Headline: OpenRouter’s new “Fusion” stacks cheap models to mimic Claude Fable 5 — just as Fable goes offline for many users
OpenRouter this week unveiled Fusion, an API that bets you can match a top-tier model by combining multiple cheaper ones and stitching their outputs together — at a fraction of the cost. The launch landed at an awkwardly perfect moment: Anthropic’s Fable 5 and Mythos 5 were suspended for foreign nationals after a U.S. export-control directive, opening a hole in the market that OpenRouter was quick to target with the claim of “Fable-level intelligence at half the price.”
How Fusion works
- A single prompt is broadcast in parallel to a panel of models. Each model can use web search and bash tools.
- A judge model analyzes the panel responses to extract consensus points, contradictions, and blind spots.
- A synthesizer (Claude Opus 4.8 by default) then writes the final, grounded answer based on the judge’s analysis.
- All of this runs server-side through OpenRouter. Users can call the default panel by setting their model string to openrouter/fusion, add a fusion tool so their own model can call Fusion selectively, or build custom panels in a no-code Fusion chatroom.
Benchmarks and pricing tradeoffs
OpenRouter tested Fusion on DRACO, Perplexity’s benchmark drawn from real user deep-research tasks. Highlights:
- A panel of Fable 5 + OpenAI GPT-5.5, synthesized by Opus, scored 69%.
- Solo Fable scored 65.3% (seven tasks were blocked by Fable’s filters).
- The “cheap” panel OpenRouter emphasizes — Google’s Gemini 3 Flash plus open-weight Chinese models Kimi K2.6 and DeepSeek V4 Pro, fused and synthesized by Opus — hit 64.7%. That beats solo GPT-5.5 (60%) and solo Opus 4.8 (58.8%), and sits within a percentage point of Fable, at roughly half the cost.
- Pairing Opus 4.8 with a separate Opus instance scored 65.5%, a 6.7-point lift vs. solo Opus. OpenRouter attributes about 75% of the improvement to the synthesis step itself and the rest to model diversity.
Quality control and limits
- One issue found: when models had live web access, they could surface DRACO’s grading rubric in search results, contaminating the benchmark. OpenRouter fixed that with a one-line config to exclude the benchmark’s hosting domains; published results reflect the cleaned runs.
- OpenRouter is candid that Fusion isn’t a drop-in Fable replacement. DRACO doesn’t evaluate long-horizon tasks where Fable reportedly still leads. For coding, Fusion is positioned as a tool a coding agent calls selectively rather than a wholesale replacement for a coding model — a limitation others testing cheaper Claude-compatible backends (e.g., DeepClaude) have also observed.
- Fusion runs entirely on models routed through OpenRouter’s infrastructure, so it doesn’t circumvent the export-control problem at its source.
Community reaction and implications for crypto developers
Reactions to the launch skewed positive (roughly 2:1 in OpenRouter’s tracking). AI researcher Andrew Trask called Fusion “a way bigger deal than it seems,” saying frontier labs won’t automatically own the frontier anymore. Skeptics noted poorer coding/tool-calling outcomes in some cases and warned that the suspension of Fable 5 makes apples-to-apples public comparisons harder.
For crypto projects and global dev teams, Fusion matters for a few reasons:
- Cost: if you need near–top-tier reasoning but want to cut bills, model composition can get you close at substantially lower cost.
- Availability: teams locked out of Fable 5 now have multiple fallbacks — Fusion panels, backend swaps like DeepClaude, or open-weight models (GLM-5.2 and others).
- Centralization risk: Fusion still relies on OpenRouter’s routing and infra, so it’s not a full decentralization answer to export controls or model monopolies.
Bottom line
Fusion is a timely demonstration that “many middling-but-cheap models + a good judge + a good synthesizer” can approach the performance of an expensive single model — and at a lower price. It won’t instantly replace Fable for every high-end task, but for teams that prioritize cost and resilience over absolute top-end reasoning today, Fusion is a compelling new option.
Read more AI-generated news on: undefined/news
