 Claude Sonnet 5 52.2% 8-needle, 128K to 256K bin, max reasoning Jun 2026 Official leaderboard
24 Grok 4.6 47.7% 8-needle, 128K to 256K bin, xhigh reasoning Aug 2026 Official leaderboard
25 GPT-5.4 Mini 47.7% Mar 2026 Lab model card
26 Gemini 3.1 Pro 47.4% 8-needle, 128K to 256K bin, high reasoning Feb 2026 Official leaderboard
27 GLM-5.2 46.5% 8-needle, 128K to 256K bin, max reasoning Jun 2026 Official leaderboard
28 Spark-X2.6-4B 45.8% Oct 2026 Lab model card
29 GPT-5.6 Luna 45.1% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
30 Gemini 3 Flash 44.9% 8-needle, 128K to 256K bin, high reasoning Dec 2025 Official leaderboard
31 Qwen3.7-Max 44.5% 8-needle, 128K to 256K bin May 2026 Official leaderboard
32 DeepSeek V4 Flash 0731 44.5% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
33 GPT-5.4 Nano 44.2% Mar 2026 Lab model card
34 Qwen3.7 Flash 43.1% 8-needle, 128K to 256K bin, enabled reasoning Jul 2026 Official leaderboard
35 Qwen 3.6 Plus 41% 8-needle, 128K to 256K bin Mar 2026 Official leaderboard
36 Gemini 3.5 Flash-Lite 40.3% 8-needle, 128K to 256K bin, medium reasoning Jul 2026 Official leaderboard
37 DeepSeek V4 Pro 0813 38.9% 8-needle, 128K to 256K bin, high reasoning Aug 2026 Official leaderboard
38 Qwen 3.6 Flash 38.1% 8-needle, 128K to 256K bin Apr 2026 Official leaderboard
39 Kimi K3 34.8% 8-needle, 128K to 256K bin, max reasoning Jul 2026 Official leaderboard
40 DeepSeek-V4-Flash 31.3% 8-needle, 128K to 256K bin, enabled reasoning Apr 2026 Official leaderboard
41 Qwen3.5-Plus 29.7% 8-needle, 128K to 256K bin, enabled reasoning Feb 2026 Official leaderboard
42 MiMo-V2.5 28% 8-needle, 128K to 256K bin, enabled reasoning Apr 2026 Official leaderboard
43 MiMo-V2-Pro 23.5% 8-needle, 128K to 256K bin Mar 2026 Official leaderboard
44 Gemini 3.1 Flash-Lite 22.5% 8-needle, 128K to 256K bin, high reasoning Mar 2026 Official leaderboard
45 Qwen3.5-Flash 20.3% 8-needle, 128K to 256K bin Feb 2026 Official leaderboard
46 MiMo-V2.5-Pro 18.6% 8-needle, 128K to 256K bin Apr 2026 Official leaderboard
47 MiniMax-M3 18.6% 8-needle, 128K to 256K bin, enabled reasoning Jun 2026 Official leaderboard
48 Grok 4.20 15.9% 8-needle, 128K to 256K bin, enabled reasoning Feb 2026 Official leaderboard
49 Solar Pro 4 14.4% 8-needle, 128K to 256K bin, enabled reasoning Aug 2026 Official leaderboard
50 Inkling-Small 10% 8-needle, 128K to 256K bin, minimal reasoning Jul 2026 Official leaderboard
51 Grok 4.3 6.64% 8-needle, 128K to 256K bin Apr 2026 Official leaderboard
52 Laguna S 2.1 4.49% 8-needle, 128K to 256K bin, enabled reasoning Jul 2026 Official leaderboard
53 Claude Opus 4.7 2.74% 8-needle, 128K to 256K bin, medium reasoning Apr 2026 Official leaderboard

Labs report scores for 14 models in their cards. Switch the scores above to rank them together.

More effort adds a median 6 points

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.

