Dogpile was only a good idea while Search Engines were mostly trash.
You needed to search all of them to find something decent.
That's roughly analogous to today. Ignoring cost, you'd be way better off asking all the LLMs to solve a problem (like coding) where you can verify the answer.
So the question is, for things like that -> can a group of models perform better than frontier models, especially at a reasonable cost?
Fable is not a great value, so unless you're trying to find answers to Erdos questions, you can probably do better on cost.
You can probably typically ask 3 or 4 of the top Chinese models for an answer and get a response for the same Fable question... Given that Fable isn't that much better, it's not surprising you can do better for a large subset of problems.
Mixture of Models, perhaps? tbf to OP, the setup they're proposing has also been recently evangelized by other "ai gateway" products (like OpenRouter, JusCode, Fireworks etc), so there's likely something useful here.
thanks to everyone for taking the time to try Echo and share feedback, this is precisely why i wanted to launch early.
i am going to try to address a couple of topics that came up often:
- i'll keep publishing stronger evals, including more difficult coding and agentic benchmarks, to map out more precisely the differences with sota
- the public eval dashboard will keep expanding and be updated (very open to more benchmark suggestions as well!)
- some people found issues in the eval dashboard ui and the sign up flow, should be now all fixed in prod
some important precisions as well:
- NO credit card is required to try Echo
- each acount includes 10$ of free credits to try on both the API and the chat
on the approach itself: the idea i'm exploring is more broader than model routing, i'm looking at how to allocate inference efficiently across open-weight models, deciding not only which models to use, but also how much computation a request deserves and how intermediate work should be combined.
ensembling by itself is not new. since random forests and probably even before in statistics/classic ml we knew that bringing multiple models together can outperform individual ones. the interesting problem for Echo is how to model and leverage this without paying the full ensemble cost at each request.
while there are conceptual similarities with systems like Fusion or Fugu, the architecture and optimization objective are different.
I am guessing this is not targeting those of us on the heavily subsidized $200/mo plans. Sure, these plans may be temporary, but none of us really know how temporary they are. Until then, 1/3rd of the published API pricing is not very appealing.
I burned though my weekly fable usage last night on the $200 plan. I had $200 in promotional usage credits and was in the middle of executing a moderate sized coding plan. Ran on usage credits for about 1h 15m and burned $120 in usage credits. I was astounded to see how fast the $ usage added up. One problem was that I was using sub-agent execution so multiple agents were running simultaneously and I realized at the end that claude had "Forgotten" my directive to use cheaper models as appropriate for sub-agent tasks so I was running multiple instances of Fable at once. Still hard to imagine paying per token. $200 a month is high, $200 per night is crazy.
All enterprises users (people using them for work and not side projects) can't get the subsidized plans. I would say subsidized plans are a minority of usage?
People _can_ get subsidized plans for work: we use Claude Teams, $100/mo premium seat, which caps at 150 seats. Not enterprise tier, but fine for SMBs.
I imagine that most small to medium sized businesses are on either individual plans or Teams plans. The vast majority of firms do not need more than 150 seats, and API rates are not sustainable for most.
I have been trying something like this with local only models. I think I’ve tried everything that will fit in 96GB alone and in concert with each other using real project data (Rails). I haven’t found much benefit to anything other than Qwen 3.6 27b dense feeding a Claude final pass. I really thought layering was going to work like the law of large numbers, but for my workload it just hasn’t.
You are right that the privacy wording was too broad. We are fixing it now so it states explicitly that Echo does not use customer prompts, files, chats, or outputs to train or fine-tune models. We are updating the matching Terms language at the same time.
Echo also starts users with free credit and does not require a credit card to try it. The current signup flow did not make that clear enough, so we are fixing that presentation too.
No benchmarks, no info on which models are used, ai generated video, just a signup page with nothing else.
Anyhow, this kinda reminds me of that quote about architecture: "We replaced our monolith with micro services so that every outage could be more like a murder mystery."
It currently exposes 907 stored rows across seven benchmark families, with prompts, outputs, grades, and cost records. More benchmarks are coming soon.
Echo does not disclose its per-request routing decision because that policy is the product. We can, however, publish some of the eligible open-weight model pool, version dates, aggregate allocation mix, and evaluation settings without exposing the request-level recipe.
> Echo does not disclose its per-request routing decision because that policy is the product
My honest advice: that's going to pull away a decent amount of potential customers, even though I think your idea/concept is fantastic.
For example, if we were to consider it for Canva, observerability and full transparency is critical requirement; we can't accept not knowing which model serves a request. Both for legal/contract reasons, co-ordinated capacity planning with API providers, or even just evaluating our prompts and harnesses; and debugging/tracing results that went wrong. So that renders it out of consideration; and also suggests some kind of adversarial relationship where customers aren't trusted with critical information.
I definitely understand you need to keep business value, but I don't think hiding which model a request is routed to, is the right one, or at least if you want to expand to bigger potential customers / more advanced LLM deployments.
