Breaking

Cognition unveils Devin Fusion, a multi-model setup cutting cost 35%

June 30, 2026 at 11:08 EDT

AI Agents
Software Dev & Coding
Foundation Models

Cognition AI on June 29, 2026 unveiled "Devin Fusion," a new architecture for its autonomous coding agent Devin. By pairing a high-performance frontier model with a smaller "sidekick" model, the company says it delivers frontier-level performance while cutting inference cost by 35%.

June 29, 2026 · Cognition AI

Devin Fusion: Frontier-Level Coding at a Fraction of the Cost

A hybrid harness pairs a frontier model with a smaller "sidekick," delegating well-scoped sub-tasks while keeping cached context — matching top models at up to 41% lower cost.

35%

lower inference cost vs. frontier models, at matching scores

41%

cost cut when combined with Fable 5

88%

of merged PRs driven fully automatically by the Fusion router

FrontierCode Extended

Same score, lower cost — cost shown as column height

$3.00

57.6

Fusion
+ Fable 5

$5.12

57.0

Fable 5
alone

$2.38

47.9

Fusion
alone

Numbers inside columns = quality score. The Fusion + Fable 5 pairing holds the top score (57.6) at the lowest cost of the two high-scoring configs.

How Fusion divides the work

Frontier model

Plans the implementation

→

Delegate sub-tasks

Well-scoped, cached context

→

Sidekick model

Fetches & skims in parallel

Mid-session routing switches models on the fly via a lightweight classifier — limiting cache misses and context loss that plagued simple routing.

Per-task savings

Test-suite delegation cut cost 62% ($3.55 → $1.37); mechanical refactoring cut 32%.

Caveats

Fable 5 inaccessible since June 12 (gov. directive); results predate the pause. Devin still starts at $500/mo and struggles with complex codebases. Now in preview.

Continue reading

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