OpenRouter’s Fusion API is a new product technology that routes a single prompt in parallel to a panel of AI models, then uses a judge model to extract consensus and a synthesizer to produce a final answer. DRACO benchmark results show Fable 5 paired with GPT-5.5 and synthesized by Opus reached 69%, solo Fable scored 65.3% (seven of 100 tasks blocked by content filters), and a cheaper Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro panel synthesized by Opus scored 64.7%.
Opus 4.8 run in a separate instance scored 65.5%, representing a 6.7-point improvement over solo Opus. OpenRouter attributed roughly three-quarters of the system’s performance lift to the synthesis step and attributed the remainder to genuine model diversity. The developers used a configuration line to exclude the benchmark hosting domains from the system’s search tools to address contamination concerns. The report also notes timing around Fable 5’s release and Anthropic’s export-control suspension for foreign nationals.
Fusion API runs server-side and is accessible via a default panel using the model string ‘openrouter/fusion’ or through a customizable Fusion chatroom that requires no code.
The description above presents the Fusion API’s multi-model routing, judging, and synthesis workflow alongside DRACO benchmark scores. The overview is limited to the product’s architecture and benchmarked results as reported.
The Fusion API sends a single prompt in parallel to a panel of models, collects each model’s response, and applies a judge model to extract consensus before the responses are merged. A separate synthesizer then produces the final answer by combining the judged outputs into a unified response. The workflow explicitly separates the consensus step from synthesis, with the judge model responsible for extracting agreement or key outputs and the synthesizer responsible for producing the merged response text. The API routes and orchestration occur server-side as part of the Fusion runtime.
OpenRouter reports that roughly three-quarters of the system’s measured performance lift comes from the synthesis step, with the remaining improvement attributed to genuine model diversity. The Fusion API can be invoked using a default panel via the model string ‘openrouter/fusion’, or it can be deployed through a customizable Fusion chatroom that requires no code. The customizable chatroom provides a way to change the panel configuration without writing integration code. The Fusion runtime and configuration options support switching between the default panel and bespoke panels at deployment time.
The preceding paragraphs summarize the Fusion API’s operational mechanism and configurable deployment options as reported. These descriptions reflect the details provided in the source material.
The DRACO benchmark reported that a Fusion configuration using Fable 5 paired with GPT-5.5 and synthesized by Opus achieved 69%. Solo Fable achieved 65.3%, with seven of 100 tasks blocked by content filters. A lower-cost panel composed of Gemini 3 Flash, Kimi K2.6 and DeepSeek V4 Pro, fused and synthesized by Opus, scored 64.7%. These benchmark results were presented as direct comparisons among individual and combined model configurations.
The 64.7% score exceeded solo GPT-5.5, which scored 60%, and solo Opus 4.8, which scored 58.8%, and was reported to be within a point of Fable at roughly half the cost. Opus 4.8 run in a separate instance scored 65.5%, a 6.7-point improvement over solo Opus. OpenRouter reported that roughly three-quarters of the system’s measured performance lift came from the synthesis step, with the remainder attributed to genuine model diversity. The comparisons included both high-end paired configurations and lower-cost fused panels.
The DRACO benchmark results summarize relative accuracy and cost-efficiency across the tested configurations. The reported figures distinguish combined-panel fusion outcomes from solo-model baselines.
A configuration line was used to exclude the benchmark hosting domains from the system’s search tools, a step included in the report to address contamination concerns. The report also notes timing around Fable 5’s release and Anthropic’s export-control suspension for foreign nationals. These operational details are presented alongside the benchmark results in the source material.
“Introducing the Fusion API, the smartest compound model in the market.”
“Fable-level intelligence at half the price.”
These contextual and operational notes accompany the benchmark findings in the report. They provide background on the testing environment and the report’s presentation.
