Breaking

Cognition Adds Claude Sonnet 5.5 to Devin Desktop and CLI as FrontierCode Score Reaches 64.4%

September 28, 2026 at 16:34 EDT

Foundation Models
AI Agents
Software Dev & Coding

Cognition has made Claude Sonnet 5.5 available in Devin Desktop and Devin CLI, putting Anthropic's new Sonnet model directly into its desktop and terminal coding-agent workflows. In Devin's FrontierCode 1.1 Main run, the model scored 64.4%, up from 56.2% for Claude Sonnet 5 and ahead of Claude Fable 5.1 at extra high reasoning effort.

Sep. 28, 2026 · Devin

Claude Sonnet 5.5 lands in Devin

The new Sonnet model is available in Devin Desktop and Devin CLI, with a 64.4% FrontierCode 1.1 Main result in Devin's coding-agent harness.

64.4%

Sonnet 5.5 in Devin

56.2%

Sonnet 5 baseline

+8.2

percentage-point gain

token context on paid Claude plans

FrontierCode 1.1 Main: same unit, proportional scale

64.4%

Sonnet 5.5

56.2%

Sonnet 5

The Devin comparison shows an 8.2-point improvement on the mergeability-focused coding benchmark.

What improves

The model brings a higher Devin score, faster generation claims and the same Sonnet token price into desktop and terminal coding workflows.

What still matters

Different harnesses and effort settings can change outcomes; too much extra work can be penalized when it exceeds the requested scope.

The story is not only model quality; it is agent behavior under real repository constraints.

The decisive question is whether the benchmark lift becomes cleaner pull requests with fewer human edits.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.