Cognition

1 months agoNews Editor

Cognition cuts Devin usage costs again as Fusion tops its in-house benchmark

Cognition has rolled out another round of cost reductions for Devin, its AI coding agent that can read code, modify it, run tests, and complete development tasks. According to the update carried by BlockBeats, usage costs for Devin’s Fusion and Normal modes were reduced by 30% to 40%, while Ultra saw a 15% to 20% cut. Devin Review posted the steepest drop, with costs reduced by as much as 70%. The company also outlined changes to Devin’s execution framework. It said the system now combines multiple commands into a single call, runs independent steps in parallel, and makes greater use of context caching to reduce repeated computation and model calls. Fusion, described as Devin’s multi-model collaboration mode, assigns stronger models to planning and checking, while lower-cost models handle execution. Alongside the lower costs, Cognition reported stronger internal benchmark results. On its self-built FrontierCode 1.1 Extended benchmark, Devin Fusion scored 68.8 points with an average cost of $0.60 per task, placing it in the top spot on that benchmark.

1640

AI Agent

3 months agoNews Editor

Fable 5 beat Opus 4.8 on both score and cost in 3,000 Devin Fusion runs

A BlockTempo report, citing a post on X by Joon Lee, says Devin’s team ran 3,000 evaluation sessions on FrontierCode 1.1 to compare Fable 5 and Opus 4.8 under four setups: each model acting as the lead agent, with and without the same lower-cost helper model. The result challenged a simple per-token pricing view. Although Fable 5 costs about twice as much per token as Opus 4.8, the Fable-led setup became cheaper once delegation was introduced. In the helper-assisted runs, Fable + helper posted a higher score and a lower average cost than Opus + helper, coming in at 60.7 points and $1.86 per run versus 54.6 points and $2.04. The article argues that total agent cost is driven less by list price per token and more by how many lead-model turns are taken, how much context is carried forward, and which tasks the lead model decides not to do itself. The write-up describes Fable as delegating earlier, writing more complete handoff briefs, and spending less time pulling work back into the expensive model context. Opus, by contrast, was described as delegating later and doing more corrective work itself. The team said the pattern does not apply to every task, especially short jobs and serial debugging work that are hard to break apart.

2210

Bit.Fan