> Echo does not disclose its per-request routing decision because that policy is the product
I have not had need for a router product thus far so excuse my ignorance if this is standard, but how could I possibly use and improve a product built on a router like this if I am not permitted to see which model served my request? If I got a bad answer back in my LLM-powered app, do I really have no way of knowing which model was responsible?
I've been working on a similar project and I found that it's easy to replicate Fable results if you use saturated benchmarks.
In my project, I wasted a huge amount of time trying to improve GPQA Diamond results above ~93% range. I realized my mistake when Fable dropped and made no improvement on this benchmark vs. Opus.
I wouldn't be surprised if ≈7% of GPQA Diamond questions simply have the wrong answer in the ground truth data, so that getting such a question correct is graded as an error. Most machine-learning benchmarks are rather badly validated.
It's basically trying to replicate OpenRouter, which works pretty well and has a lot of nice features to abstract away any single provider, such as failover, metering, autoswitching, etc. It's actually a really smart infrastructure abstraction.
I just wish this were solving an actual problem rather than being a fairly transparent attempt to say something approximating, "Hey VCs, OpenRouter just became a unicorn but I can basically vibe code it"
Calling it "Fable-level" feels intellectually lazy / dishonest, but then again, what do you expect when there's so much money on the table.
It feels a bit more like Fugu to me, which acts as a multi-LLM orchestrator (though I think Fugu combines open- and closed-weight models), but without being able to see the “secret sauce” behind how any of them decides the number of "plies" each model in the swarm gets, they all feel rather difficult to compare beyond the big public benchmarks...
Honestly not sure. I think Fugu tries to leverage multiple models simultaneously where something like NotDiamond is more about picking the most optimal singular model.
This space is so crowded it feels like I see a new "model router" pop up every few weeks.
This does not really work well, if you don't know the complexity of the problem ahead of time and ensure all future conversations go to the same model.
Else, you break the cache by doing a round robin of the same conversation across different models. Likely you'll end up paying more than what it would've cost with a cache aware system
I'm curious if there is measurable value in diversity of thought, and if there's diminishing returns on a single models thought pattern.
For example, compute X tokens with model A, then feed those into model B, etc. to get chain of thought through a diverse set of mdoels rather than chain of thought through a heterogeneous chain.
Humans seem to strongly believe echo chambers are bad. Are LLMs the same?
there are and this what you optimize for. ensemble learning has a long literature on this. you want models that have the most diverse pool of capabilities so they complement each other. in verifiable tasks or classification this is straightfroward but a bit tricker in open ended text gen or CoT. this is one of our core research question
My experience is mostly around either ensembles of weak learners that are bootstrapped to create a strong learner, or around boosted learners where you're training on residuals. With CoT you're sort of adding entropy to your answer, but you're also possibly getting a more thorough answer?
I'm curious though if these training methods are convergent or are models actually different; just like how in the stock market people think they're "diversified" but the truth is their exposure is likely much more risk correlated than one might think.
In certain situations, one right answer is better than a committee discussing the problem, but in others its sometimes nice to have some alternative methods of solving something. Fun project nonetheless.
My approach to using multiple models has been less about CoT but more about time to first token, and how you can use a small model to start interacting with the user while in parallel the more complex model is building a larger more complex thought. My work on this was primarily for voice backed interfaces before the voice models became quite a lot faster.
Fusion is a totally different approach, though similar in the sense that it leverages different models.
Fusion generates many replies then synthesizes. This adds a ton of latency and cost, so it's going to be better only for cases where you're willing to wait a lot and pay a lot more.
Routers (like this project) are a different thing, they can theoretically improve performance and cost at the same time without increasing latency much. I'm a bit skeptical though, since knowing which LLM is going to be better on a cost adjusted basis is hard (see https://artificialanalysis.ai/models/capabilities/coding?cos..., where the cost per task vs. performance is not what you expect, for example comparing Qwen 3.7 Max to GPT Sol.
A project I'm working on is aimed at improving performance without added latency but from a different angle. Instead of waiting for all replies for synthesis (like OpenRouter Fusion), it streams the "best" reply immediately (using a router to pick the best model) then synthesizes with emoji reactions and optional replies from the background models. It's free to use here with no login: http://pellmell.ai
Sounds like another token router. Not sure what is the benifit compared with codex, I used 5.6 Sol, codex already route some of the task to luna not sol.
Thanks for checking the individual rows. HumanEval+ is one small code slice, not the whole basis for the launch claim. The public evaluator currently contains 907 rows across seven benchmark families, and matched SWE-bench Verified and BigCodeBench runs are the next code evidence being added.
You also found a real UI bug: the inspector should show both stored answers and currently does not in some rows. We are fixing that.
On the row you reran: the page records a frozen matched run. It does not claim that Fable is incapable of solving that prompt on another run. We are adding repeated matched trials and making run count and variance explicit. Your rerun is exactly the kind of external check the row-level page is intended to make possible.
