> Just as a painter knows exactly what brush they need to use, and the craftsman carefully chooses their tools, engineers should understand trade-offs and subtleties of the different models.
I'm really skeptical of this idea. Pragmatically: who has time to understand the nuances of these models when there's like a new one every week? Also without any view into the training, figuring out what each model is potentially good at is more or less just throwing spaghetti against the wall, except the spaghetti is potentially very expensive and might insert subtle issues into your code base.
I just ask our fleet maintenance agent to configure the variants of the new model and then run the bakeoff. In a few hours, I get a doc containing exactly how it performed on my real tasks, against the other models we've benchmarked.
We don't do any "prompt engineering" to make specific models work better or worse, so this gives us a direct comparison that lets us objectively compare how they'll work in practice for different types of tasks.
Good enough to have guided multiple model swaps and measured increases in productivity.
There are "Hi" and "Lo" variants of the different models.
Maybe Gemini or ChatGPT is a "Hi" while running a local Mistral 7b is a "Lo" - sometimes you want lo-res technologies. Faster, lighter, does the job.
Find the same with image, video, vision, and so-on.
Probably get away with pretty weak vision models for some basic edge detection, but need "a good one" for things like OCR.
I don't find a ton of competitive difference amongst the "good ones" especially not reflected in the pricing and marketing. IMO there are pretty much just "non-AI" algorithms, "basic AI" that are getting more interesting on-device for some use cases, and "good AI" or what we'd call "frontier models" - which again, not seeing a ton of differentiation on that layer.
You don't just try every model that comes along... and for the ones you do, you have eval pipelines set up. It's engineering, you have to test stuff works.
Now, there are thousands of use cases for lots of the fontier models. I don't use them for everything, so I don't test them except on the stuff I'm using them for.
I just use whatever's cheapest for personal things. Dsv4 flash and got 5.6 Luna are great for the price and I'll use the openrouter benchmark view to compare the overall score.
I spent a lot of time researching LLM routing last year and also came to the conclusion that it's generally not worth the effort. It's too hard to understand the difficulty of a query a priori.
One specific challenge I was seeing is that difficulty depends a lot on what information is retrievable by the agent. Consider the question "what is the 5-state busy beaver number?" (https://en.wikipedia.org/wiki/Busy_beaver). In 2023 this would be a Mythos-tier research problem, but a solution was proved in 2024 so today any minimally intelligent model with a web search tool can just fetch the answer. You don't know which queries will be basic summarization and which will be deep reasoning until you get going.
I think routing should be pushed 'down the stack' so to speak.
What I mean is that most tasks can be recursively fragmented into smaller tasks, and once you've hit suitable leaf nodes -- where the task is very granular -- you can begin to deterministically show which models perform better or worse for that specific task. Then, when your agent is running a workflow, you may use various models for different steps in a workflow. For example, some models may excel at exploration, some at determining a good architectural fit for an implementation, some at actually writing the implementation, and so on.
But you don't know until you define your 'work' taxonomy, and still further, you won't know until you have a statistically significant number of runs on a given chunk of work. Once you have that, though, you can hone in on models that excel at one specific task or another and prefer those the majority of the time (say ~80%) and hold back the remainder work as a 'test corpus' just in case a different or new model does even better.
This is something I've kept in the back of my head as I've been working through my agent harness primitives -- specifically enabling different models per chunk of work.
You also have problems with short prompts whose results depend highly on the understanding of nuance. The router is going to have to mostly solve the prompt to decide which model to send it to.
What you actually want is a model that can conclude either "I know the answer to this with confidence" and answer, or "I think I don't know the answer to this, I should ask another model and I know which one". But I don't think LLMs can really bring their uncertainty to the surface in that way yet? Their internal confidence can be measured and returned, so you could probably front a more powerful model with a knowledgeable assistant, but they can't consciously mark their own homework?
> Their internal confidence can be measured and returned
It can? I was under the impression that confidence was either self-reported by the LLM or assessed by having another model interpret the output response. If there's a confidence score at the level of the actual model math, that's news to me.
You certainly can't ask an LLM for its confidence level because it will generate that answer as it sees fit.
