All Benchmarks/Long context

MRCR v2 (8-needle)

Widely citedHigher is better

A long-context test: a very long synthetic conversation hides eight nearly identical requests (such as several poems on the same topic), and the model must reproduce one specific instance, such as 'the third one'. This is Google DeepMind's MRCR v2; models are tested up to different maximum lengths, so their averages cover different ranges. Scores run from 0 to 100%. Higher is better.

Top score88.1%Qwen3.6 Max
Models tested53
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Alibaba
    Qwen3.6 Max
    2026.04.27 · enabled
    88.1%
  2. 2
    OpenAI
    GPT-5.6 Sol
    2026.07.09 · max
    85.5%
  3. 3
    Alibaba
    Qwen3.6 35B A3B
    2026.04.27 · enabled
    83.5%
  4. 4
    Google
    Gemini 3.7 Flash
    2026.08.13 · high
    83.0%
  5. 5
    OpenAI
    GPT-5.5
    2026.04.23 · xhigh
    79.8%
  6. 6
    Anthropic
    Claude Opus 5
    2026.07.24 · max
    77.3%
  7. 7
    OpenAI
    GPT-5.6 Terra
    2026.07.09 · max
    76.6%
  8. 8
    Alibaba
    Qwen3.5 397B A17B
    2026.02.16 · enabled
    75.5%
  9. 9
    xAI
    Grok 4.6
    2026.08.12 · medium
    73.6%
  10. 10
    Google
    Gemini 3.6 Flash
    2026.07.21 · high
    73.3%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.