Breaking

Cognition Integrates Anthropic's Fable 5 Into Devin Fusion, Cutting Per-Task Costs Below Opus 4.8

July 13, 2026 at 13:07 EDT

AI Agents
Software Dev & Coding
Foundation Models

Cognition AI has completed the integration of Anthropic's Claude Fable 5 into Devin Fusion, its hybrid model harness for the Devin coding agent, and says the pricier frontier model now runs at a lower cost per task than the earlier Opus 4.8. The move, confirmed on July 13, 2026, marks a notable milestone for an architecture designed to blunt the escalating cost of running top-tier models on autonomous software engineering workloads.

July 13, 2026 · Cognition AI

A pricier frontier model, a cheaper bill

Cognition has folded Anthropic's Claude Fable 5 into Devin Fusion — a hybrid harness that pairs a frontier model with cheaper "sidekick" agents. The result: Fable-level quality at a lower cost per task than the older Opus 4.8.

~41%

potential savings with the Fable 5 integration

~35%

lower cost at Fable-level quality on FrontierCode

88%

of internal merged PRs done via automatic routing

Cost per task — FrontierCode Extended

Lower is cheaper. Fusion trades a sliver of quality for a meaningful price cut.

$3.24

Opus 4.8

quality score 48.8

$2.38

Devin Fusion

quality score 47.9

~27% cheaper per task, with quality slipping less than one point.

HOW FUSION SPLITS THE WORK

Frontier model

Planning & final review (Fable 5)

→

Sidekick agents

Retrieval, code exploration, testing — separate context caches

→

Dynamic routing

Can fall back to Opus 4.8 or GPT-5.5 if needed

Offloading routine tokens to cheaper helpers lets an expensive frontier model stay economical — and softens the risk of relying on any single model.

EARLY UPSIDE

Developers call Fusion cheaper and easier to work with than calling Fable 5 directly.

THE CAVEATS

On complex multi-file work, over-delegation dropped one score from 54 to 27. Benchmarks are Cognition's own — independent verification is still limited.

Continue reading

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