The broader result remains: Echo is competitive with Fable across the evaluated task mix at materially lower measured inference cost. We are filling in the harder agentic-code evidence now rather than asking anyone to infer it from HumanEval+.
This reminds me on OpenRouters report that combining multiple different models gave comparable performance to Fable 5. I think this approach has lots of potential. Maybe OpenAi was ahead of its time with GPT-5 (it being a router to different models rather than just being one new model)
You might be interested in a project that I'm working on, which is kind of like OpenRouter Fusion, but instead of waiting for all models to synthesize, we stream the best model immediately and background the rest. The background models then reconcile with an emoji reaction and optional reply. It gets similar results to Fusion and is a lot faster! There's a free version that leverages open weight models here: http://pellmell.ai.
I like that I was able to test without an account and my prompt (What's the feasibility of a NetNavi IRL, along with a PErsonal Terminal? Create a document that outlines how to make this happen.) makes me think Pellmell is gonna be a research tool I'll continue to use.
OpenAI's router chooses between models of different sizes, which are still trained on roughly the same data. Its purpose is to reduce infra cost for OpenAI for simpler queries. No need to pay for GPT 5.6 Sol inference for "Hello" prompt.
I think approaches like this have potential. Only time will tell. This reminds me of the mixture of experts taken by deepseek r2 (I think it was r2, at least), but less specific models I guess.
I have often wondered how tools like GHCP choose the best model for the job when set to "auto".
So the word security or any topic related to it is mentioned and it flips to an older gen model? Fable is nearly useless now it you do anything around auth.
I'm not an expert on this, but this sounds a lot like a larger-scale MoE (Mixture of Experts) type of architecture.
As I understand it, in an MoE model, you essentially have hundreds of smaller sub-models ("experts") that are good at different tasks, and for every generated token, a single "master" model chooses which ones are most relevant to participate, and you only activate them.
In MoE systems the routing decision is made per-token, not per prompt or task. It’s one of ML’s many confusing naming conventions.
Even more confusingly, there are older pre-LLM MoE systems which ensemble and pool the predictions from multiple sub-components. For example in a random forest you could take the majority vote of the decision trees or the average of their numerical predictions.
After that, we developed neural net architectures for predicting a single thing like whether the user will click on your ad. An MMoE is in the same family.
And so now we are at massive MoE networks for LLMs which have similarities with MMoE in that the “decision” is about the very next token to predict.
The "ensemble of models" is nothing new. It's just that it's not really a moat that can be monetized. In the end you're always paying for something. You can engineer an elaborate harness with multiple models but it is not going to particularly solve a novel problem that the frontier models can with the same level of efficiency. You're saving money and paying with time. You're going to pay with something one way or another no matter what.
The frontier providers aren't dumb. They charge what they charge because they know this. If you think Fable is too expensive then the type of problems you are solving don't demand that level of capability.
If you are working on something cutting edge, something truly novel, the cost of frontier AI is well worth its price.
With all that being said. No one is going to complain if we can get the same capability at a lower cost. And I mean true parity. Not trading space for time.
Intuitively, your savings depend heavily on how hard the tasks are in the first place. If you have a base rate where 99% of your tasks can be routed to a cheap model, yeah, you can save a ton by not using Fable for that.
I have been using this : https://magnitude.dev/ for a while now. Is it something similar you are doing? I would love to have something that would connect to my codex, Claude, and opencode subscription rather than having to make a new subscription.
> Fable-level results at 1/3 the cost using open-weight models
But we get ~$2500/mo worth of Fable credits for $200/mo on Anthropic pan? I'm still confused why people (who don't have to use API billing) are chasing open weight models based on cost.
Enterprise plans pay API costs, they're only subsidizing individual accounts and that's only because they have to compete with open-source models.
When they are successful at making those illegal/inaccessible, both Anthropic and ChatGPT are going to rip the band-aid off and inference will only be sold to those that can afford it.
> When they are successful at making those illegal/inaccessible
This would be like trying to outlaw Linux or peer-to-peer file sharing. It's technically possible to write and pass a law, but it's basically impossible to enforce it.
Because that is a short term solution, it won’t be offered forever. Large organisations have to purchase credits at $/tokens. Eventually everyone else will too.
I think you are looking at it incorrectly. No business is buying individual accounts, because if they do, they open themselves up to considerable risk.
The $200 plans are priced so that the power-users use them and then advocate about how great the product is. If you're buying a $200 plan, you're not doing it because of the price point but rather because of the amount of work it is doing for you.
The point still stands. The chinese labs don't have super discounted plans, so if the price per task benchmarks[1] are correct, and we apply the discount, you'll actually be paying more by using cheaper chinese models and this technique.
I think most people assume the subsidized plans will go away or get more limited eventually. They are basically a loss leader and a marketing cost that is very flexible and easy to change w/o directly impacting their primary customers.
By the way, is it just me or opus 4.8 is much better at some programming tasks than fable? I've been really disappointed lately, i stopped using it evem though it's "premium on my subscription"
It's just you. Fable 5 is consistently better than Opus 4.8 at literally every single level, for me. (I always use both at xhigh, for reference.) I could go down a laundry list of various issues I have with Opus that I don't have with Fable. For me, Opus 4.8 < GPT-5.6 Sol < Fable 5.
