OpenRouter’s Fusion offers Fable-level results at half the cost

OpenRouter launched Fusion on June 12. The server-side API sends a single prompt to a configured panel of models, compares their outputs and returns one synthesized, grounded response.

Fusion dispatches prompts to each model in the panel in parallel. By default each model has access to web search and bash tools. A judge model extracts consensus points, contradictions and blind spots from the individual outputs. A synthesizer-Claude Opus 4.8 by default-then generates the final answer anchored to that analysis. The service runs on OpenRouter infrastructure; developers can call it by switching a model string to “openrouter/fusion”, adding a fusion tool for selective calls, or building a custom panel in a chatroom without code.

OpenRouter tested Fusion on Perplexity’s DRACO benchmark, a set of deep-research user requests. A panel that paired Anthropic’s Fable 5 with OpenAI’s GPT-5.5 and used Opus for synthesis scored 69% on the benchmark. Solo Fable scored 65.3%; seven of Fable’s 100 tasks were blocked by its content filters. A lower-cost panel combining Google’s Gemini 3 Flash with two open-weight Chinese models, Kimi K2.6 and DeepSeek V4 Pro, fused and synthesized by Opus, scored 64.7% and cost roughly half as much as the Fable configuration. For comparison, solo GPT-5.5 scored 60% and solo Opus 4.8 scored 58.8%.

OpenRouter reported that much of the performance gain came from the synthesizer. Pairing Opus 4.8 with a separate instance of itself produced a 65.5% score, a 6.7-point increase over a solo Opus run; OpenRouter attributes about three-quarters of that lift to the synthesizer and the rest to model diversity. Early test runs also showed a contamination risk: live web access allowed models to surface DRACO’s grading rubric in search results. OpenRouter excluded the benchmark’s hosting domains from the search tools and published cleaned results.

The launch followed a U.S. export-control directive that prompted Anthropic to suspend Fable 5 and Mythos 5 for foreign nationals. OpenRouter announced Fusion on social media the day after the suspension, promoting the product with the claim “Fusion achieves Fable-level intelligence at half the price.” Because Fusion runs models through OpenRouter’s infrastructure, it does not restore direct access to Fable for users blocked by export controls.

Reactions among researchers and engineers were mixed. AI researcher Andrew Trask wrote on social media that combining models could reduce the advantage of a single frontier model. Other observers flagged gaps in coding performance, inconsistent tool calling and the difficulty of comparing results now that Fable 5 is not widely available for side-by-side testing.

OpenRouter and third-party testers described limits. Fusion is presented as a way to cross-check responses when one model might miss information, not as a full replacement for the highest-end models on every workload. Testers reported that Fable leads on long-range planning and the most difficult reasoning tasks. For coding, Fusion is designed to be invoked as a tool by a coding model rather than to replace a full coding agent loop. Users who cannot access Fable now can choose Fusion, backend swaps that route requests to alternative models, or open-weight models such as GLM-5.2.