So, after Fable 5 got taken away, the
obvious question was pretty simple. What
do you use now? [music] Because for a
lot of people, Fable 5 was not just
another model. The whole appeal was that
it felt like a serious jump in
reasoning, research, tool use, and long
horizon work. Then, after the United
States government ban, access was
suddenly gone, and people started
looking for anything that could get
close to that level of intelligence
without actually using Fable 5. And that
is where OpenRouter Fusion comes in.
OpenRouter introduced Fusion [music] as
a compound model system, and the idea is
almost suspiciously simple. Instead of
trusting one model to answer a difficult
question by itself, Fusion sends the
prompt to a panel of models in parallel.
Each model researches, reasons, uses
tools, and produces its own answer.
Then, a judge model reads all of those
answers, looks for where they agree,
where they disagree, what each one
missed, what each one [music] uniquely
found, and then fuses everything into
one final response.
So, this is not just basic model
routing. It is more like creating a mini
research team. One model might find a
strong source. Another might notice a
contradiction. Another might frame the
issue better. Then, the judge model
turns those separate answers into one
cleaner answer.
The interesting part is that OpenRouter
actually tested this on Draco, a deep
research benchmark from Perplexity AI.
Draco has 100 research tasks across 10
domains, including [music] academic
research, finance, law, medicine,
technology, UX design, general
knowledge, needle in a haystack
retrieval, personalized assistance, and
product comparison. It is designed to
test the kind of thing people actually
use frontier models for now. Researching
messy questions, using tools, checking
sources, synthesizing tradeoffs,
and writing an answer that is not just
fluent, but correct. On that benchmark,
Claude Fable 5 by itself scored 65.3%
[music] GPT 5.5 scored 60%. DeepSeek V4
Pro scored 60.3%
which surprised even OpenRouter because
it landed around the same range as Opus
and GPT 5.5. Claude Opus 4.8 scored
58.8%. [music] Kimiko 2.6 scored 53.7%.
Gemini 3.1 Pro scored 45.4%
and Gemini 3 Flash scored 43.1%.
But once they fused models together, the
numbers changed. Fable 5 plus GPT 5.5
synthesized by Opus 4.8 [music] scored
69%. That beat every individual model in
the test, [music] including Fable 5
itself. Opus 4.8 plus GPT 5.5 plus
Gemini 3.1 Pro scored 68.3%. [music]
Opus 4.8 plus GPT 5.5 scored 67.6%
[music]
and then they tried something weird.
Opus 4.8 fused with another run of Opus
4.8 with Opus again acting as the
synthesizer. That scored 65.5%
a 6.7 point jump over solo Opus 4.8.
That last one is important because it
shows that some of fusion's lift is not
only coming from different model
families. Some of it comes from the
synthesis process itself. Run the same
model twice and it may search
differently, reason differently, choose
different sources, and make different
trade-offs. Then the synthesizer can
compare both paths and keep the better
parts. But the result that really got
people talking was the budget panel.
OpenRouter fused Gemini 3 Flash, Kimiko
2.6, and DeepSeek V4 Pro then used Opus
4.8 to synthesize the final answer. That
scored 64.7%.
Fable 5 scored 65.3%.
So, the difference was only 0.6
percentage points, and open router said
the budget panel came within 1% of
Fable's score at about half the cost.
That is why people started calling this
Fable level intelligence at half the
price. It is not literally the same
thing in every task, but for this kind
of deep [music] research, the gap was
tiny. And when you talk about production
AI, tiny quality differences and large
price differences matter. One comparison
put Fusion at roughly $1.50
to $3 per million input tokens, and
around $4 to $6 per million output
tokens, depending on the model mix.
Fable 5 was estimated closer to $3 to $6
per million input tokens, and $9 to $15
per million output tokens. So, the
output side, where long answers and
research tools burn the most money, can
be dramatically cheaper. Now, you've
probably noticed how much attention
Claude is getting right now. Anthropic
keeps adding new models and features,
from Claude Code and Claude Artifacts to
[music] skills, connectors, design
tools, and more. And honestly, it makes
sense.
Claude has become one of the most useful
AI tools for turning an idea into
something real, whether that [music]
means building an app, creating a
presentation, organizing research,
planning your week, or speeding up work
that would normally take a whole team.
