June 12 was a double event in the AI industry. The U.S. Department of Commerce issued an export-control directive: Anthropic's Fable 5 and Mythos 5 were blocked for all users outside the United States. Anthropic pulled the models the same day the directive landed.
OpenRouter launched Fusion that same day.
Fusion isn't a new model in the traditional sense. It's a compound API that routes a single prompt to multiple models at once — Google's Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro. A judge model then analyzes all the responses, spotting consensus and contradiction. A synthesizer — defaulting to Claude Opus 4.8 — turns that analysis into one coherent answer. The technique is called Compound AI.
On the DRACO benchmark, Fusion scored 64.7% against Fable 5's 65.3%, a gap of just 0.6 percentage points. GPT-5.5 topped out at 60% on the same scale; solo Opus 4.8 reached 58.8%. A panel of cheaper models, in other words, beat every flagship in the market except the one that just got banned.
The price runs at roughly half what Fable 5 cost. Access works through any OpenAI-compatible client — just set the model to "openrouter/fusion."
OpenRouter is upfront that Fusion isn't a full Fable replacement. Long-horizon planning and complex agent workflows still favor Fable based on current data. For typical everyday inference, though, the performance gap shrinks to near-zero.
The speed of the market's response signals something broader. Export controls pulled a leading model off the table; within hours, an alternative was assembled from components that remain legal everywhere. Frontier AI is no longer contained inside a single company's infrastructure — and that changes the leverage regulators thought they had.
