Headline: OpenRouter’s new “Fusion” bundles cheap AI models to mimic Claude Fable 5 — just as Fable goes offline for many users
OpenRouter this week launched Fusion, an API that stitches together multiple inexpensive models into a single “compound” model — and pitches it as a way to get near-Fable performance at a fraction of the cost. The timing is notable: Anthropic suspended access to its newly released Fable 5 and Mythos 5 for foreign nationals after a U.S. export-control directive, creating an immediate vacuum that Fusion aims to fill.
How Fusion works
- Parallel calls: A prompt is sent in parallel to a panel of models. Each model can use web search and bash tools.
- Judge stage: A dedicated judge model scans responses for consensus, contradictions and blind spots.
- Synthesis: A synthesizer (Claude Opus 4.8 by default) composes the final, grounded answer from the judge’s analysis.
- Deployment: The whole pipeline runs server-side. Developers can use the default panel by switching their model string to "openrouter/fusion," add a fusion tool to call Fusion selectively, or build a custom panel in a no-code Fusion chatroom.
Benchmark results (DRACO, Perplexity)
- Top lineup: Fable 5 + GPT-5.5, synthesized by Opus, scored 69%.
- Solo Fable 5: 65.3% (seven tasks blocked by Fable’s content filters).
- Cost-conscious winner: Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro, fused and synthesized by Opus, scored 64.7% — within a point of solo Fable but at roughly half the cost.
- Other comparators: Solo GPT-5.5 (60%), solo Opus 4.8 (58.8%); pairing Opus 4.8 with itself reached 65.5% — OpenRouter attributes ~75% of that lift to synthesis and the remainder to model diversity.
A quick correction: giving panels live web access briefly let models surface DRACO’s own grading rubric during runs, a contamination risk OpenRouter fixed by excluding the benchmark’s host domains from search tools; published numbers reflect the cleaned runs.
Limits and trade-offs
- Not a full Fable replacement: OpenRouter concedes Fusion lags on long-horizon tasks and harder reasoning work where Fable reportedly excels. For coding, Fusion is intended as a tool called selectively by a coding model, not an outright swap.
- Practical role: Fusion is positioned as a safety net — a second opinion layer for queries where a single model may miss important contradictions or facts. OpenRouter says the approach shines on deep research, complex planning, and tasks where cross-checking matters.
- Infrastructure and export controls: Fusion runs on models routed through OpenRouter’s infrastructure, so it doesn’t bypass the underlying export-control or access restrictions that sidelined Fable 5.
Market reaction and alternatives
- Community split: Sentiment on the launch thread tracked roughly two-to-one positive. AI researcher Andrew Trask called Fusion “a way bigger deal than it seems,” suggesting frontier labs won’t uniquely own the lead. Critics flagged weaker coding outputs, tooling gaps, and limited transparency now that Fable 5 is unavailable for side-by-side testing.
- Options for blocked users: OpenRouter positions Fusion as one option; others include backend swaps like DeepClaude or open-weight models such as GLM-5.2, which may be cheaper though not necessarily better at the toughest tasks.
Why crypto and Web3 builders should care
- Cost efficiency: For teams running on tight budgets — common in crypto and Web3 startups — Fusion’s promise of near-top-tier performance at roughly half the price could materially lower AI infrastructure costs.
- Redundancy & resilience: Multi-model synthesis reduces dependence on any single proprietary model, which matters if regulatory moves or vendor policy changes interrupt access.
- Developer flexibility: APIs that let you swap panels or call a fusion tool selectively can fit into agent workflows, smart-contract tooling, and research pipelines where accuracy and cost both matter.
Bottom line: Fusion demonstrates that intelligently combining several cheaper models plus a judge-and-synthesizer pipeline can close much of the gap with a single high-end model. It’s not a drop-in replacement for Fable 5 across every use case, but for many research and production workloads — and particularly for cost-conscious crypto projects — it’s a pragmatic alternative worth testing.
Read more AI-generated news on: undefined/news