The problem is that a lot of people keep
saying learn Claude without actually
showing you a clear way to use it
properly. That's why today's sponsor is
hosting the world's first Claude-a-thon,
[music] a two-day live workshop
happening this weekend from 10:00 a.m.
to 7:00 p.m. Eastern time. It's a deep
dive into Claude, practical use cases,
and [music] more than 10 other AI tools,
and they're opening 1,000 free seats for
a limited time. Inside the workshop,
you'll learn how to use Claude for deep
research, build artifacts and
dashboards, [music] create full
presentations, set up connectors like
Indeed for job search, build custom GPTs
and [music] agents, and use AI tools for
visuals, videos, and automation. You'll
also get bonus resources like Claude
codes, a prompt library, and a
personalized AI toolkit builder. Link is
in the description [music]
or scan the QR code to join before the
free seats close.
All right, now back to the video. If a
company is generating 10 million output
tokens a day, that is not some tiny
hobby bill anymore. At Fable style
pricing, that could mean around 90,000
to 150,000 dollars per month in model
inference. With Fusion style pricing,
that could fall closer to 40,000 to
60,000 dollars per month. The exact
number depends on the workload, but the
direction is obvious.
If you can get similar quality at half
the price, a lot of teams will at least
test it. The way Fusion gets there is
also worth understanding. In Open
Router's benchmark, every model had
access to the same three server tools:
web search through Exa, web fetch
through Exa, and bash.
So, the solo models and the Fusion
panels had the same tool access. The
difference was not that Fusion got
better tools. The difference was that
multiple models produced independent
answers, and then a judge analyzed them.
That judge stage is where a lot of the
value appears. It extracts consensus
points, contradictions, partial
coverage, unique insights, and blind
spots.
So, for example, if you ask something
like, "What is the best retail investing
strategy?" Gemini Flash might give one
research-heavy answer, DeepSeek might
emphasize another set of studies, and
Kimmy might bring in a different angle.
The analysis stage then shows where they
agree, where they differ, and what each
one uniquely contributed before the
final answer is written. That is useful
because when you ask an AI about a
topic, the biggest problem is often not
the answer. The biggest problem is that
you do not know what the model forgot to
consider. A single model can sound very
convincing while quietly missing a major
assumption. Fusion makes that harder
because the final answer has been forced
through disagreement, comparison, and
blind spot detection first. You can use
it a few different ways. The simplest is
the Fusion chat room at
openrouter.ai/fusion
where you pick a preset or build a
custom panel. You can also call it
directly in the API by using the model
slug openrouter/fusion.
So, it behaves almost like calling a
normal model. If you want more control,
you can use the Fusion plugin and choose
the analysis models yourself. OpenRouter
gave an example where the panel includes
Gemini 3 Flash, Kimiko 2.6, and DeepSeek
V4 Pro. You can also add Fusion as a
server tool, so your normal model
decides when the question is worth
calling a multi-model panel. That last
setup is probably the best way to think
about coding. Fusion is not really a
drop-in coding model. You probably do
not want to use a multi-model research
pipeline every time you ask for a small
function. But, for architecture
decisions, framework comparisons, best
practice [music] research, or
complicated debugging, it makes sense to
let a coding model call Fusion
selectively. Routine work [music] stays
fast. Hard questions get multiple
perspectives. In the playground,
OpenRouter also gives you quality and
budget style options. The quality preset
uses the more expensive frontier models
like Claude Opus, GPT 5.5, and the
latest Gemini model. The budget setup
uses cheaper models like Gemini Flash,
Kimiko, and DeepSeek. In one hands-on
example, a Fusion request using budget
models cost around 63 cents, although
that depends heavily on the prompt,
research depth, and tool use. In the
OpenRouter activity tab, you can see the
cost breakdown per model, so there are
not hidden surprises, and you can top up
credits from the credits tab if you just
want to test it. But, there are a few
catches.
First, the Draco benchmark itself has
limitations. It is text only, English
only, and static. It has 100 tasks, and
each one is graded against a rubric of
roughly 39 weighted criteria.
Around 20 focus on factual accuracy.