Plus Fable is way less annoying to talk to than Opus 4.8. Opus 4.8's writing style is absolutely insufferable. Fable has some of the same quirks but it's way less bad.
We see no difference between opus 4.8 and fable except that fable is slower. So +1 for not just GP. We use neither interactive, just via our own tooling so the ‘talk to’ doesn’t apply and we use it 24/7. We currently run 25% of tasks on both fable and opus and the rest only opus. The 25% are being code reviewed side by side and we do detect when fable switches to opus for ‘security concerns’ nonsense. Opus generally finishes sooner, results are similar quality.
They're obviously not claiming this is a new model that is fable-like at 1/3. It's a router that saves money by only using Fable when necessary. I don't think you should be calling someone's post a "scam" without doing a minimum of research. (I'm not associated with the company, but very interested in this.)
I don’t find the recent spate of blog posts and systems delegating and combining LLMs to get better performance particularly interesting. Especially given that anyone who’s taken an ML 101 course has learned about ensemble methods.
While an LLM isn’t what you’d traditionally consider a weak learner, the theorems on learning systems clearly point to them being so in this context. The feigned surprise at combining them to yield better results seems disingenuous.
Even so, the work to predict which models are best suited for which task, how to delegate, and how to combine their outputs is interesting, especially if you’re placing a cost minimization objective on it. That said, this isn’t too far off from what many AI labs are already doing.
The entirety of "agents" and tool calls is a process of combining LLMs to get better results. Is it the same LLM in many cases? Yes. But it doesn't have to be.
It's the natural move that happened after people realized you couldn't throw away half a century of AI research.
Most of these focus on costs. But it is simply the case that the one-shot output did not scale for harder problems on workflows.
You definitely still would, but you need to pay the full input cost twice, so the equation really depends on how much first message vs repeat message matter.
So this is the dogpile.com of the askjeeves, alta vista, and lycos approach? Time is a flat circle?
Good ideas are usually still good across time and tools
Dogpile was only a good idea while Search Engines were mostly trash.
You needed to search all of them to find something decent.
That's roughly analogous to today. Ignoring cost, you'd be way better off asking all the LLMs to solve a problem (like coding) where you can verify the answer.
So the question is, for things like that -> can a group of models perform better than frontier models, especially at a reasonable cost?
Fable is not a great value, so unless you're trying to find answers to Erdos questions, you can probably do better on cost.
You can probably typically ask 3 or 4 of the top Chinese models for an answer and get a response for the same Fable question... Given that Fable isn't that much better, it's not surprising you can do better for a large subset of problems.
There is no reason a smarter model will not build an internal smarter/efficient router itself
> Dogpile was only a good idea while Search Engines were mostly trash.
Precisely.
Exactly
High quality, fast & cheap (all 3 combined) - is a formula success.
It’s just way easier said than done.
There's a reason most pros will tell you pick two of the three.
Well, you can try to make triangular wheels, but they are round for a reason
ensemble models always did the best at Kaggle
we did the same: https://trustedrouter.com/blog/prometheus-2-new-draco-state-...
Mixture of Models, perhaps? tbf to OP, the setup they're proposing has also been recently evangelized by other "ai gateway" products (like OpenRouter, JusCode, Fireworks etc), so there's likely something useful here.
Ensemble Methods:
https://en.wikipedia.org/wiki/Ensemble_learning
This sounds like networking.
Did OP invent an “intelligence router?”
Indeed. What a deep cut
metacrawler.com
timecube?
thanks to everyone for taking the time to try Echo and share feedback, this is precisely why i wanted to launch early.
i am going to try to address a couple of topics that came up often:
- i'll keep publishing stronger evals, including more difficult coding and agentic benchmarks, to map out more precisely the differences with sota
- the public eval dashboard will keep expanding and be updated (very open to more benchmark suggestions as well!)
- some people found issues in the eval dashboard ui and the sign up flow, should be now all fixed in prod
some important precisions as well:
- NO credit card is required to try Echo
- each acount includes 10$ of free credits to try on both the API and the chat
on the approach itself: the idea i'm exploring is more broader than model routing, i'm looking at how to allocate inference efficiently across open-weight models, deciding not only which models to use, but also how much computation a request deserves and how intermediate work should be combined.
ensembling by itself is not new. since random forests and probably even before in statistics/classic ml we knew that bringing multiple models together can outperform individual ones. the interesting problem for Echo is how to model and leverage this without paying the full ensemble cost at each request.
while there are conceptual similarities with systems like Fusion or Fugu, the architecture and optimization objective are different.
thanks again for all the thoughtful feedback.
> Fable-level results at 1/3 the cost
I am guessing this is not targeting those of us on the heavily subsidized $200/mo plans. Sure, these plans may be temporary, but none of us really know how temporary they are. Until then, 1/3rd of the published API pricing is not very appealing.