My understanding was that models can be post-trained to assess their own confidence on short prompts within a level of accuracy (there was an article about this a few weeks ago that I can't find), but can't use it to reason, and that research has shown that internally they effectively have a measurable sense of truth but can't surface it:
I mangled what I was trying to say but the point I guess is, it is effectively in there, and researchers can see it and therefore score it, but it is not something the LLM can use.
The problem seems to me (layman's understanding at best) that the LLM is inherently confident within the words of its answer, because of what the model is trained to do and how it is trained. So you can never get an accurate "I don't know this" from a model while it is answering; it is bullshitting.
(One of the things that is most interesting to me at the moment is asking a small local LLM what it knows about a topic and then testing it. It can have no idea of what it wasn't trained on, but any question about what it knows is seemingly going to trigger it to work through all the summary word associations it did find in training, so it can sort of summarise its knowledge that way, with some likelihood of success, without any sense of introspection)
Ahh you found the link I couldn't find with a search.
Yes — they claim that it is likely to be able to assess confidence in whether the statement it just made is correct. But it isn't going to be capable of that while it is answering.
I mean can't you just directly test some variations of the query against many models at once and pick the cheapest model that hits your accuracy goal? If it's a one time run then ofc this is all pointless, but for ongoing tasks it makes sense. This seems more logical to me than making another AI model of some sorts intuit the right LLM for the job.
I agree with one distinction - coding agent workflows can use defined subagent roles that are pinned to specific models and I have found this very effective. The orchestrator is building all the context to make these assignments - it’s not a dumb router. Using Minimax M3 for exploration and librarian tasks for example is fast and cheap - my $10 plan lasts all month and saves a lot of tokens for my main coding plan.
Weird tense combination. I'm not sure if it's officially wrong but it feels weird to change tense mid-sentence after so many words. "X will do ABCDEFGHIJKLMNOPQRSTUVWXY and does Z."
I thought it could be a typo for "X will do ... and do Z" but only when I went to write why it's weird, I realised it doesn't have to be a typo.
It doesn't even matter here. It would mean the same thing either way.
The 'personality' of the model matters a lot even for stuff like Code Gen. If you're building something with Fable and then get routed to Opus cause the router decided your task is not 'demanding enough' it's like: great, now I have to read complex prose like "That's the thing, and the thing is the point:..." Opus-isms when I just wanted to make a shopping cart
I think this should probably be scoped to 'generic router systems that don't understand query context' are not useful. We have had lots of good results with routers that understand the context of the types of workloads they process and can route requests to the most efficient models.
> Complexity cannot be deduced from the prompt alone. Let’s take an example: “evaluate the tests for the repo $GIT_REPO and improve them” can be a very simple task if you mention a personal website written in plain HTML5; or an incredibly complex task if you target the Linux kernel repo.
I don't think this a good example. The first step would be reading the documentation, reading an overview of the tests, then executing commands to run the tests. A cheap model could do that. After that, though, the router will have to figure out how complex the tests are, which is the hard part, but I can't imagine it's that hard to determine a bunch of C code is tougher than some HTML from looking at it. Unless they want to select one model at start and never change it, because that's pretty clearly not the right way to go about routing.
Regardless I agree that routers usually aren't worthwhile, at least in the form of something that's meant to be universal. It's probably more efficient to just change something in the repo code, whether that's skills or instructions or something else. Benefit of that being it's persistent, portable, and more well tuned than guessing complexity on every turn.
I think the author need to rethink out of the box what's LLM routers in a traditional sense ( input in, route, output ) and move to think how a router would work in agentic workflow. See cognition Devin Fusion design.
I recently wrote about first principles of model routing that I've learnt building a model router.
The model pool should be kept small, and models in the pool should be clearly differentiated. For example, one large frontier model for quality, one small, fast and cheap model like DeepSeek V4 Flash for routing work.
These two principles by themselves solve the issues with caching, with routing decision making. I routinely hit >99% cache while routing between GPT 5.4 and DeepSeek.
I find this routing problem to be opaque and I’m generally skeptical that the label people are trying to predict is meaningful.
