Summary

  • OpenRouter introduced Fusion on June 12, a server-side API that queries a group of models, then employs a judge and synthesizer to create the best output.
  • This budget-friendly panel of various AIs performed within 1% of Fable 5 on Perplexity's DRACO benchmark, at approximately half the expense.
  • The launch coincided with a U.S. export control directive that led Anthropic to halt Fable 5 and Mythos 5.

OpenRouter has unveiled an API based on a straightforward premise: that a cost-effective assembly of AI models, when properly combined, can rival a single high-end model, specifically Claude Fable 5.

The new offering is named Fusion. It sends a prompt to several models simultaneously, utilizing a judge model and a synthesizer to consolidate the responses into a coherent answer.

The timing is particularly advantageous. After the release of Fable 5 and Mythos 5 last week, a U.S. export control directive compelled Anthropic to suspend these models for international users, citing issues with jailbreak findings. OpenRouter took to X the following day, advertising "Fable-level intelligence at half the cost."

Introducing the Fusion API, the leading compound model available.

Fusion delivers Fable-level intelligence at half the cost.

How it works 👇 pic.twitter.com/OTUQAdTQjU

— OpenRouter (@OpenRouter) June 13, 2026

Accessing Affordable Fable-Level Performance

A judge model then identifies consensus points, contradictions, and gaps in the responses. Following this, a synthesizer—typically Claude Opus 4.8—compiles the final answer based on this evaluation.

This entire process occurs server-side. Users can switch their model string to "openrouter/fusion" for a default setup, incorporate a fusion tool for selective model calls, or create a custom panel in the Fusion chatroom without coding.

OpenRouter assessed this using DRACO, Perplexity’s benchmark derived from actual user deep research requests. The combination of Fable 5 and OpenAI’s GPT-5.5, synthesized by Opus, achieved a score of 69%. In contrast, solo Fable scored 65.3%, although seven of its 100 tasks were blocked due to its content filters.

The more economical combination is what OpenRouter wants to highlight: The affordable Gemini 3 Flash paired with the open-source Chinese models Kimi K2.6 and DeepSeek V4 Pro, fused and synthesized by Opus, scored 64.7%—outperforming solo GPT-5.5 (60%) and solo Opus 4.8 (58.8%) while costing about half.

Even pairing Opus 4.8 with another instance of itself yielded a score of 65.5%, a 6.7-point increase over solo Opus; OpenRouter claims that approximately 75% of this improvement arises from the synthesis phase, with the remainder resulting from true model diversity.

However, a potential issue arises from granting the panel live web access, which allows models to inadvertently surface DRACO's grading rubric in search results, a risk OpenRouter attributes to coincidence rather than intent. They resolved this by excluding the benchmark's hosting domains from search tools, ensuring all published numbers reflect this adjusted run.

Is It Worth Trying?

OpenRouter makes it clear that Fusion is not a complete substitute for Fable. DRACO does not account for long-horizon tasks, where Fable is believed to excel, and Fusion serves as a tool that a coding model can call selectively rather than a direct replacement—echoing findings from Decrypt regarding DeepClaude, a more affordable backend alternative that maintains Claude Code's agent loop while still lagging behind Opus on challenging reasoning tasks.

The standard model continues to manage everyday tasks, while Fusion is designed for scenarios where a single model might overlook crucial details, making the cross-checking of multiple perspectives beneficial.

For in-depth research, intricate planning, or situations where contradictions are important, the collaborative approach appears to enhance results.

The data clearly indicates that for such tasks, the costly standalone model is no longer the sole option for effective synthesis. A combined set of models that are still accessible can achieve comparable results while significantly reducing costs.

The launch generated a roughly two-to-one positive sentiment ratio. AI researcher Andrew Trask described it as "a way bigger deal than it seems," arguing that frontier labs will no longer monopolize cutting-edge technology. However, some skeptics challenged this view, pointing to unsatisfactory coding results, poor tool utilization, and a lack of transparency since Fable 5 is no longer available for comparison.