Build the model-proof stack — the Agent OS inside AIPB
Join Now
Breaking — OpenRouter Fusion API · June 2026
I.The headline — cost & performance vs Fable

It beats Fable 5. At half the price.

OpenRouter just shipped the most important AI release of the week. One API call sends your prompt to a panel of up to 8 models at once — they answer in parallel with live web search, and a judge fuses the best of all of them into one reply. On the DRACO deep-research benchmark it scored higher than Fable 5 — for roughly half the cost. Here's the proof, in charts.

0%
DRACO score
beats Fable's 65.3%
0
of Fable's cost
0
models in parallel
Opus Gemini Grok Fable +more JUDGE fuses one answer
II.The only numbers that matter

Higher score. Lower bill.

This is the whole story in three charts. On deep-research tasks, the premium Fusion panel out-scores Fable 5 outright — and even the cheap "budget" panel lands within a point of it, for about half the spend.

DRACO deep-research benchmark
Higher is better. Fusion (premium) wins at 69%. Axis starts at 60 to show the gap. Hover a bar.
72
69
66
63
60
✦ winner
69%
Fusion
premium
69.0%
65.3%
Fable 5
solo
65.3%
64.7%
Fusion
budget
64.7%
The premium panel beats Fable 5 (+3.7 pts). The budget panel ties it within ~1 point — at roughly half the price.
Cost to run — same job
Lower is better. Fusion does it for about half.
Solo Fable 5100%
Fusion panel~50%
Same deep-research quality — better, actually — for half the bill.
The sweet spot
Top-left is the dream: high performance, low cost. Only Fusion sits there. Hover the dots.
best
quadrant
Performance → Cost → cheap expensive
Fusion premium69% · ½ cost
Fusion budget64.7% · cheapest
Fable 565.3% · full price
Single modelspricier · lower
Everyone else makes you choose: cheap or good. Fusion is the only dot in the top-left.

That's the pitch in the founders' own words.

Alex Atallah @alexatallah · CEO, OpenRouter

"Fable-level performance on deep research tasks, at half the cost. Better-than-SOTA performance using panels. The future of AI is neurodiversity, not single-model takeovers."

jacky @jjacky

"we combined a panel of models and came within 1% of fable 5's perf at half the cost. 👉 simply 'model': 'openrouter/fusion'"

III.How it pulls that off

A boardroom of models, with a chairman.

One request does four things behind the scenes. You never see the mess — just the verdict.

The Fusion loop
  • Fan out. Your prompt hits up to 8 models at once — Opus 4.8, Gemini 3, Grok, Fable and more.
  • Search live. Every panel member runs with web search on, so answers are current.
  • Judge. A judge model reads them all and maps the consensus, contradictions and blind spots.
  • Fuse. It writes one clean answer — the disagreement already resolved. That's why the panel beats any single member.

Two dials let you tune it: analysis_models swaps who's on the panel, and model swaps the judge. Quality preset by default; flip to Budget for the cheaper mix.

IV.Wiring it into your stack

Three ways to plug it in.

It's OpenAI- and Anthropic-compatible, so it drops into anything. One slug: openrouter/fusion.

1 · Hermes profile

~/.hermes/profiles/fusion/config.yaml
model:
  default: openrouter/fusion
  provider: openrouter

2 · Inside Free Claude Code

Fusion has an Anthropic endpoint — so a panel deliberates inside Claude Code. Every plan gets 8 brains.

point Claude Code at Fusion
export ANTHROPIC_BASE_URL="https://openrouter.ai/api/v1"
export ANTHROPIC_API_KEY="$OPENROUTER_API_KEY"

3 · Its own Agent OS section

✅ Already live in your Agent OS
  • A "Fusion" section in the sidebar — its own page, like GLM and Kimi.
  • A Boardroom chat with 6 preset workflows, streaming the verdict in.
  • A "Fusion ✦" workspace bucket for everything it builds.
  • Server-side key — uses your existing OpenRouter key, nothing to paste.
V.Real output, not a demo

I asked the board one real question.

I gave Fusion a genuinely useful prompt: research this week's AI news and give me the 5 highest-leverage videos to film. It searched the web, deliberated, and the judge wrote this — word for word.

Generated by openrouter/fusion · judged by Opus 4.8

The situation in one breath: Two top Claude models got switched off by the government overnight, an open-weight model matched the frontier for ~$18/month, and a panel of cheap models beat the frontier on quality and price. One signal: single-model dependency is now a business risk — the people who route around it win.

1. "Claude Got Banned. Fix Your Agents Now"48 chars

Why now: the highest-panic, highest-search moment of the week. Be the answer before the algorithm cools.

