Breaking

Cognition Launches Multi-Model 'Devin Fusion,' Cuts Cost 35%

June 29, 2026 at 11:59 EDT

AI Agents
Software Dev & Coding
Foundation Models

Cognition unveiled Devin Fusion, a hybrid-model harness for agentic coding, on June 29, 2026. The company says it reduces the cost of Fable-level intelligence by 35% while maintaining performance.

June 29, 2026 · Cognition AI

Devin Fusion: same coding smarts, ~35% cheaper

A new hybrid-model harness splits roles across multiple models — a frontier model plans and reviews while a parallel "sidekick" handles mechanical work — holding "Fable 5-level" quality at a fraction of the cost.

~35%

lower cost while holding quality

88%

of merged PRs driven by Fusion's router (internal test)

62%

savings on mechanical tasks (e.g. ES6 refactor)

Cost per task — FrontierCode Extended

Score held flat (57.0 → 57.6) while cost drops

$5.12

Fable 5
alone
score 57.0

→

$3.00

Fusion
+ Fable 5
score 57.6

How the hybrid harness works

Frontier model
plans, reads ambiguity, final review

→

Sidekick agent
runs in parallel on mechanical work

→

Dynamic routing
lightweight classifier switches models mid-session

The claim

Measured on its own FrontierCode benchmark for mergeable, production-quality code — not just correctness — Fusion keeps frontier-level quality at lower cost.

The caveats

Risk of lower scores on judgment-heavy tasks. Preview only; reference model Fable 5 access suspended since June 12. External evaluations remain limited.

Continue reading

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