I burned though my weekly fable usage last night on the $200 plan. I had $200 in promotional usage credits and was in the middle of executing a moderate sized coding plan. Ran on usage credits for about 1h 15m and burned $120 in usage credits. I was astounded to see how fast the $ usage added up. One problem was that I was using sub-agent execution so multiple agents were running simultaneously and I realized at the end that claude had "Forgotten" my directive to use cheaper models as appropriate for sub-agent tasks so I was running multiple instances of Fable at once. Still hard to imagine paying per token. $200 a month is high, $200 per night is crazy.
All enterprises users (people using them for work and not side projects) can't get the subsidized plans. I would say subsidized plans are a minority of usage?
People _can_ get subsidized plans for work: we use Claude Teams, $100/mo premium seat, which caps at 150 seats. Not enterprise tier, but fine for SMBs.
I imagine that most small to medium sized businesses are on either individual plans or Teams plans. The vast majority of firms do not need more than 150 seats, and API rates are not sustainable for most.
The subsidized userbase is large enough that cheema is right to call this out this distinction for other readers.
I have been trying something like this with local only models. I think I’ve tried everything that will fit in 96GB alone and in concert with each other using real project data (Rails). I haven’t found much benefit to anything other than Qwen 3.6 27b dense feeding a Claude final pass. I really thought layering was going to work like the law of large numbers, but for my workload it just hasn’t.
No single signin. Privacy policy allows training. No try it first without credit card. It's a good idea, but this looks premature.
You are right that the privacy wording was too broad. We are fixing it now so it states explicitly that Echo does not use customer prompts, files, chats, or outputs to train or fine-tune models. We are updating the matching Terms language at the same time.
Echo also starts users with free credit and does not require a credit card to try it. The current signup flow did not make that clear enough, so we are fixing that presentation too.
Thanks for calling both out.
If you want to try a similar idea that you can chat with immediately (free tier uses open weight models), I'm working on this: http://pellmell.ai
I assume trying without a credit card will bankrupt him immediately.
No benchmarks, no info on which models are used, ai generated video, just a signup page with nothing else.
Anyhow, this kinda reminds me of that quote about architecture: "We replaced our monolith with micro services so that every outage could be more like a murder mystery."
The evaluator is public here: https://echo.tracerml.ai/eval/
It currently exposes 907 stored rows across seven benchmark families, with prompts, outputs, grades, and cost records. More benchmarks are coming soon.
Echo does not disclose its per-request routing decision because that policy is the product. We can, however, publish some of the eligible open-weight model pool, version dates, aggregate allocation mix, and evaluation settings without exposing the request-level recipe.
New video is also being made.
> Echo does not disclose its per-request routing decision because that policy is the product
My honest advice: that's going to pull away a decent amount of potential customers, even though I think your idea/concept is fantastic.
For example, if we were to consider it for Canva, observerability and full transparency is critical requirement; we can't accept not knowing which model serves a request. Both for legal/contract reasons, co-ordinated capacity planning with API providers, or even just evaluating our prompts and harnesses; and debugging/tracing results that went wrong. So that renders it out of consideration; and also suggests some kind of adversarial relationship where customers aren't trusted with critical information.
I definitely understand you need to keep business value, but I don't think hiding which model a request is routed to, is the right one, or at least if you want to expand to bigger potential customers / more advanced LLM deployments.
> Echo does not disclose its per-request routing decision because that policy is the product
I have not had need for a router product thus far so excuse my ignorance if this is standard, but how could I possibly use and improve a product built on a router like this if I am not permitted to see which model served my request? If I got a bad answer back in my LLM-powered app, do I really have no way of knowing which model was responsible?
Isn't the relevant benchmark RouterBench? https://arxiv.org/html/2403.12031v2#S7
> No benchmarks, no info on which models are used, […]
The benchmarks are here → https://echo.tracerml.ai/eval/
They are not good benchmarks but at least they exist.
I've been working on a similar project and I found that it's easy to replicate Fable results if you use saturated benchmarks.
In my project, I wasted a huge amount of time trying to improve GPQA Diamond results above ~93% range. I realized my mistake when Fable dropped and made no improvement on this benchmark vs. Opus.
I wouldn't be surprised if ≈7% of GPQA Diamond questions simply have the wrong answer in the ground truth data, so that getting such a question correct is graded as an error. Most machine-learning benchmarks are rather badly validated.
Yep! I found this interesting article after banging my head against a wall for a long time: https://epoch.ai/gradient-updates/gpqa-diamond-whats-left
It's basically trying to replicate OpenRouter, which works pretty well and has a lot of nice features to abstract away any single provider, such as failover, metering, autoswitching, etc. It's actually a really smart infrastructure abstraction.
I just wish this were solving an actual problem rather than being a fairly transparent attempt to say something approximating, "Hey VCs, OpenRouter just became a unicorn but I can basically vibe code it"
Calling it "Fable-level" feels intellectually lazy / dishonest, but then again, what do you expect when there's so much money on the table.
