Key facts
- OpenRouter launched its Fusion API on June 12, a server-side tool that combines multiple AI models.
- Fusion aims to match the performance of expensive models like Claude Fable 5 at a significantly lower cost.
- On Perplexity's DRACO benchmark, a Fusion panel achieved 64.7% performance, close to Fable 5's 65.3%, at roughly half the price.
- Anthropic's Fable 5 and Mythos 5 were suspended for international users due to a U.S. export control directive.
- Fusion synthesizes responses from a panel of models, using a judge model for consensus and a synthesizer for the final answer.
OpenRouter has introduced Fusion, a new server-side API designed to leverage multiple AI models in parallel to achieve performance comparable to high-end, single models like Anthropic's Claude Fable 5, but at a reduced cost. The launch coincides with U.S. export control directives that led Anthropic to suspend Fable 5 and Mythos 5 for international users.
Fusion works by sending a user's prompt to a panel of AI models simultaneously. A 'judge' model then analyzes these responses to identify consensus points, contradictions, and gaps. Finally, a 'synthesizer' model, by default Claude Opus 4.8, crafts a unified and grounded answer based on this analysis.
In testing on Perplexity's DRACO benchmark, a budget panel comprising Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro, synthesized by Opus, achieved a score of 64.7%. This performance is within 1% of Fable 5's score of 65.3% and comes at approximately half the cost. This Fusion panel outperformed solo models like GPT-5.5 (60%) and Opus 4.8 (58.8%). Even pairing Opus 4.8 with itself resulted in a score of 65.5%, with OpenRouter attributing most of this gain to the synthesis step and the diversity of the models.
OpenRouter acknowledges that Fusion is not a direct replacement for Fable 5, particularly for complex tasks like coding or long-horizon work where Fable reportedly still excels. The company suggests Fusion is best suited for questions where a single model might miss crucial details, and cross-checking multiple perspectives can yield better results. The API can be accessed by swapping the model string to 'openrouter/fusion', using a fusion tool, or building custom panels.
Separately, users have noted significant improvements in ChatGPT, leading to speculation that OpenAI is quietly testing GPT-5.6. AI researcher Andrew Trask viewed Fusion as a major development, suggesting that leading AI labs may no longer have exclusive access to frontier capabilities. However, some critics raised concerns about Fusion's coding performance, tool-calling abilities, and a lack of transparency due to Fable 5's unavailability for direct comparison.