Fusion API targets Fable 5 research quality at half the cost

A cheaper path to deep-research answers

OpenRouter is testing whether the next jump in research agents comes from model orchestration rather than a single larger model. In a June 13, 2026 X thread, the company wrote, “Fusion achieves Fable-level intelligence at half the price.” The source tweet is available on X.

The benchmark claim centers on DRACO, Perplexity’s deep-research benchmark with 100 hard tasks across 10 domains, including law, medicine, finance, and product comparison. OpenRouter says panels of models consistently outperformed individual models, and that a budget panel landed within 1% of Claude Fable 5 while costing roughly half as much. That is a concrete pricing challenge for frontier-model vendors if the result holds up outside OpenRouter’s own tests.

Fusion works by sending a prompt to multiple models in parallel, with web search and bash tools enabled. A judge model then reads the answers and extracts structure: consensus, contradictions, partial coverage, unique insights, and blind spots. A synthesizer writes the final response from that analysis. OpenRouter says roughly three quarters of Fusion’s lift came from synthesis, with the remaining quarter coming from model diversity.

The developer interface is designed to look like an ordinary model call. Teams can call openrouter/fusion as a server-side model slug, or add {"type":"openrouter:fusion"} to a tools array so another model can decide when to invoke it. The thread also points to OpenRouter’s Fusion page and its API documentation.

One detail matters for credibility: OpenRouter says it discovered that models with web search could surface the DRACO rubric online, then excluded those domains and reran the numbers. The next thing to watch is independent replication, especially on private enterprise research tasks where rubric leakage is less likely and latency, cost, and citation quality are easier to measure.