Cognition AI releases Devin Fusion, using mid-session model routing to cut agentic coding costs by 35%

Conventional model routing sucks. It passes benchmarks but fails to write code you'd actually merge.

Introducing Devin Fusion, a new hybrid-model harness for agentic coding.

In testing, it reduces the cost of Fable-level intelligence by 35% and still feels good to use.

especially with fable-level models, i'm *very* bullish on mid-session routing.

(e.g. more routing via sub-agents, etc.)

this is how you do a great "model routing" release btw :)

https://cognition.com/blog/devin-fusion

Cache misses usually make model routing very expensive.

The "sidekick" approach avoids this in several clever ways.

This is a neat idea around doing model routing and sub-agent delegation while ensuring cache hits on accumulated context for all agents.

It makes a lot of sense: you want to ensure that all subagents can also make use of accumulated context in their cache

So exciting to see the innovation on the efficiency front.

Devin Fusion uses a hybrid-model harness built around two ideas:

First, a “sidekick” agent: a smaller agent runs in parallel with the frontier agent. The frontier agent delegates work, monitors progress, and keeps ownership of planning, ambiguity, and final review.

This lets Fusion stay intelligent while avoiding unnecessary frontier-model spend.

For more research on how we built Devin Fusion, including examples, check out our blog:

https://cognition.com/blog/devin-fusion

We developed FrontierCode to try to measure whether you’d actually merge a PR to production.

Devin Fusion builds on this research while avoiding common pitfalls of model routing, like expensive cache misses and lack of generalization.

@cognition We've found this sort of "sidekick" architecture to be very effective at cutting LLM spend because it allows you to do context control and not spend expensive tokens on simple tasks.

Here's a 200-line example of how to do it in the OpenHands SDK :) https://gist.github.com/neubig/412ab8df8e6fd0b2bdf10602d77f9d86

Second, Fusion routes dynamically mid-session.

Engineering tasks often reveal their difficulty late: a “simple bug” might become a multi-file race condition investigation. Fusion can shift work back toward the smarter model when needed, including around context-compaction moments where switching can be much cheaper.

Devin Fusion is available today in Devin. Try it out:

http://app.devin.ai/signup?source=fusion_socials

> Devin Fusion builds on this research while avoiding common pitfalls of model routing, like expensive cache misses and lack of generalization. Cognition finally showing serious engineering

@walden_yan Really like this concept.

@cognition FrontierCode is not an easy bench to win - let's go!

Cool idea! When the big model is close to compaction, we can choose between continuing, or switching for “free” to a small model to complete simpler subtasks + compacting in parallel. But I guess it only works on tasks that involve alt. sequences of hard > easy > hard subtasks. But if the task at hand doesn’t fit this structure, subagents feel more natural (if the subtasks are independent).

@cognition Cool, there's an open source version https://github.com/autohandai/routes

@Ali_Shobeiri There’s some careful tuning of this we do. It’s worth remembering that context is often still valuable. But of course cached tokens aren’t free either

@iAmHenryMascot @cognition 😬😬😬