AI-modeller
Alla topplistor
Kod & mjukvara · Scale AI (Scale Labs)

SWE-Bench Pro V2

Hur ofta en AI-agent på egen hand lyckas rätta ett verkligt fel eller bygga en ny funktion i ett stort, verkligt programvaruprojekt, så att alla tester går igenom.

Placeringar (16, varav 14 i vår lista)

  1. 1
    Anthropic Opus 5 (Claude Code) xhigh Inaktuell
    Claude Code, tankenivå xhigh
    99.4% ± 0.4[1]
  2. 2
    Google Gemini 4 Argon (mini-swe-agent) high
    mini-swe-agent, tankenivå high
    99.22% ± 0.45[1]
  3. 3
    Anthropic Fable 5.1 (Claude Code) high
    Claude Code, tankenivå high
    99.1% ± 0.5[1]
  4. 4
    Moonshot AI Kimi-K3 (mini-swe-agent) max Inaktuell
    mini-swe-agent, tankenivå max
    97.7% ± 0.9[1]
  5. 5
    OpenAI GPT-6-Astra (Codex) high
    Codex, tankenivå high
    96.9% ± 1.1[1]
  6. 6
    Meta Muse Wonder 1.0 (mini-swe-agent) xhigh
    mini-swe-agent, tankenivå xhigh
    96.73% ± 1.12[1]
  7. 7
    OpenAI GPT-6.1-Sol (mini-swe-agent) xhigh
    mini-swe-agent, tankenivå xhigh
    96.11% ± 1.24[1]
  8. 8
    Anthropic Sonnet 5.5 (mini-swe-agent) xhigh
    mini-swe-agent, tankenivå xhigh
    95.79% ± 1.30[1]
  9. 9
    Z.ai (Zhipu) GLM-5.3 (mini-swe-agent) max
    mini-swe-agent, tankenivå max
    95.6% ± 1.3[1]
  10. 10
    OpenAI GPT-5.6-sol (Codex) xhigh Inaktuell
    Codex, tankenivå xhigh
    95.5% ± 1.4[1]
  11. 11
    Meta Muse Spark Open (mini-swe-agent) xhigh
    mini-swe-agent, tankenivå xhigh
    95.02% ± 1.43[1]
  12. 12
    Google Gemini 3.8 Flash (mini-swe-agent) high
    mini-swe-agent, tankenivå high
    94.86% ± 1.46[1]
  13. 13
    Anthropic Sonnet 5 (Claude Code) xhigh
    Anthropic · Claude Code, tankenivå xhigh
    93.15% ± 1.71[1]
  14. 14
    OpenAI GPT-5.6-Terra (Codex) xhigh
    Codex, tankenivå xhigh
    92.37% ± 1.81[1]
  15. 15
    Inkling (mini-swe-agent) xhigh
    Thinking Machines · mini-swe-agent, tankenivå xhigh
    89.88% ± 2.10[1]
  16. 16
    Anthropic Haiku 5.5 (mini-swe-agent) high
    mini-swe-agent, tankenivå high
    87.23% ± 2.36[1]

Källor

  1. [1]
    SWE-Bench Pro V2 Leaderboard ↗
    Scale AITopplistalabs.scale.com
    Publicerad 16 okt 2026 Läst 25 okt 2026
  2. [2]
    SWE-bench Leaderboards (Verified m.fl.) ↗
    SWE-bench (Princeton/Stanford)Topplistaswebench.com
    Publicerad 26 feb 2026 Läst 25 okt 2026

Historik

  1. 25 okt 2026
    Fler placeringar länkade till modeller som lagts till i listan.
  2. 25 okt 2026
    Tillagd. Avläst 25 okt 2026 från SWE-Bench Pro V2 (Full), som ersatt SWE-bench Verified som det test fältet använder.