All Benchmarks/Agents

CyberGym

Often citedHigher is better

A UC Berkeley security test: from a description alone, the AI must write code that reproduces 1,507 real software vulnerabilities. It has values measured by the maintainers with open agents such as OpenHands, plus maker-reported values; maker values are with safety mitigations turned off. Scores run from 0 to 100%. Higher is better.

Top score95.1%MiMo-V2.6-Flash
Models tested29
Last updated2026.09.22
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Xiaomi
    MiMo-V2.6-FlashSelf-reported
    2026.09.22
    95.1%
  2. 2
    Xiaomi
    MiMo-V2.6-ProSelf-reported
    2026.09.21
    94.0%
  3. 3
    DeepSeek
    DeepSeek V4.1 FlashSelf-reported
    2026.09.10 · max (reasoning_effort=100)
    88.1%
  4. 4
    Stepfun
    Step-5Self-reported
    2026.09.20 · High
    84.7%
  5. 5
    Z.ai
    GLM 5.3Self-reported
    2026.08.18 · max
    84.5%
  6. 6
    OpenAI
    GPT-5.6 SolSelf-reported
    2026.07.09 · with tools
    83.6%
  7. 7
    OpenAI
    GPT-5.5Self-reported
    2026.04.23 · with tools · xhigh
    81.8%
  8. 8
    xAI
    Grok 4.5Self-reported
    2026.07.08 · Mean Reproduced · with tools · high
    80.4%
  9. 9
    xAI
    Grok 4.7Self-reported
    2026.09.21 · Mean Reproduced · with tools · high
    80.3%
  10. 10
    xAI
    Grok 4.6Self-reported
    2026.08.12 · Mean Reproduced · with tools · high
    79.7%

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.

Vendor-reported numbers use varying setups — treat them as indicative.