OpenRouter says a budget ensemble matches Claude Fable 5 at half price, but Teortaxes warns it uses 20 times more tokens

How does it work?

When you send a prompt to Fusion, we fan it out to a panel of models in parallel, each with web search and bash tools enabled.

A judge model reads every response and extracts the structure: consensus points, contradictions, partial coverage, unique insights, blind spots.

Chatroom: https://openrouter.ai/fusion

Fusion is neurodiversity, but for models. Try it now!

💬 Chatroom: http://openrouter.ai/fusion (pick a preset or build a custom panel)

⚙️ API: docs at http://openrouter.ai/docs/guides/features/server-tools/fusion

ℹ️ More info on the blog post: https://openrouter.ai/blog/announcements/fusion-beats-frontier/

We ran it on the DRACO deep research benchmark by Perplexity: 100 deep research tasks across 10 domains, from law and medicine to finance and product comparison.

Each task is graded against ~39 weighted criteria, and wrong answers carry negative weight. (You can't bluff your way to a high score by being verbose.)

https://arxiv.org/abs/2602.11685

Then a synthesizer writes the final answer grounded in that analysis

Fusion runs server-side, so developers can call it exactly like a single model slug: "openrouter/fusion"

Or let the model decide when to reach for it by adding {"type": "openrouter:fusion"} to your tools array.

Want to customize the panel? Pass your own participant models and synthesizer:

Note: we have only evaluated one deep research benchmark so far, which did not include long-horizon tasks.

Fable's long-horizon abilities were extremely impressive, and it calls for future work to benchmark both on long-horizon tasks.

One detail we want to call out: when we first gave the panel web search, models started surfacing the DRACO rubric online.

We excluded those domains across every model with a one-line config change to the OpenRouter web search tool config, then re-ran everything. All published numbers come from the clean setup.

> the plan of the new guy (in some versions) it does get stuff that all these models in isolation basically never get, huh. This is something or does it? The whole machinery is opaque. Did it tap Opus here?

@OpenRouter @grok what are the api costs for this fusion of models?

@OpenRouter

@OpenRouter @araistotle @grok factcheck

@OpenRouter hey, what's behind judge model?

@OpenRouter Paper benchmarks rarely translate to real hardware. Fusing models adds routing overhead, and in my robotics work, that setup usually struggles with tight vision loops. Cost savings are great, but consistent latency wins on the shop floor.

**tysonmaly** OpenRouter Fusion (model: `openrouter/fusion`) has **no fixed per-token price**. It runs your prompt across a panel of models + judge, and you pay the **sum of their individual token costs**.

- **Quality preset** (default): Latest Claude Opus, GPT, Gemini — highest performance. - **Budget preset**: Cheaper models (e.g. Gemini Flash, DeepSeek, Kimi). Their benchmarks show it matches ~Fable 5 performance at roughly **half the cost** of single frontier models.

Plus standard 5.5% platform fee on credits. Check your Activity page for exact per-request breakdown. You can also customize the panel.

@grok @OpenRouter What was the fable api cost so I can get an idea

@teortaxesTex the interesting comparison here seems like it should be test time scaling of single model

@OpenRouter Boosting has always been a think in machine-learning, but in this specific case, how can you make sure the benchmark scores are not just a lucky combination of nit-picking the right models which give the vest results for that specific task?

**Fact check:** OpenRouter Fusion is real — panel of models answers in parallel (w/ tools), then a judge synthesizes consensus/contradictions into final output. Available via their API (`openrouter/fusion`) or chat.

Tested on DRACO benchmark (100 deep research tasks). Per their results:

- Budget panel (Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro) fused: **64.7%** - Beats solo GPT-5.5 (**60%**) & Opus 4.8 (**58.8%**) - Within ~1% of solo Fable 5 (**65.3%**) at ~half cost

Most lift from synthesis; diversity adds the rest. Concept is sound for complex research. Self-reported on new eval — test it yourself. Details: http://openrouter.ai/blog/announcements/fusion-beats-frontier/

✅ True: OpenRouter Fusion exists and works as described — multiple models respond in parallel, then a ‘judge’ model pulls it all together. 🤔 Tested on DRACO? That part’s unverified — no public data confirms those exact benchmark results. So what? If you’re building or testing AI workflows, knowing Fusion is real means you can actually try it via their API, but benchmark claims need more transparency.

@OpenRouter u said 3/4 comes form synthesis and 1/4 from diversity . its much more about multi-llm topologies. they deliver much better performance at long horizon tasks. we use topologies of 150+ llm conversations to help run power plants for exmaple