What OpenRouter’s Fusion API Is and Why It Matters
OpenRouter’s Fusion API is a compound AI service that routes a single prompt through several underlying language models, compares their outputs with a judge model, and synthesizes a single higher-quality answer, giving developers Fable-like performance while using cheaper building blocks. This approach emerged as Anthropic’s Fable and Mythos models were pulled from broad access after an export control directive, leaving many teams without the high-end models they had started to depend on. Anthropic’s Fable 5 had quickly set a new bar for capability, scoring 80% on the SWE-Bench Pro coding benchmark and outperforming Anthropic Opus 4.8, OpenAI GPT 5.5, and Google Gemini 3.1 Pro. At the same time, Fable 5 carried a premium price of USD 10 (approx. RM46) per million input tokens and USD 50 (approx. RM230) per million output tokens, and has now been withdrawn for most users.
How the Fusion API Works: Multi-Model, Judge, and Synthesizer
Fusion is an OpenRouter multi-model system that fans each request out to a panel of AI models in parallel, with web search and bash tools switched on. A dedicated judge model then reads every response, maps agreement and disagreement, and notes what each model covers or misses. A synthesizer model writes the final answer based on that map, instead of simply picking a winner. OpenRouter says roughly three-quarters of Fusion’s performance gain comes from this synthesis step, with the remaining quarter coming from the diversity of participating models. Developers can treat Fusion as a single model slug, openrouter/fusion, or expose it as a tool so an upstream model decides when to invoke it. Panel composition is flexible: teams can plug in their own preferred participants and even swap the synthesizer, turning Fusion into an experiment platform for AI model cost performance and behavior.
Benchmark Results: Fable-Level Output at Frontier AI Cheaper Rates
OpenRouter benchmarked Fusion on DRACO, Perplexity’s deep research suite of 100 tasks across domains such as law, medicine, finance, and product comparisons. Each task uses around 39 weighted criteria, with penalties for wrong answers so vague rambling does not help scores. In the top-end configuration, combining Fable 5 with GPT-5.5 through Fusion reached about 69%, slightly ahead of Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro and of Opus 4.8 + GPT-5.5 alone, while solo Fable 5 scored about 65%. The more important result for Fusion API pricing is a budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro. Fused together, these Fable alternative models outperformed solo GPT-5.5 and solo Claude Opus 4.8, landing within 1% of Fable 5’s DRACO score at roughly half of Fable 5’s cost, according to OpenRouter’s internal testing.
From Anthropic’s Withdrawals to Composite Model Strategies
Anthropic’s decision to withdraw Fable and Mythos for most users, following an export control directive and concerns over jailbreak risks, left many teams searching for frontier AI cheaper options. Fable 5 had demonstrated impressive performance on coding, tool use, and long-horizon tasks, and Anthropic had positioned it as a flagship model with strong safeguards, including routing some sensitive queries back to Opus 4.8. When that capability disappeared for global users, OpenRouter’s Fusion became a practical workaround: instead of relying on a single high-end vendor model, developers can mix several mid-priced or "below frontier" options and still approach Fable-class output. This shift points to a broader market trend. Rather than betting everything on one foundation model, more teams are adopting composite model strategies, combining multiple providers and using orchestration layers like Fusion to balance cost, safety, and performance.
