Devin Gets Up to 40% Cheaper Across Fusion, Normal, Ultra, and Devin Review

Cognition announced on September 28, 2026 that Devin is now substantially cheaper to run across every mode, crediting newer underlying models and a series of harness-level efficiency improvements rather than a simple price cut.

Details

• Pricing changes by mode: Fusion and Normal modes are 30-40% cheaper, Ultra mode is 15-20% cheaper, and Devin Review is up to 70% cheaper than before
• Model lineup driving the change: Devin now draws on newer models including SWE-2, Claude Opus 5.5, GPT-6 Sol, GPT-6 Astra, and GPT-6 Luna, with Cognition noting that “different models have different strengths, and new releases can change what’s possible at a given cost”
• Tool call batching: newer models can bundle multiple operations, such as formatting, linting, and testing, into a single request, which Cognition says cut token consumption by roughly 49% in illustrative scenarios
• Prompt caching: reusing context across agent sessions cut tokens processed from scratch by about 71% in sample workflows, aided by improved cache pricing from providers including Anthropic and OpenAI
• Benchmark performance: Devin Fusion leads the FrontierCode 1.1 Extended leaderboard with a score of 68.8, at an average cost of about $0.60 per task

What happened next

The price cut arrives the same day Cognition made Claude Sonnet 5.5 available inside Devin, and just three days after the company disclosed it had crossed $1 billion in annualized revenue run rate. Rather than resting on that growth, Cognition is using its expanding menu of frontier models and infrastructure optimizations to push Devin’s cost-per-task down while keeping benchmark performance at the top of its own FrontierCode leaderboard, a combination aimed at making autonomous coding agents cheaper to run at enterprise scale even as usage grows.