Lowest effort to highest, light to dark
Gemini 3.7 Flash
8-needle, 1… 8-needle, 1…
89%
GPT-5.5
8-needle, 1… 8-needle, 1… 8-needle, 1…
75.3%
Gemini 3.6 Flash
8-needle, 1… 8-needle, 1… 8-needle, 1…
73.9%
Claude Sonnet 4.6
8-needle, 1… 8-needle, 1…
72.7%
Claude Opus 4.6
8-needle, 1… 8-needle, 1…
71.8%
Qwen 3.8 Max
8-needle, 1…
71.2%
GLM-5.3-Flash
8-needle, 1… 8-needle, 1… 8-needle, 1…
60.1%
Muse Spark 1.2
8-needle, 1… 8-needle, 1… 8-needle, 1…
59%
GLM-5.3
8-needle, 1…
57.8%
Gemini 3.5 Flash
8-needle, 1… 8-needle, 1… 8-needle, 1…
54.6%
GPT-5.4
8-needle, 1… 8-needle, 1… 8-needle, 1…
52.7%
Muse Spark 1.1
8-needle, 1… 8-needle, 1…
52.7%

Source: Context Arena’s leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 3.7 Flash 8-needle, 128K to 256K bin, high reasoning 89%
GPT-5.6 Sol 8-needle, 128K to 256K bin, max reasoning 83.5%
Gemini 3.7 Flash 8-needle, 128K to 256K bin, medium reasoning 82.1%
Gemini 3.7 Flash 8-needle, 128K to 256K bin, low reasoning 79.5%
GPT-5.5 8-needle, 128K to 256K bin, high reasoning 75.3%
Gemini 3.6 Flash 8-needle, 128K to 256K bin, high reasoning 73.9%
GPT-5.5 8-needle, 128K to 256K bin, xhigh reasoning 73.2%
Claude Sonnet 4.6 8-needle, 128K to 256K bin, low reasoning 72.7%
GPT-5.5 8-needle, 128K to 256K bin, medium reasoning 71.9%
Claude Opus 4.6 8-needle, 128K to 256K bin, medium reasoning 71.8%
Qwen 3.8 Max 8-needle, 128K to 256K bin, low reasoning 71.2%
Qwen 3.8 Max 8-needle, 128K to 256K bin, medium reasoning 70.5%
Claude Sonnet 4.6 8-needle, 128K to 256K bin, high reasoning 69.3%
Claude Opus 4.6 8-needle, 128K to 256K bin, high reasoning 69.1%
Claude Sonnet 4.6 8-needle, 128K to 256K bin, medium reasoning 68.5%
Claude Opus 4.6 8-needle, 128K to 256K bin, low reasoning 68.5%
GPT-5.6 Terra 8-needle, 128K to 256K bin, max reasoning 67.9%
Qwen 3.8 Max 8-needle, 128K to 256K bin, xhigh reasoning 67.7%
GPT-5.5 8-needle, 128K to 256K bin, low reasoning 67%
Gemini 3.6 Flash 8-needle, 128K to 256K bin, medium reasoning 66.4%
Gemini 3.6 Flash 8-needle, 128K to 256K bin, low reasoning 65.7%
Claude Opus 5 8-needle, 128K to 256K bin, max reasoning 65.6%
Claude Opus 4.8 8-needle, 128K to 256K bin, max reasoning 61.8%
GLM-5.3-Flash 8-needle, 128K to 256K bin, low reasoning 60.1%
Muse Spark 1.2 8-needle, 128K to 256K bin, high reasoning 59%
GLM-5.3 8-needle, 128K to 256K bin, max reasoning 57.8%
Muse Spark 1.2 8-needle, 128K to 256K bin, medium reasoning 55.8%
GLM-5.3 8-needle, 128K to 256K bin, high reasoning 55.8%
GLM-5.3-Flash 8-needle, 128K to 256K bin, max reasoning 55.6%
Gemini 3.5 Flash 8-needle, 128K to 256K bin, high reasoning 54.6%
Muse Spark 1.2 8-needle, 128K to 256K bin, xhigh reasoning 52.8%
GPT-5.4 8-needle, 128K to 256K bin, medium reasoning 52.7%
Muse Spark 1.1 8-needle, 128K to 256K bin, minimal reasoning 52.7%
GLM-5.3 8-needle, 128K to 256K bin, low reasoning 52.5%
Muse Spark 1.2 8-needle, 128K to 256K bin, low reasoning 52.2%
Claude Sonnet 5 8-needle, 128K to 256K bin, max reasoning 52.2%
Muse Spark 1.2 8-needle, 128K to 256K bin, minimal reasoning 52.2%
Muse Spark 1.1 8-needle, 128K to 256K bin, medium reasoning 51.3%
GPT-5.4 8-needle, 128K to 256K bin, high reasoning 51.2%
Muse Spark 1.1 8-needle, 128K to 256K bin, low reasoning 49.1%
GLM-5.3-Flash 8-needle, 128K to 256K bin, high reasoning 48.6%
Muse Spark 1.1 8-needle, 128K to 256K bin, high reasoning 48%
GPT-5.4 8-needle, 128K to 256K bin, xhigh reasoning 47.8%
Grok 4.6 8-needle, 128K to 256K bin, xhigh reasoning 47.7%
Gemini 3.1 Pro 8-needle, 128K to 256K bin, high reasoning 47.4%
Grok 4.6 8-needle, 128K to 256K bin, medium reasoning 46.8%
Muse Spark 1.1 8-needle, 128K to 256K bin, xhigh reasoning 46.7%
GLM-5.2 8-needle, 128K to 256K bin, max reasoning 46.5%
Claude Sonnet 4.6 8-needle, 128K to 256K bin 46.1%
Grok 4.6 8-needle, 128K to 256K bin, high reasoning 45.8%
GPT-5.6 Luna 8-needle, 128K to 256K bin, max reasoning 45.1%
Gemini 3 Flash 8-needle, 128K to 256K bin, high reasoning 44.9%
Qwen3.7-Max 8-needle, 128K to 256K bin 44.5%
DeepSeek V4 Flash 0731 8-needle, 128K to 256K bin, max reasoning 44.5%
Gemini 3.1 Pro 8-needle, 128K to 256K bin, low reasoning 44.1%
GPT-5.4 8-needle, 128K to 256K bin, low reasoning 43.4%
Qwen3.7 Flash 8-needle, 128K to 256K bin, enabled reasoning 43.1%
Grok 4.6 8-needle, 128K to 256K bin, low reasoning 41.8%
Gemini 3.5 Flash 8-needle, 128K to 256K bin, medium reasoning 41.7%
Qwen 3.6 Plus 8-needle, 128K to 256K bin 41%
Gemini 3.5 Flash-Lite 8-needle, 128K to 256K bin, medium reasoning 40.3%
Qwen3.7 Flash 8-needle, 128K to 256K bin 40.1%
Claude Opus 4.6 8-needle, 128K to 256K bin 40%
Gemini 3 Flash 8-needle, 128K to 256K bin, medium reasoning 39.8%
Gemini 3 Flash 8-needle, 128K to 256K bin, low reasoning 39.6%
DeepSeek V4 Flash 0731 8-needle, 128K to 256K bin, high reasoning 39.6%
Qwen 3.6 Plus 8-needle, 128K to 256K bin, enabled reasoning 39%
DeepSeek V4 Pro 0813 8-needle, 128K to 256K bin, high reasoning 38.9%
Qwen 3.6 Flash 8-needle, 128K to 256K bin 38.1%
Gemini 3.5 Flash-Lite 8-needle, 128K to 256K bin, high reasoning 37.6%
DeepSeek V4 Pro 0813 8-needle, 128K to 256K bin, max reasoning 36.1%
Gemini 3.5 Flash-Lite 8-needle, 128K to 256K bin, low reasoning 35.8%
Kimi K3 8-needle, 128K to 256K bin, max reasoning 34.8%
Gemini 3.6 Flash 8-needle, 128K to 256K bin, minimal reasoning 34.2%
DeepSeek V4 Pro 0813 8-needle, 128K to 256K bin, low reasoning 34.1%
DeepSeek-V4-Pro 8-needle, 128K to 256K bin, enabled reasoning 33.2%
GPT-5.4 Nano 8-needle, 128K to 256K bin, xhigh reasoning 33%
DeepSeek-V4-Flash 8-needle, 128K to 256K bin, enabled reasoning 31.3%
Qwen3.5-Plus 8-needle, 128K to 256K bin, enabled reasoning 29.7%
DeepSeek V4 Flash 0731 8-needle, 128K to 256K bin, low reasoning 29.5%
Kimi K3 8-needle, 128K to 256K bin, high reasoning 29.4%