It feels a bit more like Fugu to me, which acts as a multi-LLM orchestrator (though I think Fugu combines open- and closed-weight models), but without being able to see the “secret sauce” behind how any of them decides the number of "plies" each model in the swarm gets, they all feel rather difficult to compare beyond the big public benchmarks...
https://github.com/SakanaAI/fugu
Isn't Fugu the same kind of thing as NotDiamond[1] (which I believe OpenRouter uses) except not as good?
[1] https://www.notdiamond.ai/
Honestly not sure. I think Fugu tries to leverage multiple models simultaneously where something like NotDiamond is more about picking the most optimal singular model.
This space is so crowded it feels like I see a new "model router" pop up every few weeks.
OpenRouter's model router for coding isn't really sophisticated. This is very different.
Unrelated - but reminds me of my favorite quote by tenderlove:
"microservices turn function calls into distributed computing problems"
grug wonder why big brain take hardest problem, factoring system correctly, and introduce network call too
seem very confusing to grug
It's easy enough to copy and paste in a prompt, no?
Eval tests while giving general indicators might not be similar for each use case.
This does not really work well, if you don't know the complexity of the problem ahead of time and ensure all future conversations go to the same model.
Else, you break the cache by doing a round robin of the same conversation across different models. Likely you'll end up paying more than what it would've cost with a cache aware system
Replace "Show HN:" with "Advertisement:" ?
I'm curious if there is measurable value in diversity of thought, and if there's diminishing returns on a single models thought pattern.
For example, compute X tokens with model A, then feed those into model B, etc. to get chain of thought through a diverse set of mdoels rather than chain of thought through a heterogeneous chain.
Humans seem to strongly believe echo chambers are bad. Are LLMs the same?
there are and this what you optimize for. ensemble learning has a long literature on this. you want models that have the most diverse pool of capabilities so they complement each other. in verifiable tasks or classification this is straightfroward but a bit tricker in open ended text gen or CoT. this is one of our core research question
My experience is mostly around either ensembles of weak learners that are bootstrapped to create a strong learner, or around boosted learners where you're training on residuals. With CoT you're sort of adding entropy to your answer, but you're also possibly getting a more thorough answer?
I'm curious though if these training methods are convergent or are models actually different; just like how in the stock market people think they're "diversified" but the truth is their exposure is likely much more risk correlated than one might think.
In certain situations, one right answer is better than a committee discussing the problem, but in others its sometimes nice to have some alternative methods of solving something. Fun project nonetheless.
My approach to using multiple models has been less about CoT but more about time to first token, and how you can use a small model to start interacting with the user while in parallel the more complex model is building a larger more complex thought. My work on this was primarily for voice backed interfaces before the voice models became quite a lot faster.
we did it with 1/10 the cost with Prometheus-1.0 and further https://trustedrouter.com/blog/prometheus-2-new-draco-state-...
Seems similar to Openrouter Fusion - https://openrouter.ai/docs/guides/routing/routers/fusion-rou...
Fusion is a totally different approach, though similar in the sense that it leverages different models.
Fusion generates many replies then synthesizes. This adds a ton of latency and cost, so it's going to be better only for cases where you're willing to wait a lot and pay a lot more.
Routers (like this project) are a different thing, they can theoretically improve performance and cost at the same time without increasing latency much. I'm a bit skeptical though, since knowing which LLM is going to be better on a cost adjusted basis is hard (see https://artificialanalysis.ai/models/capabilities/coding?cos..., where the cost per task vs. performance is not what you expect, for example comparing Qwen 3.7 Max to GPT Sol.
A project I'm working on is aimed at improving performance without added latency but from a different angle. Instead of waiting for all replies for synthesis (like OpenRouter Fusion), it streams the "best" reply immediately (using a router to pick the best model) then synthesizes with emoji reactions and optional replies from the background models. It's free to use here with no login: http://pellmell.ai
Sounds like another token router. Not sure what is the benifit compared with codex, I used 5.6 Sol, codex already route some of the task to luna not sol.
Thanks for checking the individual rows. HumanEval+ is one small code slice, not the whole basis for the launch claim. The public evaluator currently contains 907 rows across seven benchmark families, and matched SWE-bench Verified and BigCodeBench runs are the next code evidence being added.
You also found a real UI bug: the inspector should show both stored answers and currently does not in some rows. We are fixing that.
On the row you reran: the page records a frozen matched run. It does not claim that Fable is incapable of solving that prompt on another run. We are adding repeated matched trials and making run count and variance explicit. Your rerun is exactly the kind of external check the row-level page is intended to make possible.
The broader result remains: Echo is competitive with Fable across the evaluated task mix at materially lower measured inference cost. We are filling in the harder agentic-code evidence now rather than asking anyone to infer it from HumanEval+.
