GPT-4 at $24.7 per million tokens vs Mixtral at $0.24 - that's a 100x cost difference! Even if routing gets it wrong 20% of the time, the economics still work. But the real question is how you measure 'performance' - user satisfaction doesn't always correlate with technical metrics.

It's trivial to get better score than GPT-4 with 1% of the cost by using my propertiary routing algorithm that routes all requests to Gemini 2.5 Flash. It's called GASP (Gemini Always, Save Pennies)

Does anyone working in an individual capacity actually end up paying for Gemini (Flash or Pro)? Or does Google boil you like a frog and you end up subscribing?

If I actually had time to work on my hobby projects Gemini pro would be the first thing I’d spend money on. As is, it’s amazing how much progress you can squeeze out of those 5 chats every 24h; I can get a couple hours of before-times hacking done in 15 minutes, which is incidentally when free usage gets throttled and my free time runs out.

I've used Gemini in a lot of personal projects. At this point I've probably made tens of thousands of requests, sometimes exceeding 1k per week. So far, I haven't had to pay a dime!

How come you don't need to pay? Do you get it for free somehow?

There's free tier for API.

"When you use Unpaid Services, including, for example, Google AI Studio and the unpaid quota on Gemini API, Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies, including Google's enterprise features, products, and services, consistent with our Privacy Policy.

To help with quality and improve our products, human reviewers may read, annotate, and process your API input and output. Google takes steps to protect your privacy as part of this process. This includes disconnecting this data from your Google Account, API key, and Cloud project before reviewers see or annotate it. Do not submit sensitive, confidential, or personal information to the Unpaid Services."

Reference: https://ai.google.dev/gemini-api/terms

You get 1500 prompts on AIStudio across a few Gemini flash models. I think I saw 250 or 500 for 2.5. It’s basically free and beats the consumer rate limits of big apps (Claude, ChatGPT, Gemini, meta). I wonder when they’ll cut this off.

I've paid a few dollars a month for my API usage for about 6 months.

PPT (price-per-token) is insufficient to compute cost. You will also need to know an average tokens-per-interaction (TPI). They multiply to give you a cost estimate. A .01x PPT is wiped out by 100x TPI.

Are you saying that some models will take 100x more tokens than other (models in the same ballpark) for the same task? Is the 100 a real measured metric or just random numbers to illustrate a point?

With thinking models, yes 100x is not just possible, but probable. You get charged for the intermediate thinking tokens, even if you don't see them (which is the case for Grok, for example). And even if you do see them, they won't necessarily add value.

> With thinking models, yes 100x is not just possible, but probable

So the answer is no then, because I don't put reasoning and non-reasoning models in the same ballpark when it comes to token usage. You can just turn off reasoning.

the GPT 5 models use ~10x more tokens depending on the reasoning settings.

number of complaints / million tokens?

> How you measure 'performance'

I heard the best way is through valuations

> GPT-4 at $24.7 per million tokens

While technically true why would you want to use it when OpenAI itself provides a bunch of many times cheaper and better models?

RouterBench is from March 2024.

The framing in the headline is interesting. As far as I recall, spending 4x more compute on a model to improve performance by 7% is the move that has worked over and over again up to this point. 101 % of GPT-4 performance (potentially at any cost) is what I would expect an improved routing algorithm to achieve.

(The submitted title was "93% of GPT-4 performance at 1/4 cost: LLM routing with weak bandit feedback")

Incredible that they are using contextual bandits, and named it: Preference-prior Informed Linucb fOr adaptive rouTing (PILOT)

Rather than the much more obvious: Preference-prior Informed Linucb For Adaptive Routing (PILFAR)

That's pretty funny. I might need to pilfer it.

Is there a reason human preference data is even needed? Don't LLMs already have a strong enough notion of question complexity to build a dataset for routing?

> a strong enough notion of question complexity

Aka Wisdom. No, LLMs don't have that. Me neither, I usually have to step in the rabbit holes in order to detect them.

"Do you think you need to do high/medium/low amount of thinking to answer X?" seems well within an LLMs wheelhouse if the goal is to build an optimized routing engine.