GPT-5.4 Mini 8-needle, 128K to 256K bin, high reasoning 28.3%
MiMo-V2.5 8-needle, 128K to 256K bin, enabled reasoning 28%
Gemini 3 Flash 8-needle, 128K to 256K bin 27.7%
GPT-5.5 8-needle, 128K to 256K bin 27.5%
Gemini 3.5 Flash 8-needle, 128K to 256K bin, low reasoning 27.5%
Qwen3.5-Plus 8-needle, 128K to 256K bin 26.4%
GPT-5.4 Nano 8-needle, 128K to 256K bin, high reasoning 26.3%
DeepSeek V4 Flash 0731 8-needle, 128K to 256K bin 26%
DeepSeek V4 Pro 0813 8-needle, 128K to 256K bin 25.9%
GPT-5.4 Mini 8-needle, 128K to 256K bin, xhigh reasoning 25.2%
Gemini 3.5 Flash 8-needle, 128K to 256K bin 24.7%
MiMo-V2-Pro 8-needle, 128K to 256K bin 23.5%
MiMo-V2.5 8-needle, 128K to 256K bin 23.4%
GPT-5.4 Mini 8-needle, 128K to 256K bin, medium reasoning 23%
DeepSeek-V4-Pro 8-needle, 128K to 256K bin 22.6%
Gemini 3.1 Flash-Lite 8-needle, 128K to 256K bin, high reasoning 22.5%
DeepSeek-V4-Flash 8-needle, 128K to 256K bin 22.4%
Gemini 3.1 Flash-Lite 8-needle, 128K to 256K bin, low reasoning 22.1%
Kimi K3 8-needle, 128K to 256K bin, low reasoning 21.8%
MiMo-V2-Pro 8-needle, 128K to 256K bin, enabled reasoning 21.5%
GPT-5.4 8-needle, 128K to 256K bin 20.9%
Gemini 3.1 Flash-Lite 8-needle, 128K to 256K bin, medium reasoning 20.4%
Qwen3.5-Flash 8-needle, 128K to 256K bin 20.3%
GPT-5.4 Nano 8-needle, 128K to 256K bin, medium reasoning 20.2%
Gemini 3.1 Flash-Lite 8-needle, 128K to 256K bin 20%
MiMo-V2.5-Pro 8-needle, 128K to 256K bin 18.6%
MiniMax-M3 8-needle, 128K to 256K bin, enabled reasoning 18.6%
GPT-5.4 Mini 8-needle, 128K to 256K bin, low reasoning 18.1%
MiMo-V2.5-Pro 8-needle, 128K to 256K bin, enabled reasoning 17%
GPT-5.4 Nano 8-needle, 128K to 256K bin, low reasoning 16.9%
Grok 4.20 8-needle, 128K to 256K bin, enabled reasoning 15.9%
Gemini 3.5 Flash-Lite 8-needle, 128K to 256K bin, minimal reasoning 15%
Solar Pro 4 8-needle, 128K to 256K bin, enabled reasoning 14.4%
Solar Pro 4 8-needle, 128K to 256K bin 13.2%
GPT-5.4 Mini 8-needle, 128K to 256K bin 12.2%
Grok 4.20 8-needle, 128K to 256K bin 10.4%
Inkling-Small 8-needle, 128K to 256K bin, minimal reasoning 10%
GPT-5.4 Nano 8-needle, 128K to 256K bin 9.81%
Inkling-Small 8-needle, 128K to 256K bin, medium reasoning 8.22%
Inkling-Small 8-needle, 128K to 256K bin, low reasoning 8.17%
Inkling-Small 8-needle, 128K to 256K bin, high reasoning 7.55%
Inkling-Small 8-needle, 128K to 256K bin, max reasoning 7.52%
Inkling-Small 8-needle, 128K to 256K bin 6.85%
Grok 4.3 8-needle, 128K to 256K bin 6.64%
Laguna S 2.1 8-needle, 128K to 256K bin, enabled reasoning 4.49%
Claude Opus 4.7 8-needle, 128K to 256K bin, medium reasoning 2.74%
Claude Opus 4.7 8-needle, 128K to 256K bin, xhigh reasoning 2.16%
Claude Opus 4.7 8-needle, 128K to 256K bin, low reasoning 1.97%
Claude Opus 4.7 8-needle, 128K to 256K bin, high reasoning 1.43%
Claude Opus 4.7 8-needle, 128K to 256K bin 1.24%