This reminds me on OpenRouters report that combining multiple different models gave comparable performance to Fable 5. I think this approach has lots of potential. Maybe OpenAi was ahead of its time with GPT-5 (it being a router to different models rather than just being one new model)
You might be interested in a project that I'm working on, which is kind of like OpenRouter Fusion, but instead of waiting for all models to synthesize, we stream the best model immediately and background the rest. The background models then reconcile with an emoji reaction and optional reply. It gets similar results to Fusion and is a lot faster! There's a free version that leverages open weight models here: http://pellmell.ai.
I like that I was able to test without an account and my prompt (What's the feasibility of a NetNavi IRL, along with a PErsonal Terminal? Create a document that outlines how to make this happen.) makes me think Pellmell is gonna be a research tool I'll continue to use.
This is really cool, super fun!
OpenAI's router chooses between models of different sizes, which are still trained on roughly the same data. Its purpose is to reduce infra cost for OpenAI for simpler queries. No need to pay for GPT 5.6 Sol inference for "Hello" prompt.
I think approaches like this have potential. Only time will tell. This reminds me of the mixture of experts taken by deepseek r2 (I think it was r2, at least), but less specific models I guess.
I have often wondered how tools like GHCP choose the best model for the job when set to "auto".
So the word security or any topic related to it is mentioned and it flips to an older gen model? Fable is nearly useless now it you do anything around auth.
I can get it to write win32 unsafe rust code.
I can't get it to review win32 unsafe rust code.
Make it make sense.
I'm not an expert on this, but this sounds a lot like a larger-scale MoE (Mixture of Experts) type of architecture.
As I understand it, in an MoE model, you essentially have hundreds of smaller sub-models ("experts") that are good at different tasks, and for every generated token, a single "master" model chooses which ones are most relevant to participate, and you only activate them.
In MoE systems the routing decision is made per-token, not per prompt or task. It’s one of ML’s many confusing naming conventions.
Even more confusingly, there are older pre-LLM MoE systems which ensemble and pool the predictions from multiple sub-components. For example in a random forest you could take the majority vote of the decision trees or the average of their numerical predictions.
After that, we developed neural net architectures for predicting a single thing like whether the user will click on your ad. An MMoE is in the same family.
And so now we are at massive MoE networks for LLMs which have similarities with MMoE in that the “decision” is about the very next token to predict.
"In MoE systems the routing decision is made per-token, not per prompt or task."
Have there been experiments with doing it per task? Like, "oh this is python project, use this model" "oh this is about writing fantasy, use this"?
It’s a good idea. The results probably depend a lot on how close your task is to the benchmarks though.
I think OpenAI already has (had?) a feature like this called “auto” mode for thinking.
The "ensemble of models" is nothing new. It's just that it's not really a moat that can be monetized. In the end you're always paying for something. You can engineer an elaborate harness with multiple models but it is not going to particularly solve a novel problem that the frontier models can with the same level of efficiency. You're saving money and paying with time. You're going to pay with something one way or another no matter what.
The frontier providers aren't dumb. They charge what they charge because they know this. If you think Fable is too expensive then the type of problems you are solving don't demand that level of capability.
If you are working on something cutting edge, something truly novel, the cost of frontier AI is well worth its price.
With all that being said. No one is going to complain if we can get the same capability at a lower cost. And I mean true parity. Not trading space for time.
There's so many people reimplementing Sakana fugu from its two ICLR papers but no open source version of it.
Intuitively, your savings depend heavily on how hard the tasks are in the first place. If you have a base rate where 99% of your tasks can be routed to a cheap model, yeah, you can save a ton by not using Fable for that.
So “1/3 the cost” really depends.
Sometimes expensive models are cheaper on easier tasks because they use fewer tokens, too.
Yeah apparently Opus and Sonnet are like that.
Yep! There are a lot of models like this: https://artificialanalysis.ai/models/capabilities/coding?cos...
I’m very confused on what this is, my initial thought was “oh nice, open source router.”
I go to the website…and it’s a sign up. I expected a repo. Otherwise how do I use it? As a SaaS? Yeah right.
Oh well I guess at least the benchmarks are good…I find the benchmarks and many are either not present or are not what the title claims.
My main question is how this has so many updoots from HN, probably the passerby not looking closer for sure.
I mean no offense and I really do wish you best on this, but it seems like what we used to call back in the day, vaporware.
Can't sign up: `too many authentication attempts`
looking at it now
I can't sign up: too many authentication attempts
looking at it now
What signals does Echo use to decide which models should participate before it has seen their outputs?
I have been using this : https://magnitude.dev/ for a while now. Is it something similar you are doing? I would love to have something that would connect to my codex, Claude, and opencode subscription rather than having to make a new subscription.
> Fable-level results at 1/3 the cost using open-weight models
But we get ~$2500/mo worth of Fable credits for $200/mo on Anthropic pan? I'm still confused why people (who don't have to use API billing) are chasing open weight models based on cost.
Enterprise plans pay API costs, they're only subsidizing individual accounts and that's only because they have to compete with open-source models.
When they are successful at making those illegal/inaccessible, both Anthropic and ChatGPT are going to rip the band-aid off and inference will only be sold to those that can afford it.
