SWE-Bench Pro V2 leaderboard
- Claude Opus 5 leads
- 99.4%
- models measured
- 15
- lead over #2
- +0.2 pts
- median
- 95.6%
Leader on SWE-Bench Pro V2 · read October 22
# SWE-Bench Pro V2, read October 22
leaders = {
"Claude Opus 5": 99.4%, "Gemini 4 Argon": 99.2%, "Claude Fable 5.1": 99.1%,
}
On October 22, 2026, 15 of our catalog's models had a SWE-Bench Pro V2 result, and Anthropic Claude Opus 5 led with 99.4%, ahead of Google Gemini 4 Argon at 99.2%.
SWE-Bench Pro V2 measures the share of software engineering tasks a coding agent resolves. Scale AI runs each model inside an agent harness such as Claude Code, Codex or mini-swe-agent; we show the best result per model and name the harness. This is the full set, not the HARD one.
Role in the index
A coding board we show beside the Vibe Coding Index, not inside it: the index keeps the six boards it was defined on, so a score does not move every time a board is added. Read the methodology
| # | Model | Result | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Opus 5Anthropic | 99.4% | 83.0 | $25.00 | |
| 2 | Gemini 4 ArgonGoogle | 99.2% | 81.0 | $10.00 | |
| 3 | Claude Fable 5.1Anthropic | 99.1% | 84.2 | $50.00 | |
| 4 | Kimi K3Moonshot AI | 97.7% | 62.3 | $15.00 | |
| 5 | GPT-6 AstraOpenAI | 96.9% | 85.1 | $50.00 | |
| 6 | GPT-6.1 SolOpenAI | 96.1% | 82.4 | $10.00 | |
| 7 | Claude Sonnet 5.5Anthropic | 95.8% | 87.3 | $10.00 | |
| 8 | GLM 5.3Z.ai | 95.6% | 70.6 | $4.40 | |
| 9 | GPT-5.6 SolOpenAI | 95.5% | 73.5 | $20.00 | |
| 10 | Muse Spark OpenMeta | 95.0% | 47.4 | $4.25 | |
| 11 | Gemini 3.8 FlashGoogle | 94.9% | 61.6 | $3.75 | |
| 12 | Claude Sonnet 5Anthropic | 93.2% | 57.7 | $10.00 | |
| 13 | GPT-5.6 TerraOpenAI | 92.4% | 63.9 | $12.00 | |
| 14 | T InklingThinking Machines | 89.9% | 31.5 | $4.05 | |
| 15 | Claude Haiku 5.5Anthropic | 87.2% | 70.7 | $0.50 |
Only models in our catalog are listed, and each appears once with its best published run. 66 rows were read from the source on October 22, 2026, and rows that match no catalog model are left out.