Cognition replaced Opus 4.8 with Fable 5 as the lead model in Devin and said the bill went down—even though Fable costs about 2× more per token. The July 13 post from Joon Hee Lee argues that once Fusion’s sidekick architecture is in play, the right metric is cost-per-agent-run, not sticker price per million tokens.
On FrontierCode 1.1, Cognition ran about 3,000 evaluation sessions across four setups: Fable or Opus in the lead seat, each with and without the same cheaper sidekick model. Figures below are as published by Cognition (measured on their benchmark), not independent replication.
What Cognition published
Pure (no-sidekick) runs: Fable outscored Opus (60.8 vs 55.4) and cost more per run. With the same sidekick, Cognition reports cost order flipped: Fable + sidekick averaged $1.86 per run at score 60.7, versus Opus + sidekick at $2.04 and 54.6. Versus pure Fable, Fusion cut cost roughly in half while nearly holding the score.
Mechanism claim (quoted style): both leads handed off about three times per run; Fable handed off earlier. Cognition reports Fable leads took ~11.5 turns per run vs Opus ~26.5, and that in 81% of Fable-led runs the lead never made a code edit itself (24% for Opus).
Implications
Anyone pricing autonomous coding agents has been trained to stare at model list prices. Cognition’s telemetry argues agent cost is dominated by how the lead model plans—so “Fable is 2× Opus” is the wrong comparison for Fusion customers if their workloads behave like FrontierCode 1.1.
Extrapolation (labeled): if your Devin Cloud overage is billed at underlying model API prices (devin.ai/pricing), run-level Fusion savings published here still depend on your task mix matching Cognition’s harness behavior.
Sources
- Official notes: Making Fable Cheaper Than Opus
- Plan overage policy: Devin pricing
Fan coverage from Devin Central — not an official Cognition announcement. Devin is a trademark of Cognition.