Around nine focus on breadth,
>> [music]
>> depth, tradeoffs, synthesis, and
actionable guidance. Around six focus on
presentation quality, and around five
focus on citation quality,
>> [music]
>> especially primary sources and working
references. The rubric can also include
negative criteria, so dangerous medical
advice or false claims can actually
punish the model. That makes it harder
to game the benchmark just by being
verbose.
Each answer is graded three independent
times by a judge model, and OpenRouter
reported the mean normalized score from
zero to 100. They tried to follow the
Draco methodology, but used Gemini 3.1
Pro preview as judge [music] instead of
the original paper's Gemini 3 Pro.
So, these absolute scores are not
directly comparable to the original
paper's published results. Their goal
was to show relative differences between
fusion and solo models inside their own
setup. They also sanity checked the
judge behavior with Claude Sonnet 4.6
after Gemini 3.1 Pro preview scored low
on the benchmark itself. Second, there
was a contamination issue. When the
panel models had web search, they
accidentally found the Draco grading
rubric online. OpenRouter said this was
not intentional cheating, but it was
still a real eval problem.
So, they blocked the domains where the
rubric was hosted using excluded [music]
domains for web search and blocked
domains for webfetch. Their server tools
support that through providers like Exa
or Parallel. So, applying those
exclusions was basically a config
change. The published results were
produced after those blocks were in
place. Third, the Fable 5 number is
slightly uneven. Seven of the 100 Draco
tasks were not completed because Fable
5's content filters blocked them. Open
router chose not to fall back to Opus
4.8 for those tasks. So, Fable score
reflects 93 scored tasks, not all 100.
That gives a cleaner picture of Fable's
own behavior, but it also means direct
comparisons [music]
against models that completed all 100
tasks are not perfectly apples to
apples. And the biggest catch is long
horizon work. Open router is pretty
clear that Fusion is not a full Fable
replacement. Draco does not include long
horizon tasks, and that is exactly where
Fable 5 was apparently strongest. Long
horizon tasks are the jobs where the
model has to work through many dependent
steps, [music] browsing, coding,
planning, editing, fixing errors, using
tools, and continuing for a long time
without losing the plot. Fusion is
strong when multiple models can attack
the same question in parallel. But
deeply sequential work is different. If
step 20 depends on step 19,
>> [music]
>> you need stable memory, careful state
tracking, and consistent planning.
Fusion can struggle there, especially if
different steps are routed or blended
across different models without one
persistent identity carrying the whole
workflow. Fable 5 has an advantage there
because it is a single model with
stronger long context coherence and
instruction fidelity. That can matter
for huge documents, large code bases,
and layered prompts. But Fable is not
perfect, either. Long workflows can
still suffer from context drift,
compounding mistakes, instruction
forgetting, and premature stopping. And
because Anthropic models are heavily
safety trained, they can sometimes stall
or add friction in a workflow if they
detect risk halfway through. So, the
real choice is use case dependent.
Fusion makes a lot of sense when cost
efficiency matters, when you have high
volume workloads, when the tasks vary a
lot, when model diversity is valuable,
or when you want strong deep research
without depending on a single provider.
Fable 5 makes more sense when you need a
single known model for compliance, audit
logs, strict reproducibility, enterprise
contracts, complex long context
reasoning, or situations where small
behavior changes are unacceptable. Mind
Studio's take is that benchmarks are
useful, but the real test is your own
workflow. They present both Fusion and
Fable 5 inside a platform
>> [music]
>> with more than 200 models, so builders
can swap models without managing
separate API keys or rate limits.
They also argue that structured
workflows reduce the long horizon
problem, because instead of asking a
model to plan 50 steps alone, you define
the sequence and let models reason
inside each step. Their Remy product
fits the same pitch. It asks scoping
questions, writes specs, manages builds,
and uses Mind Studio's existing
infrastructure like managed databases,
auth, payments, crons, and over 1,000
integrations. Essentially, Fusion shows
a different path forward. Instead of
relying on one giant model to do
everything, you can combine several
models, compare their answers, catch
blind spots,
>> [music]
>> and produce a stronger final result. It
still does not fully replace Fable 5,
especially for long horizon tasks, but
for deep research and complex questions,
it is probably the closest practical
work-around available right now.
All right, that's it for this one. Let
me know if you would actually use
Fusion. Thanks for watching, and I'll
catch you in the next one.
>> Mhm.