How do you think that an LLM could come by that information? Do you think that LLM vendors are logging performance and feeding that back into the model or some other mechanism?

Why not something dumb like this: https://chatgpt.com/share/68b60199-b6ac-8009-b50d-3e7cfff1d7... (gpt-4o)

Yes, that's why they keep getting better and why Anthropic is switching privacy policy defaults to eat my data please.

LLMs don't have notions ... they are pattern matchers against a vast database of human text.

Please do a SELECT * from this database

What was the name of the rocket that brought the first humans into space?

This is like asking someone to make you a sandwich and expect them to read your mind to determine what kind of sandwich you want.

I'm very curious whether a) anecdotally, anyone has encountered a real enterprise cost-cutting effort focused on LLM APIs and b) empirically, whether anyone has done any research on price elasticity in LLMs of different performance scales.

So far, my experience has been that it's just too early for most people / applications to worry about cost - at most, I've seen AI to be accountable for 10% of cloud costs. But very curious if others have other experiences.

LLM is far from the highest AI related cost, so we basically don't care about optimizing LLMs.

Obviously we don't use the super expensive ones like GPT4.5 or so. But we don't really bother with mini models, because GPT4.1 etc.. are cheap enough.

Stuff like speech to text etc.. are still way more expensive, and yes there we do focus on cost optimization. We have no large scale image generation use cases (yet)

In which context? The serious engineering folks are still in exploration phase, costs is mostly not a concern as long as shipping velocity increases. Reselling repackaged tokens is a different beast, no experience here

> We instantiate this idea through Preference-prior Informed Linucb fOr adaptive rouTing (PILOT), a novel extension of LinUCB

Academics are pretty creative at naming their creations

I almost named my LoRA replacement BEMO, but that felt too cute, so it's just BEM (Bolt-on Expert Modules).

These router papers are popping up hard now. I have a gradient boosted router I've been playing with that ties into retrieval to provide adaptive routing. The truth about these routers is that you have to tune them on your workloads to get the full benefit, otherwise they test way better than they work in production. That was why I added the retrieval aspect to mine, otherwise your top line slice and reality are very different.

Unless your application is relatively trivial you would always want consistent behaviour as much as possible than some random metric that is used to proxy as "performance", routing is NOT the solution.

From last neurips https://automix-llm.github.io/automix/

Is this really the frontier of LLM research? I guess we really aren't getting AGI any time soon, then. It makes me a little less worried about the future, honestly.

Edit: I never actually expected AGI from LLMs. That was snark. I just think it's notable that the fundamental gains in LLM performance seem to have dried up.

First, I don't think we will ever get to AGI. Not because we won't see huge advances still, but AGI is a moving ambiguous target that we won't get consensus on.

But why does this paper impact your thinking on it? It is about budget and recognizing that different LLMs have different cost structures. It's not really an attempt to improve LLM performance measured absolutely.

I can totally see "it's not really AGI because it doesn't consistently outperform those three top 0.000001% outlier human experts yet if they work together".

It'll be a while until the ability to move the goalposts of "actual intelligence" is exhausted entirely.

Well right now, my niece of 7 years outperforms all LLM contenders in drawing a Pelican on a bicycle

I know this was a joke, but LLMs are quite good at this now. If your niece draws better then she’s a good artist.

I tried it in Gemini just now, it seems to have done a decent job: https://g.co/gemini/share/b6fef8398c01

So you don't expect AGI to be possible ever? Or is your concern mainly with the wildly different definitions people use for it and that we'll continue moving goal posts rather than agree we got there?

Got it, and yeah I agree with you there. I've been frustrated by a different view of it though, many people seem to have a definition and they are often wildly different.

even AI does not have a concrete definition.

Doesn't mean there aren't practical definitions depending on the context.

In essence, teaching an AI using recources meant for humans, and nothing more, would be considered AGI. That could be a practical definition, without needing much more rigour.

