/Tech8h ago

Aaron Levie agrees AI workloads will split between high-end frontier models and cheaper, high-volume alternatives

Coinbase predicts 80% of workloads will shift to cheaper models.

844191517071.2K

#14

Original post

kache@yacineMTB#244inTech

The frontier models are pretty retarded, and are only useful for chainsaw level tasks that don't require that much intelligence like mathematics

Brian Armstrong@brian_armstrong

Good take

My guess is - demand for intelligence is near infinite - but 80% of workloads will be running on 99% cheaper models within 12-18 months - 20% of workloads will still run on latest gen models where IQ maxing is important (scientific breakthroughs, higher level ochestrator agents?) - rough analogy might be what % of macbooks or gaming PCs sold have the maxed out specs for CPU/GPU, prices are falling much faster than Moore's law here though - this leads me to think the limiting factor will be energy and compute, not better models

At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.

7:38 PM · Jun 7, 2026 · 15.2K Views

/Tech8h ago

Aaron Levie agrees AI workloads will split between high-end frontier models and cheaper, high-volume alternatives

Coinbase predicts 80% of workloads will shift to cheaper models.

844191517071.2K

#14

Original post

kache@yacineMTB#244inTech

The frontier models are pretty retarded, and are only useful for chainsaw level tasks that don't require that much intelligence like mathematics

Brian Armstrong@brian_armstrong

Good take

At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.

7:38 PM · Jun 7, 2026 · 15.2K Views

Sentiment

Positive users praise cost optimization and model routing as key AI differentiators, while negative users dismiss the cheaper-models shift predictions as cope or bait.

Pos

55.6%

Neg

44.4%

10 comments with sentiment.

Cluster Engagement

Posts from X

Most Activity

VIEWS47.4KBOOKMARKS132LIKES178RETWEETS9REPLIES41

Aaron Levie@levie

The numbers may be a bit extreme here, but unquestionably use-cases have to stratify in the next year or two between model families.

We’ll see a split between frontier intelligence for high end tasks and work, and much cheaper models for high volume workloads that can sufficiently be peeled off to cheaper models. Frontier will still be far bigger than today because the use-cases will demand it, but the low-end will get quite a bit larger as well.

The big update here is that the layer that can efficiently route the workload to the right model will then become increasingly valuable since that becomes one of the new hard problems in AI agents. Agent orchestration that can cost optimize while still performing the task successfully will be in a strong position.

Brian Armstrong@brian_armstrong

Good take

At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.

6h47.4K178132

Jerry Liu@jerryjliu0

I really do think we'll see a lot of value accrue in AI startups building "model routing as a service"

Not just OpenRouter - this includes a much broader set of verticalized agents and infrastructure. * For us it's document infrastructure: parsing, extraction, search * Another infra analogy is web search: Exa/Parallel * For verticalized apps it could be anything from Cognition to Harvey

The frontier labs own the underlying models, and their main application at the app layer is the enormous amounts of $$ that they have; but their main disadvantage is that they only own a subset of all the points on the Pareto curve.

Speaking from our team's experience, it is both non-trivial and extremely important to find a point on the pareto curve of accuracy and cost. There's enormous amounts of ML time spent on model evaluations and benchmarking, and infra time making sure that the service can scale reliability, without rate limits, and without blowing up cost. At the same time it's extremely important not just for cost reasons but also latency, and oftentimes long-tail accuracy in use cases that demand it.

Brian Armstrong@brian_armstrong

Good take

At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.

5h3K2212

kache@yacineMTB

Idk. The more this plays out the more I realize that the quality of intelligence isn't really what it seems to be. To me, AGI is the evolutionary process, the spread, the sampling, the search, the spamming

kache@yacineMTB

The frontier models are pretty retarded, and are only useful for chainsaw level tasks that don't require that much intelligence like mathematics

8h2.5K484

Julien Chaumond@julien_c

Model routing is growing a lot these days

Brian Armstrong@brian_armstrong

Good take

At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.

2h1.7K134

Aaron Levie@levie

@lessin Routing in a standalone sense isn’t valuable. But being an applied AI company that is neutral/agnostic to the model, and thus *can* route as needed, is the point.

3h94223

sam lessin 🏴‍☠️@lessin

@levie that isn't a big update ... this has been clear for a long time -- but why do you think the routing will be valuable ... it will exist and be used but likely commodity.

Aaron Levie@levie

The numbers may be a bit extreme here, but unquestionably use-cases have to stratify in the next year or two between model families.

6h2.4K82

kache@yacineMTB

One of the things that I cope about: it takes barely any effort to have a creative spark, and the cost of building something that has that would be too much to be super human. And we like it, so why? So automate the cheap drudgery, so we can create!

8h888151

kache@yacineMTB

It's just not that impressive you know? Not as impressive as some humans that I know. Some very, very smart humans, remarkable people I've had the opportunity to meet

8h5359

Josh Clemm@joshclemm

@levie You're also going to see a lot more companies post-training open weight models with their own data.

A lot of inference shops are starting to build out that capability to make it easier and easier for any decent tech company

6h4431

Darbeci@onceuponquantum

@yacineMTB On the other hand Demis hassabis said a few days ago we just see the surface level of what this frontier models can do

7h912

Tuning Engines : C³ Action Plane for Agentic AI@tuning_engines

@levie Even a 50% or worst case assuming 30% is still a very significant overall reduction in cost to your point.

6h93

donkey_j_kong@slashslashdjk

@yacineMTB They need more context. They are too forgetful.

7h68

Dylan@df00z

@yacineMTB Just use a tiny open source model and RAG in relevant documentation that is specific to your esoteric needs. AI can read a book real quick and actually not lie to you.

4h49

tk@terrific_kid

@levie I don't think this is a good take. Armstrong is talking corpo while his whole thing was selling people open source public good commodities. Let's say y'all are right, then super AI completely destroys the world order. Ok, fuck SCRUM

4h39

Dylan@df00z

@yacineMTB The politically correct term for retarded is overfit to hell. All those parameters lost in a sea of white noise, burning your tokens and creating crushing debt for the hyperscalers.

4h11

Alan Sass@alansass

What % of enterprises actively route workflows this way at this point from your experience? I agree that the agent orchestration layer seems missing and much needed at this point but it also seems very difficult to get to your point above when even AI companies selling services aren’t properly routing their/customer traffic yet.

6h106

AJ Asver@_aj

@levie 100% - we have started experimenting with cross-harness orchestration and are seeing real savings with code generation in our custom IDE:

6h301

max@mSanterre

@yacineMTB Bro wrote the first part and knew he needed to add bait at the end to make it a banger.

Top 10 minds of the century.

7h82

juggernaut@curlysaarthak

@yacineMTB "like mathematics" i guess someone was googling whats difference between sine cosine a while ago

8h221

Nigel Koh@nigelkoh

@levie the layer also needs a good harness. @coworkerapp we've also found that it's not just routing to the right model but the instructions and task structure have to adapt to the model too, otherwise you're just switching engines without changing the fuel.

4h69