2. "I Replaced Claude With an $18 Model"40 chars

Why now: the counter-punch — GLM 5.2, MIT-licensed, drop-in. Show a real rebuild.

3. "Beat the Frontier for Half the Price"38 chars

Why now: the "I can't unsee this" video — Fusion, explained with a live run.

4. "Build a Ban-Proof AI Agent"24 chars

Why now: the strategic flagship — a router, fallback chain, cost ceiling, one-click swap. Evergreen.

5. "1M Tokens Changes Your Agents"35 chars

Why now: the high-retention deep-dive — run a genuine long-context test, not a hype reel.

The one theme: Stop renting your business from a single AI. Become the router.

It even flagged its own weak claim: GLM's 1M-context isn't independently benchmarked yet — demonstrate it with your own test.

judge: claude-4.8-opus · tokens 7,465 in / 2,009 out · cost $0.46 · web search on

Read that last line again — the panel caught its own weak claim. That's blind-spot detection. One model rarely tells you where it's wrong.

VI.What it's actually for

The workflows worth paying for.

It runs a whole panel, so it costs more per call than one model. Use it where being wrong is expensive.

i.

Never-blocked fallback

A model gets banned or goes down — Fusion routes around it. Work never stops.

ii.

SEO content council

Live-searched brief that reconciles what ranks — plus the angle competitors missed.

iii.

Title + thumbnail brain

Each model proposes hooks; the judge ranks them and says why. A vote, not one taste.

iv.

Fact-check before camera

Consensus-vs-contradiction tells you if the models agree or it's shaky.

v.

Deep research

Web search baked in, sources cross-checked, blind spots flagged. Built for DRACO depth.

vi.

Red-team my offer

The judge's blind-spot output is a free senior reviewer before you launch.

💰 The cost reality
  • You pay for the whole panel + judge — a few cents to ~50¢ per serious question.
  • Not for bulk. GLM 5.2 and Kimi stay your workhorses for volume and code.
  • Use it for "this matters" — research, strategy, fact-checks, the ban fallback.
VII.The lesson under all of it

Stop renting from one model. Become the router.

Same thread as the Claude ban and the GLM 5.2 drop. The model layer churns weekly now. Fusion is that whole idea in one API call: don't bet on one brain, orchestrate many.

The old way
one model
  • Wire everything to the "best" model
  • One provider, one point of failure
  • It gets banned, gated or rate-limited
  • You trust one model's blind spots
  • Work stops when it does
The new way
a panel
  • Send the question to many at once
  • A judge resolves the disagreement
  • One goes down — the panel routes on
  • Blind spots caught by other models
  • Work keeps running, sharper + cheaper
Claude, GLM, Kimi and Fusion connected into one Agent Operating System
The Agent Operating System

I wired Fusion in over an afternoon. You can run the same stack.

New model drops, new router launches — you bolt it on and your stack gets stronger instead of breaking. Inside the AI Profit Boardroom you get the full system: Claude, OpenClaw, GLM, Kimi and Fusion in one dashboard, one memory, agents that know your business.

What you get when you join
  • The full Agent OS zip — every profile, prompt and config
  • The Fusion Boardroom + the model-routing setup from this guide
  • The Obsidian memory setup so your AI knows your business
  • Weekly coaching calls where I wire it up with you live
  • 3,100 founders building alongside you — someone's online 24/7
Get the Agent OS →link in the description
VIII.Read it yourself

Every source, first-hand.

IX.The whole thing in 6 tiles

Recap — what you now know.

i.

The headline

Fusion beats Fable 5 on DRACO (69% vs 65.3%) — for ~half the cost.

ii.

Even the budget panel

Lands within 1 point of Fable (64.7%) at half the spend.

iii.

What it is

Up to 8 models in parallel + a judge that fuses one answer. Slug: openrouter/fusion.

iv.

Plugs in anywhere

OpenAI + Anthropic compatible. Hermes, Free Claude Code, its own Agent OS section.

v.

When to use it

High-stakes only — research, strategy, fact-checks, fallback. Workhorses stay GLM + Kimi.

vi.

The lesson

Don't rent your business from one model. Become the router.

Better answers. Half the cost. One API call.
The future of AI isn't one genius model — it's a panel that catches its own mistakes.

Want the whole stack, wired and ready?

Claude, GLM, Kimi, OpenClaw and Fusion — one dashboard, one memory, agents that know your business.

  • The full Agent OS zip — every profile, prompt and config
  • The Fusion Boardroom + model-router from this guide
  • Weekly coaching calls — we set it up together
  • 3,100 founders already building this way

See the 158 pages of member wins →

Get the Agent OS →
Inside the AI Profit Boardroom · aiprofitboardroom.com

Built for operators · used in 38 countries