AI agent developer Cognition has announced the release of "Fusion," a new harness designed to optimize the performance and cost-efficiency of AI agents such as OpenAI's GPT-6 Astra and Anthropic's Claude Fable. The system is now available in Devin Desktop and CLI.
Fusion utilizes a multi-agent architecture where two models work in parallel: a "lead" model and a "sidekick" model. The lead model, typically a frontier model, handles high-level planning, interpretation, and review, while the sidekick model, a more cost-effective model like SWE-2, handles task execution. Instead of passing entire conversation histories, the models exchange only briefs, results, and feedback, which allows them to maximize the use of prompt caching.
According to benchmarks conducted by Cognition in partnership with Artificial Analysis and Vals AI, the Fusion harness provides significant savings. For instance, when using Claude Fable 5.1 with the SWE-2 sidekick, the cost was 36% lower compared to using Claude Fable 5.1 with Claude Code, while maintaining comparable performance. In cases involving GPT-6 Astra, Fusion achieved results equivalent to Codex while reducing costs by up to 39%.
Sources: