Devin cuts costs by up to 40% with SWE-2 and improved harness

Devin has implemented significant cost reductions, becoming 30-40% cheaper in Fusion and Normal modes, 15-20% cheaper in Ultra, and up to 70% cheaper in Devin Review. These improvements stem from integrating the latest models, including SWE-2, and optimizing Devin’s Cloud harness to maximize efficiency.

Model Independence and Performance

Devin leverages a mix of models to balance intelligence and cost. Devin Fusion now leads FrontierCode 1.1 Extended with a score of 68.8 at an average cost of $0.60 per task, outperforming competitors like Opus 5.5 (high) and GPT-6 Astra (high). The system utilizes different models for specific strengths:

• Opus 5.5 and GPT-6 Sol provide strong intelligence with good price-performance.
• GPT-6 Astra offers excellent computer use capabilities.
• GPT-6 Luna lowers costs for supporting tasks.
• SWE-2 and other models are used alongside these to fit specific parts of the workflow.

Harness Optimizations

Beyond model selection, engineering improvements in the harness drive efficiency. The system now supports fewer, smarter tool calls, allowing Devin to batch actions like formatting, linting, and testing into a single request. This reduces the number of turns and tokens required, with illustrative examples showing 49% fewer tokens sent for certain sequences.

Additionally, the harness is optimized for prompt caching. By maintaining a stable shared prefix, Devin reuses saved calculations across turns. Recent provider updates have made this more effective, with Opus 5.5 cache reads costing 60% less than Opus 5, and GPT-6 Sol and Luna halving cache-read prices compared to their predecessors. These changes allow Devin to process 71% fewer tokens from scratch in cached sessions.