If you really need more discrimination of the complexity of an input to get an efficient response, sft or rl tuning something for your harness would be more effective.
To add: I doubt frontier labs will build routers - they are not financially incentivized to optimize token usage (though in the short term they may be incentivized by constrained GPU capacity to reduce load).
What was the architecture of your router? If it was based on GRPO/RL, it would be interesting to hear why your router performance capped.
I think the truth is that it's not an efficient cost cutting method. Your router has to be at least as 'smart' as all the but the smartest of your models (models do poorly when asked 'is this a task you're well suited to'), and that means you're caching multiple prompt histories including kv-filling/prefix caching on your expensive router model. Most of the time, not super great for savings.
Insider take: routing will not be a (successful, durable) thing, at least not externally to model providers.
The labs are incentivized to solve this problem themselves, since they’re competing on a 2D cost-intelligence frontier. If they can reduce cost without harming intelligence they will do that and pass on (some of) the cost reduction to the user. There are nicer solutions available to them because they can cut into lower levels of abstraction.
E.g. you should consider speculative decoding to be one (very conservative) form of routing and note that you can’t implement that for the labs from the outside.
This arbitrage isn't durable either though, as it's just a race to the bottom with open research, and the long term destination of everything is custom models for different tasks, since it's going to become increasingly apparent that some areas of knowledge have anti-synergy as we push up the diminishing return curve.
From a big-O notation perspective, if you're serving N queries, working the kinks out of your routing system costs perhaps O(logN) in developer time, whereas LLM provider savings grow as O(N).
In other words, the more queries you're serving, the more worthwhile it looks to figure out a viable method of model routing.
Even the simplest router that differentiates models between plan & execute will ensure that this is consistently followed. Folks have too high FOMO to choose themselves.
Yeah routers suck, been doing this myself for the best part of a year now and it’s really difficult to make it behave. A good model for your task is the best bet.
You cant fail over to a different model though. Even failing over to a different provider isn't always as simple as set it and forget it. For example bedrock doesn't support the tool search tool feature from anthropic.
Their router classified prompts into difficulty buckets. This obviously won't work. Consider a senior developer routing work based only on the task description. Clearly you need to dig a bit deeper into the task.
Saying routers don't work is sort of like saying serverless doesn't work. It depends on when and how!
One routing implementation that recently launched here is interesting (https://news.ycombinator.com/item?id=49099143). It routes based on the models' initial trajectories. This is like having multiple developers get started, seeing what they're doing, then pulling all but one off the project. It should work, but doesn't seem ideal!
I've gotten routing working well for typical chatbot prompts in http://pellmell.ai. This is fine because prompts are easy to classify into category buckets (for example: legal, medical, general knowledge, code). And models definitely have strengths and weaknesses. You want Gemini to answer General Knowledge and you want Claude to answer coding.
Some routing services not just route to different LLMs, they also handle all the legal issues (GDPR compliance, ISO certification, guaranteed Zero-Data-Retention, domestic data processing/European based clouds, etc.). In regulated industries, these things matter a lot, especially when processing of sensitive data is involved.
By "guaranteed" they mean "we promise really hard" right? Data retention has so much profit potential that any provider would have to be completely stupid not to retain data, even after promising they didn't.
Now that even the smaller models from labs (Luna, deepseek v4 flash) are getting powerful, I think the orchestrator pattern of a smart model coordinating smaller models for work will end up being the way to go.
> Just as a painter knows exactly what brush they need to use, and the craftsman carefully chooses their tools, engineers should understand trade-offs and subtleties of the different models.
I'm really skeptical of this idea. Pragmatically: who has time to understand the nuances of these models when there's like a new one every week? Also without any view into the training, figuring out what each model is potentially good at is more or less just throwing spaghetti against the wall, except the spaghetti is potentially very expensive and might insert subtle issues into your code base.
I just ask our fleet maintenance agent to configure the variants of the new model and then run the bakeoff. In a few hours, I get a doc containing exactly how it performed on my real tasks, against the other models we've benchmarked.
We don't do any "prompt engineering" to make specific models work better or worse, so this gives us a direct comparison that lets us objectively compare how they'll work in practice for different types of tasks.
