SWE-Interact: leaderboard
SWE-Interact (2026) turns 75 tasks from SWE-bench Pro, SWE Atlas and DeepSWE into multi-turn sessions in which a tool-using simulated user starts vague, reveals requirements gradually and reviews the workspace of the agent. The board is the resolve rate in percent under the original task verifiers, averaged over two user-simulator models; higher is better.
Metric: Resolve rate (%) of 75 multi-turn software-engineering tasks (25 each adapted from SWE-bench Pro, SWE Atlas refactoring and DeepSWE) in which a tool-using simulated user reveals requirements progressively and inspects the agent's workspace; the original task verifiers, mean of two runs with Claude Opus 4.7 and GPT 5.5 as the user simulator; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 (High) | 26.7 | #56 (Claude Opus 4.8) |
| 2 | GPT-5.5 (High) | 24.7 | #49 (GPT-5.5) |
| 3 | Claude Sonnet 4.6 (High) | 18.8 | #161 (Claude Sonnet 4.6) |
| 4 | Gemini 3.5 Flash (High) | 17.3 | #101 (Gemini 3.5 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=swe-interact · How It Works · Data refreshed daily, snapshot 2026-11-13.