Comparing 13 models and 4 agents on SWE Tasks: Go, Java, Python, Rust, TS

Over the period May 15 - July 1, 2026, model and coding agent performance and cost were compared on 111 SWE problems from 65 repositories.

The top performers by resolve rate are as follows.

• Fable 5 [high]: 64.5%, $4.40 per problem
• Grok 4.5 [high]: 63.8%, $1.47
• Claude Opus 5 [high]: 63.4%, $3.47
• GLM-5.2 [high]: 62.9%, $1.40
• GPT-5.6 Sol [medium]: 62.3%, $0.85

Among agents, the order was Junie (61.8%), Claude Code (60.4%), Codex (58.0%), Cursor (51.7%). Cursor was the cheapest among the measured agents at $0.41 per problem.

By Pass@5, GLM-5.2 was highest at 81.1%, followed by GPT-5.6 Sol at 79.3% and Fable 5 at 78.4%. Resolve rate and Pass@5 don't necessarily align, making it difficult to rank based on a single metric alone.

In cost efficiency, MiMo V2.5 Pro ($0.10), GPT-5.6 Luna ($0.11), and DeepSeek-V4 Pro ($0.15) showed low costs, but their resolve rates were 46.5%, 43.6%, and 40.2% respectively. Many existing models and some of the newest models had their performance and cost data marked as N/A.