Good enough to have guided multiple model swaps and measured increases in productivity.
There are "Hi" and "Lo" variants of the different models.
Maybe Gemini or ChatGPT is a "Hi" while running a local Mistral 7b is a "Lo" - sometimes you want lo-res technologies. Faster, lighter, does the job.
Find the same with image, video, vision, and so-on.
Probably get away with pretty weak vision models for some basic edge detection, but need "a good one" for things like OCR.
I don't find a ton of competitive difference amongst the "good ones" especially not reflected in the pricing and marketing. IMO there are pretty much just "non-AI" algorithms, "basic AI" that are getting more interesting on-device for some use cases, and "good AI" or what we'd call "frontier models" - which again, not seeing a ton of differentiation on that layer.
I feel like nearly anything done by a "Lo" is better done myself. I use "Hi" for pushing the limits of my expertise.
You don't just try every model that comes along... and for the ones you do, you have eval pipelines set up. It's engineering, you have to test stuff works.
Now, there are thousands of use cases for lots of the fontier models. I don't use them for everything, so I don't test them except on the stuff I'm using them for.
Just don’t switch models every week and you’ll be fine. Especially in production.
You do still need to know though, different prompts have different affects on them.
I just use whatever's cheapest for personal things. Dsv4 flash and got 5.6 Luna are great for the price and I'll use the openrouter benchmark view to compare the overall score.
Found the drywaller
My landlord just muds over cob webs and open space. They're over 100 years old so they legally count as part of the building's foundation.
I'm not surprised that people who use LLMs all day for everything get a feel for the different ones.
I spent a lot of time researching LLM routing last year and also came to the conclusion that it's generally not worth the effort. It's too hard to understand the difficulty of a query a priori.
One specific challenge I was seeing is that difficulty depends a lot on what information is retrievable by the agent. Consider the question "what is the 5-state busy beaver number?" (https://en.wikipedia.org/wiki/Busy_beaver). In 2023 this would be a Mythos-tier research problem, but a solution was proved in 2024 so today any minimally intelligent model with a web search tool can just fetch the answer. You don't know which queries will be basic summarization and which will be deep reasoning until you get going.
I think routing should be pushed 'down the stack' so to speak.
What I mean is that most tasks can be recursively fragmented into smaller tasks, and once you've hit suitable leaf nodes -- where the task is very granular -- you can begin to deterministically show which models perform better or worse for that specific task. Then, when your agent is running a workflow, you may use various models for different steps in a workflow. For example, some models may excel at exploration, some at determining a good architectural fit for an implementation, some at actually writing the implementation, and so on.
But you don't know until you define your 'work' taxonomy, and still further, you won't know until you have a statistically significant number of runs on a given chunk of work. Once you have that, though, you can hone in on models that excel at one specific task or another and prefer those the majority of the time (say ~80%) and hold back the remainder work as a 'test corpus' just in case a different or new model does even better.
This is something I've kept in the back of my head as I've been working through my agent harness primitives -- specifically enabling different models per chunk of work.
You also have problems with short prompts whose results depend highly on the understanding of nuance. The router is going to have to mostly solve the prompt to decide which model to send it to.
What you actually want is a model that can conclude either "I know the answer to this with confidence" and answer, or "I think I don't know the answer to this, I should ask another model and I know which one". But I don't think LLMs can really bring their uncertainty to the surface in that way yet? Their internal confidence can be measured and returned, so you could probably front a more powerful model with a knowledgeable assistant, but they can't consciously mark their own homework?
> Their internal confidence can be measured and returned
It can? I was under the impression that confidence was either self-reported by the LLM or assessed by having another model interpret the output response. If there's a confidence score at the level of the actual model math, that's news to me.
You certainly can't ask an LLM for its confidence level because it will generate that answer as it sees fit.
My understanding was that models can be post-trained to assess their own confidence on short prompts within a level of accuracy (there was an article about this a few weeks ago that I can't find), but can't use it to reason, and that research has shown that internally they effectively have a measurable sense of truth but can't surface it:
https://arxiv.org/abs/2410.02707
I mangled what I was trying to say but the point I guess is, it is effectively in there, and researchers can see it and therefore score it, but it is not something the LLM can use.
