Cognition Launches Devin Fusion, a Coding Agent 35 Percent Cheaper

Cognition, the company behind the Devin coding agent, has launched Devin Fusion, a new architecture that pairs two language models on a single programming task. The setup is designed to deliver code quality close to that of the most expensive frontier models, but at a noticeably lower token bill.

How the model duo works

In Devin Fusion, two agents work in parallel on a single task. The first runs a frontier-class model and handles planning, interpreting ambiguous instructions, and the final code review. The second, called the sidekick, runs a cheaper model and takes over the mechanical work: searching through files, pulling information from documentation, and executing well-defined subtasks.

Both agents keep separate, cached contexts, which lets the system switch models mid-session without an expensive cache wipe. Lightweight classifiers decide on the fly whether a given instruction can be handled by the cheaper model or needs the main agent to step in. In practice, the expensive model only kicks in when the task genuinely calls for it.

Benchmark numbers

On an extended version of the FrontierCode test, the Fusion and Fable 5 combination scored 57.6 points at a cost of $3 per task, while Fable 5 alone in medium mode scored 57.0 points for $5.12. Fusion without Fable 5 scored 47.9 points for $2.38, beating both Opus 4.8 (48.8 points for $3.24) and GPT-5.5 (44.8 points for $3.64), as well as China's GLM-5.2 (43.0 points for $2.70), on cost-effectiveness.

Differences of one or two dollars per task look small, but across thousands of requests a day at a large engineering organization, they add up to real savings in an AI budget. Cognition argues that it's this arithmetic, not the benchmark score alone, that persuades engineering teams to switch tools.

Conventional model routing is worthless. It passes benchmarks, but it can't write code that anyone would actually merge - Cognition, Devin Fusion launch announcement

Cognition's valuation context

Devin Fusion arrives a month after Cognition raised more than $1 billion at a $26 billion valuation, double what it was eight months earlier. The round was led by Lux Capital, General Catalyst, and 8VC, with Founders Fund and Ribbit Capital also among the investors.

The most telling figure from that announcement is Devin's share of the company's own code: 89 percent, up from just 13 percent in December 2025. Over that period, Cognition's revenue grew from around $37 million to an annualized run rate close to $492 million, with the company naming Mercedes-Benz, NASA, Goldman Sachs, and Santander among its clients.

The coding agent price race

Devin Fusion is entering a market where the fight is no longer just about the quality of generated code, but about the price per unit of work. Cursor, acquired this year by SpaceX for $60 billion, China's Z.ai with its free ZCode, and Anthropic with Claude Code are now competing mainly on the cost of handling a single task, not just on leaderboard scores.

For Polish software companies and software houses, this translates into a concrete purchasing decision: at similar code quality, the difference in a monthly coding agent bill can run into tens of percent. Devin Fusion is currently available only as a preview, so full subscription pricing may still change once testing wraps up.

Cognition says the sidekick architecture will keep evolving alongside new frontier models, with routing set to cover additional model combinations later in the second half of 2026.

Sources: Devin Fusion announcement on the Cognition blog (cognition.com), analysis of Cognition's funding round and valuation by The Decoder (the-decoder.com).