Breaking

Cognition Previews Devin Fusion, a Hybrid-Model Harness for Agentic Coding

June 29, 2026 at 11:54 EDT

AI Agents
Software Dev & Coding
Foundation Models

Cognition AI on June 29, 2026 previewed Devin Fusion, a new hybrid-model harness for agentic coding that runs frontier and cheaper models in parallel, cutting the cost of Fable-level intelligence by 35% while maintaining practical code quality.

June 29, 2026 · Cognition AI

Devin Fusion: frontier-level coding for up to 35% less

A new hybrid-model harness pairs a cheap "sidekick" agent with a frontier model and switches models mid-session — cutting cost while holding merge-worthy code quality.

35%

cost cut for "Fable-level intelligence" across testing

88%

of merged PRs auto-driven by the Fusion router (internal users)

62%

cost cut when test-suite execution was offloaded to the sidekick

Average task cost on the FrontierCode benchmark

Same merge-worthy quality (57.6 vs 57.0) — 41% lower cost.

$3.00

Fusion + Fable 5

$5.12

Fable 5 alone

How the hybrid harness splits the work

Frontier model

Planning, disambiguation & final review

→

Sidekick agent

Cheaper helper runs in parallel on mechanical work

→

Dynamic routing

Switches models mid-session, preserving cache efficiency

Where it fits

Delegating test-suite execution to the sidekick cut cost 62% while improving quality. Strong at prototyping and iterative work.

Where it doesn't

Delegating a judgment-heavy feature caused a sharp score drop. Still needs human steering on large codebases and complex tasks.

Continue reading

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