The problem seems to me (layman's understanding at best) that the LLM is inherently confident within the words of its answer, because of what the model is trained to do and how it is trained. So you can never get an accurate "I don't know this" from a model while it is answering; it is bullshitting.
(One of the things that is most interesting to me at the moment is asking a small local LLM what it knows about a topic and then testing it. It can have no idea of what it wasn't trained on, but any question about what it knows is seemingly going to trigger it to work through all the summary word associations it did find in training, so it can sort of summarise its knowledge that way, with some likelihood of success, without any sense of introspection)
At least one model can today. https://github.com/cactus-compute/cactus-hybrid
Ahh you found the link I couldn't find with a search.
Yes — they claim that it is likely to be able to assess confidence in whether the statement it just made is correct. But it isn't going to be capable of that while it is answering.
I mean can't you just directly test some variations of the query against many models at once and pick the cheapest model that hits your accuracy goal? If it's a one time run then ofc this is all pointless, but for ongoing tasks it makes sense. This seems more logical to me than making another AI model of some sorts intuit the right LLM for the job.
I agree with one distinction - coding agent workflows can use defined subagent roles that are pinned to specific models and I have found this very effective. The orchestrator is building all the context to make these assignments - it’s not a dumb router. Using Minimax M3 for exploration and librarian tasks for example is fast and cheap - my $10 plan lasts all month and saves a lot of tokens for my main coding plan.
Classifying complexity from the prompt alone was always the weak point. Most of what makes a task hard shows up after the first tool call.
Ironically, my confidence that a human had at least an active part in writing/editing this article went up because of this train wreck of a sentence:
> "A cache-aware model router will take that into account by adding stickiness to the initially chosen model and keeps querying it."
what’s wrong with the sentence? reads fine to me
Weird tense combination. I'm not sure if it's officially wrong but it feels weird to change tense mid-sentence after so many words. "X will do ABCDEFGHIJKLMNOPQRSTUVWXY and does Z."
I thought it could be a typo for "X will do ... and do Z" but only when I went to write why it's weird, I realised it doesn't have to be a typo.
It doesn't even matter here. It would mean the same thing either way.
'keeps' is the wrong tense and could be 'will keep'
i think it should be "by adding stickiness to the initially chosen provider of that model"
The 'personality' of the model matters a lot even for stuff like Code Gen. If you're building something with Fable and then get routed to Opus cause the router decided your task is not 'demanding enough' it's like: great, now I have to read complex prose like "That's the thing, and the thing is the point:..." Opus-isms when I just wanted to make a shopping cart
I think this should probably be scoped to 'generic router systems that don't understand query context' are not useful. We have had lots of good results with routers that understand the context of the types of workloads they process and can route requests to the most efficient models.
> Complexity cannot be deduced from the prompt alone. Let’s take an example: “evaluate the tests for the repo $GIT_REPO and improve them” can be a very simple task if you mention a personal website written in plain HTML5; or an incredibly complex task if you target the Linux kernel repo.
I don't think this a good example. The first step would be reading the documentation, reading an overview of the tests, then executing commands to run the tests. A cheap model could do that. After that, though, the router will have to figure out how complex the tests are, which is the hard part, but I can't imagine it's that hard to determine a bunch of C code is tougher than some HTML from looking at it. Unless they want to select one model at start and never change it, because that's pretty clearly not the right way to go about routing.
Regardless I agree that routers usually aren't worthwhile, at least in the form of something that's meant to be universal. It's probably more efficient to just change something in the repo code, whether that's skills or instructions or something else. Benefit of that being it's persistent, portable, and more well tuned than guessing complexity on every turn.
I think the author need to rethink out of the box what's LLM routers in a traditional sense ( input in, route, output ) and move to think how a router would work in agentic workflow. See cognition Devin Fusion design.
I recently wrote about first principles of model routing that I've learnt building a model router.
The model pool should be kept small, and models in the pool should be clearly differentiated. For example, one large frontier model for quality, one small, fast and cheap model like DeepSeek V4 Flash for routing work.