There is indeed no evidence we'll get there. But there is also no evidence LLM's should work as well as they do

Well, if a human is GI, we just need to make it Artificial. Easy.

I like to say that it's not AI -- it's just A.

Given OpenAI definition I’d expect AGI to be around in a decade or two. I don’t expect skynet, though maybe it’s a more realistic vision outcome that just droids mixing with humans.

LLMs are not on the road to AGI, but there are plenty of dangers associated with them nonetheless.

Agreed, broadly. I never really thought they were, but seeing people work on stuff like this instead of even trying to improve the architecture really makes it obvious.

Just 2 days ago Gemini 2.5 Pro tried to recommend me tax evasion based on non-existing laws and court decisions. The model was so charming and convincing, that even after I brought all the logic flaws and said that this is plain wrong, I started to doubt myself, because it is so good at pleasing, arguing and using words.

And most would have accept the recommendation because the model sold it as less common tactic, while sounding very logical.

Or you could understand the tool you are using and be skeptical of any of its output.

So many people just want to believe, instead of the reality of LLMs being quite unreliable.

Personally it's usually fairly obvious to me when LLMs are bullshitting probably because I have lots of experience detecting it in humans.

LLM is only useful if it gives shortcut to information with reasonable accuracy. If I need to double check everything, it is just extra step.

In this case I just happened to be domain expert and knew it was wrong. It would have required significant effort to verify everything with some less experienced person.

> even after I brought all the logic flaws and said that this is plain wrong

Once you've started to argue with an LLM you're already barking up the wrong tree. Maybe you're right, maybe not, but there's no point in arguing it out with an LLM.

There are cases when they are actually correct, instead of the human.

Yes, and there's a substantial chance they'll apologize to you anyway even when they were right. There's no reason to expect them to be more likely to apologize when they're actually right vs actually wrong- their agreeableness is really orthogonal to their correctness.

Yes, they over-apologize. But my main reason for using LLMs is seeking out things that I missed myself or my own argumentation was not good. Sometimes they are really good at bringing new perspectives. Whether they are correct or incorrect is not the point - are they giving argument or perspective that is worth inspecting more with my own brains?

Is a random paper from Fujitsu Research claiming to be the frontier of anything?

Not just this paper, but model working shenanigans also seem to have been a big part of GPT-5, which certainly claims to be frontier work.

I'm not following this either. You'd think this would be frontier back in 2023

That and LLMs are seemingly plateauing. Earlier this year, it seemed like the big companies were releasing noticeable improvements every other week. People would joke a few weeks is “an eternity” in AI…so what time span are we looking at now?

That's just the thing. There don't seem to have been any breakthroughs in model performance or architecture, so it seems like we're back to picking up marginal reductions in cost to make any progress.

There have been very large improvements in code generation in the last 6 months. A few weeks without improvement are not necessarily a plateau.

Wait until it ramps up so much that people will say "it's a plateau, for real this time" when they go 3 days without a +10% capability jump.

I mean I wish there were a plateau, without one we're well onto our way into techno-feudalism. I just don't see it.

That's what it is: wishful thinking. A lot of people really, really want AI tech to fail - because they don't like the alternative.

Yeah, obviously nobody that actually though about the consequences wants a large part of the population to become unemployed. Even if your job is not threatened by automation, it will be threatened by a lot of people looking for new jobs.

And the kind of automation brought by LLMs is decidely different than automation in the past which almost always created new (usually better) jobs. LLMs won't do this (at least to extent where it would matter) I think. Most people in ten years will have worse jobs (more physically straining, longer hours, less pay) unless there will be a political intervention.

just because it’s on arxiv doesn’t mean anything

arxiv is essentially a blog under an academic format, popular amongst asian and south asian academic communities

currently you can launder reputation with it, just like “white papers” in the crypto world allowed for capital for some time

this ability will diminish as more people catch on

arxiv should really have a big red banner "NOT REVIEWED - DON'T USE AS A SOURCE" or something

Or we'll realize that human intelligence and machine intelligence is apple and oranges.

Would there be advantages to routing to models according to cost in conjunction with prompt rewriting?