OpenRouter just launched Fusion, a multi-model routing API that fans your prompt out to multiple LLMs simultaneously, synthesizes their responses through a judge model, and returns a single answer. The pitch: frontier-level intelligence at half the price of Claude Fable 5. The Hacker News reality check: 7× slower and 4× the cost of just calling a single top model directly.
📖 Read the full version with charts and embedded sources on ComputeLeap →
So which is it?
The timing is not a coincidence. With Anthropic's Fable 5 freeze still reverberating — the model pulled barely a week ago over export control concerns — operators are scrambling for a single-vendor-risk hedge. OpenRouter is selling exactly that: don't depend on one frontier model when you can blend several. But the economics of multi-model routing are more nuanced than the marketing suggests.
How Fusion Actually Works
Fusion operates in three sequential phases, documented in OpenRouter's plugin guide:
Panel Phase. Your prompt goes out to up to 8 models in parallel. The default Quality preset sends to Fable 5 + GPT-5.5; the Budget preset uses Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro.
Judge Phase. A designated judge model (Claude Opus by default) receives all panel responses and performs comparative analysis — consensus, contradictions, partial coverage, unique insights, and blind spots.
Synthesis Phase. Your primary model crafts the final response from the judge's analysis.
The critical detail: you pay for every underlying completion plus the judge call. A 3-model panel = roughly 4–5× the cost of a single completion.
⚠️ Fusion pricing is cumulative — Quality costs 3.2× what a single Opus 4.8 call costs. Budget is the cost-efficient option at 0.40× of solo Fable 5.
The DRACO Numbers
| Configuration | DRACO Score | Cost per Prompt |
|---|---|---|
| Fusion Quality (Fable 5 + GPT-5.5) | 69.0% | $0.29 |
| Claude Fable 5 (solo) | 65.3% | ~$0.10 |
| Fusion Budget (Flash + Kimi + DeepSeek) | 64.7% | $0.04 |
| GPT-5.5 (solo) | 60.0% | ~$0.06 |
Quality beats solo Fable 5 by 3.7 points. Budget comes within 0.6 points at 40% of the cost.
Caveats: Fable 5 completed only 93/100 tasks (content filters), DRACO is text-only English-only, and Fusion showed "no advantage for long-horizon tasks."
The HN Reality Check: 7× Slower, 4× the Cost
The HN thread (200 pts, 78 comments) tells a sobering story. Top comment from a dev who built a similar system: "Fusion was 7× slower and 4× the cost compared to calling Opus 4.7 directly."
Deeper concern: having one model judge another essentially asks "how closely does this resemble the answer you would have given me." Additional rounds = "cranking up the temperature" without better answers.
💡 HN consensus: multi-model judging works for verifiable answers but poorly for ambiguous domains.
Most interesting finding from a related thread: fusing identical models also boosts performance — suggesting gains come from test-time compute, not model diversity.
Budget vs Quality: Two Very Different Products
TokenMix's review breaks down the annual math at 10K prompts/month:
- Quality Fusion: $34,800/yr (2.9× more than Fable 5 for 3.7 DRACO points)
- Solo Fable 5: $12,000/yr
- Budget Fusion: $4,800/yr (60% less, within 0.6 DRACO points)
Budget Fusion is a genuine cost play. Quality Fusion only pencils out for high-stakes domains (legal, compliance, medical).
When to Use (and Skip) Fusion
Use Quality when: output value >$1/task, need cross-model consensus, verifiable answers, 1-3s latency OK.
Use Budget when: high-volume batch, frontier-adjacent on a budget, vendor diversification matters.
Skip entirely for: real-time (<500ms), code completion, chat, long-horizon tasks.
💡 Heuristic: if a skilled human reviewer would consult three experts before answering, Fusion fits. If that's overkill, single-model wins.
The Verdict
Budget Fusion delivers on the "half the price" promise for batch workloads. Quality Fusion costs 3× more and only makes sense for high-value-per-task domains. Use it surgically, not as your default router.
Originally published at ComputeLeap
Top comments (0)