These two principles by themselves solve the issues with caching, with routing decision making. I routinely hit >99% cache while routing between GPT 5.4 and DeepSeek.
https://try.works/first-principles-of-model-routing
I find this routing problem to be opaque and I’m generally skeptical that the label people are trying to predict is meaningful.
If you really need more discrimination of the complexity of an input to get an efficient response, sft or rl tuning something for your harness would be more effective.
To add: I doubt frontier labs will build routers - they are not financially incentivized to optimize token usage (though in the short term they may be incentivized by constrained GPU capacity to reduce load).
What was the architecture of your router? If it was based on GRPO/RL, it would be interesting to hear why your router performance capped.
I think the truth is that it's not an efficient cost cutting method. Your router has to be at least as 'smart' as all the but the smartest of your models (models do poorly when asked 'is this a task you're well suited to'), and that means you're caching multiple prompt histories including kv-filling/prefix caching on your expensive router model. Most of the time, not super great for savings.
Insider take: routing will not be a (successful, durable) thing, at least not externally to model providers.
The labs are incentivized to solve this problem themselves, since they’re competing on a 2D cost-intelligence frontier. If they can reduce cost without harming intelligence they will do that and pass on (some of) the cost reduction to the user. There are nicer solutions available to them because they can cut into lower levels of abstraction.
E.g. you should consider speculative decoding to be one (very conservative) form of routing and note that you can’t implement that for the labs from the outside.
This arbitrage isn't durable either though, as it's just a race to the bottom with open research, and the long term destination of everything is custom models for different tasks, since it's going to become increasingly apparent that some areas of knowledge have anti-synergy as we push up the diminishing return curve.
From a big-O notation perspective, if you're serving N queries, working the kinks out of your routing system costs perhaps O(logN) in developer time, whereas LLM provider savings grow as O(N).
In other words, the more queries you're serving, the more worthwhile it looks to figure out a viable method of model routing.
I'm still working on mine: https://github.com/rush86999/atom/blob/main/docs/architectur...
>Our router was classifying each request into one of four different tiers of complexity: simple, standard, complex and reasoning.
Seems like a naive classification model lacking context?
Routing belongs on the client side
Even the simplest router that differentiates models between plan & execute will ensure that this is consistently followed. Folks have too high FOMO to choose themselves.
Yeah routers suck, been doing this myself for the best part of a year now and it’s really difficult to make it behave. A good model for your task is the best bet.
nobody else is putting a router inside a Trusted Execution Environment https://trustedrouter.com/
I think fail-over for a production critical service is an equally important responsibility of the router.
You cant fail over to a different model though. Even failing over to a different provider isn't always as simple as set it and forget it. For example bedrock doesn't support the tool search tool feature from anthropic.
They beta support for it now I believe
Their router classified prompts into difficulty buckets. This obviously won't work. Consider a senior developer routing work based only on the task description. Clearly you need to dig a bit deeper into the task.
Saying routers don't work is sort of like saying serverless doesn't work. It depends on when and how!
One routing implementation that recently launched here is interesting (https://news.ycombinator.com/item?id=49099143). It routes based on the models' initial trajectories. This is like having multiple developers get started, seeing what they're doing, then pulling all but one off the project. It should work, but doesn't seem ideal!
I've gotten routing working well for typical chatbot prompts in http://pellmell.ai. This is fine because prompts are easy to classify into category buckets (for example: legal, medical, general knowledge, code). And models definitely have strengths and weaknesses. You want Gemini to answer General Knowledge and you want Claude to answer coding.
Some routing services not just route to different LLMs, they also handle all the legal issues (GDPR compliance, ISO certification, guaranteed Zero-Data-Retention, domestic data processing/European based clouds, etc.). In regulated industries, these things matter a lot, especially when processing of sensitive data is involved.
By "guaranteed" they mean "we promise really hard" right? Data retention has so much profit potential that any provider would have to be completely stupid not to retain data, even after promising they didn't.
Now that even the smaller models from labs (Luna, deepseek v4 flash) are getting powerful, I think the orchestrator pattern of a smart model coordinating smaller models for work will end up being the way to go.