Devin adds Claude Haiku 5.5 and names it a Fusion sidekick option

Cognition, the company behind the Devin coding agent, announced on October 7, 2026 that Anthropic's Claude Haiku 5.5 is live in Devin and has joined the sidekick lineup in Devin Fusion, its setup that pairs a lead model with a sidekick model. The company's blog index carries the same summary line. Both are Devin's own pages, and no independent source we found confirms the specifics, so everything below is the company's account.

The headline number is a benchmark result. Cognition says Haiku 5.5 scores 58.4% on FrontierCode 1.1, which it describes as its proprietary benchmark grading real-world engineering tasks on quality and mergeability. It says that places the model ahead of Sonnet 5 at about an eighth of the cost per task. The post's text states no dollar figure for either model, although the page includes a score-versus-cost chart whose values we could not read from the saved text.

The second claim concerns Fusion. Cognition says that with Haiku 5.5 as the sidekick, Fusion holds a FrontierCode score of 66.2 while cutting cost and latency, and that users can try the setup in the Devin CLI with Opus 5.5 as the lead. The two scores measure different things: 58.4% is the model alone, and 66.2 is the combined system. The post gives no figures for the cost or latency reductions.

What is confirmed is limited to what the company published: the availability date, the model's place in Devin and Fusion, and the CLI configuration it recommends. The scores come from a benchmark Cognition itself designed and runs, and we found no third-party replication of them. The post's chart is labelled "FrontierCode 1.1 Extended" without saying how that relates to the 58.4% headline figure, so readers should not assume the numbers are directly comparable to other FrontierCode results.

The announcement matters mostly as a signal about how coding agents are being assembled. Rather than choosing one model, Devin is promoting a lead-and-sidekick design in which a cheaper helper model is paired with the lead, which Cognition says cuts cost and latency. Whether the savings hold on a given team's code is unverified, since the cost and latency gains are unquantified in the post. Builders weighing the option should treat the figures as vendor claims and test the configuration on their own tasks.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.