6 models have held number one on Context Arena’s leaderboard since its release in December 2025 6 models have held number one with claims included since its release in December 2025

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for Context Arena’s leaderboard and what labs claim in their cards.

Google DeepMind Anthropic OpenAI LG AI Research
Context Arena leaderboard Passed 85% in Aug 2026, by Gemini 3.7 Flash
Gemini 3.7 Flash 89%
Lab claims Passed 85% in Apr 2026, by GPT-5.5
GPT-6 Astra 100%

Sources: Context Arena’s leaderboard and labs’ model cards

OpenAI’s own number for GPT-5.5 passed 85% in April 2026

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

On Context Arena’s leaderboard Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra K-EXAONE 2.0 GPT-5.5 GPT-5.4 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra GPT-5.6 Sol

Sources: Context Arena’s leaderboard and labs’ model cards

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
Oct 22, 2026 Spark-X2.6-4B 45.8 Mean of Openai_MRCR v2 256k-512k 65.6 / 512k-1M 26.0
Sep 3, 2026 GPT-6 Astra 100 OpenAI MRCR v2 8-needle 256K-512K 100.0% (512K-1M 96.3%)
Jul 31, 2026 K-EXAONE 2.0 94.4 OpenAI-MRCR (bins up to 128K, macro-average)
Jul 9, 2026 GPT-5.6 Luna 41.3 OpenAI MRCR v2 8-needle 256K-512K 41.3 (512K-1M 41.3)
Jul 9, 2026 GPT-5.6 Terra 89.6 OpenAI MRCR v2 8-needle 256K-512K 89.6 (512K-1M 72.5)
Jul 9, 2026 GPT-5.6 Sol 91.5 OpenAI MRCR v2 8-needle 256K-512K 91.5% (512K-1M 73.8%)
Apr 23, 2026 GPT-5.5 87.5 OpenAI MRCR v2 8-needle 128K-256K 87.5% (512K-1M 74.0%)
Mar 17, 2026 GPT-5.4 Nano 44.2 OpenAI MRCR v2 8-needle 64K-128K
Mar 17, 2026 GPT-5.4 Mini 47.7 OpenAI MRCR v2 8-needle, 64K-128K
Mar 5, 2026 GPT-5.4 79.3 OpenAI MRCR v2 8-needle 128K-256K (36.6% at 512K-1M)
Dec 31, 2025 K-EXAONE 52.3 OpenAI-MRCR
Dec 11, 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
Jun 2, 2026 MAI-Thinking-1 technical report 60 OpenAI MRCR, avg_similarity@256K (Appendix J.3 text) Capabilities
Apr 26, 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
Apr 16, 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
Feb 17, 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
Feb 5, 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

The lab's number comes out ahead in 7 of 8 pairs that differ

Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. Every benchmark.

Model Lab says Board says Gap
DeepSeek-V4-Pro 83.5% MRCR 1M (MMR), OpenAI MRCR, DS-V4-Pro Max (Table 6). Figure 9 prints 8-needle MMR by context length 33.2% Leaderboard, 8-needle, 128K to 256K bin, enabled reasoning +50 pts
GPT-5.4 79.3% OpenAI MRCR v2 8-needle 128K-256K (36.6% at 512K-1M) 52.7% Leaderboard, 8-needle, 128K to 256K bin, medium reasoning +27 pts
GPT-5.6 Terra 89.6% OpenAI MRCR v2 8-needle 256K-512K 89.6 (512K-1M 72.5) 67.9% Leaderboard, 8-needle, 128K to 256K bin, max reasoning +22 pts
GPT-5.4 Mini 47.7% OpenAI MRCR v2 8-needle, 64K-128K 28.3% Leaderboard, 8-needle, 128K to 256K bin, high reasoning +19 pts
GPT-5.5 87.5% OpenAI MRCR v2 8-needle 128K-256K 87.5% (512K-1M 74.0%) 75.3% Leaderboard, 8-needle, 128K to 256K bin, high reasoning +12 pts
GPT-5.4 Nano 44.2% OpenAI MRCR v2 8-needle 64K-128K 33% Leaderboard, 8-needle, 128K to 256K bin, xhigh reasoning +11 pts
GPT-5.6 Sol 91.5% OpenAI MRCR v2 8-needle 256K-512K 91.5% (512K-1M 73.8%) 83.5% Leaderboard, 8-needle, 128K to 256K bin, max reasoning +8.0 pts
GPT-5.6 Luna 41.3% OpenAI MRCR v2 8-needle 256K-512K 41.3 (512K-1M 41.3) 45.1% Leaderboard, 8-needle, 128K to 256K bin, max reasoning −3.8 pts

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 Google DeepMind Anthropic
OpenAI MRCR v2 Passed 85% in Aug 2026, by Gemini 3.7 Flash
Gemini 3.7 Flash 89%
OpenAI MRCR Passed 85% in Aug 2026, by K-EXAONE 2.0
K-EXAONE 2.0 94.4%
OpenAI MRCR v2 Passed 85% in Aug 2026, by Gemini 3.7 Flash
Gemini 3.7 Flash 89%

Sources: Official leaderboards and labs’ model cards

Replaces OpenAI MRCR

More in Long context

All 79 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%
  • 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%