The good ole American way.
> When they are successful at making those illegal/inaccessible
This would be like trying to outlaw Linux or peer-to-peer file sharing. It's technically possible to write and pass a law, but it's basically impossible to enforce it.
Enough to make it a non-started at the organizations that pay their bills. Everyone else isn't big enough to matter.
Going to be an interesting world where big enterprises have to spend 100X the cost for the same value of AI as startups and small businesses.
I think they call that inflation :).
Because that is a short term solution, it won’t be offered forever. Large organisations have to purchase credits at $/tokens. Eventually everyone else will too.
This is what OpenAI and Anthropic are trying to make everyone believe. Most accountants will flinch at this (they already are).
The $200 odd plans are already out of reach of many, many people.
The attrition of customers if they were to get rid of these subscriptions plans would be untenable.
I think you are looking at it incorrectly. No business is buying individual accounts, because if they do, they open themselves up to considerable risk.
The $200 plans are priced so that the power-users use them and then advocate about how great the product is. If you're buying a $200 plan, you're not doing it because of the price point but rather because of the amount of work it is doing for you.
All companies we interact with have 200 plans.
Lots of businesses are buying and using these plans. Basically every small business I interact with.
You may want to consider the incomes of developers outside the US, students, unemployed. $200/month is a lot to a lot of people.
The point still stands. The chinese labs don't have super discounted plans, so if the price per task benchmarks[1] are correct, and we apply the discount, you'll actually be paying more by using cheaper chinese models and this technique.
https://artificialanalysis.ai/agents/coding-agents#artificia...
The point doesn't stand, and this comment appears to be spam.
They can still write code.
I think most people assume the subsidized plans will go away or get more limited eventually. They are basically a loss leader and a marketing cost that is very flexible and easy to change w/o directly impacting their primary customers.
doesnt exist for enterprise plans
M-o-MoE
By the way, is it just me or opus 4.8 is much better at some programming tasks than fable? I've been really disappointed lately, i stopped using it evem though it's "premium on my subscription"
It's just you. Fable 5 is consistently better than Opus 4.8 at literally every single level, for me. (I always use both at xhigh, for reference.) I could go down a laundry list of various issues I have with Opus that I don't have with Fable. For me, Opus 4.8 < GPT-5.6 Sol < Fable 5.
Plus Fable is way less annoying to talk to than Opus 4.8. Opus 4.8's writing style is absolutely insufferable. Fable has some of the same quirks but it's way less bad.
We see no difference between opus 4.8 and fable except that fable is slower. So +1 for not just GP. We use neither interactive, just via our own tooling so the ‘talk to’ doesn’t apply and we use it 24/7. We currently run 25% of tasks on both fable and opus and the rest only opus. The 25% are being code reviewed side by side and we do detect when fable switches to opus for ‘security concerns’ nonsense. Opus generally finishes sooner, results are similar quality.
such a scam, there is only one fable-like model, that somewhat behind, it cost half, not 3x. so from here you can stop reading.
They're obviously not claiming this is a new model that is fable-like at 1/3. It's a router that saves money by only using Fable when necessary. I don't think you should be calling someone's post a "scam" without doing a minimum of research. (I'm not associated with the company, but very interested in this.)
That’s basically the same idea IBM advertises with Bob?
This looks different, it's not a main agent delegating to subagents, it's a router.
I don’t find the recent spate of blog posts and systems delegating and combining LLMs to get better performance particularly interesting. Especially given that anyone who’s taken an ML 101 course has learned about ensemble methods.
While an LLM isn’t what you’d traditionally consider a weak learner, the theorems on learning systems clearly point to them being so in this context. The feigned surprise at combining them to yield better results seems disingenuous.
Even so, the work to predict which models are best suited for which task, how to delegate, and how to combine their outputs is interesting, especially if you’re placing a cost minimization objective on it. That said, this isn’t too far off from what many AI labs are already doing.
The entirety of "agents" and tool calls is a process of combining LLMs to get better results. Is it the same LLM in many cases? Yes. But it doesn't have to be.
It's the natural move that happened after people realized you couldn't throw away half a century of AI research.
Most of these focus on costs. But it is simply the case that the one-shot output did not scale for harder problems on workflows.
how does this differs from OpenRouter fusion?
how does this differs from Sakana Fugu?
enjoying how people are re-discovering ensemble methods.
Fable-level, yea, but can it run gstack?
but u wouldnt get caching savings
You definitely still would, but you need to pay the full input cost twice, so the equation really depends on how much first message vs repeat message matter.
Ttl 1 hour maybe. 5min? Never
Might want to rethink the name to avoid an Amazon issue.
If you copy Perplexity, they let you have the first few rounds of chat for free to get you going before asking to sign up.
Is this yet another Sakana Fugu / OpenRouter Fusion?
Looks like it but with open-weights only.
"Backed by YCombinator"
https://www.ycombinator.com/companies?query=tracerml
I don't see it?
Not all YC companies have launched publicly yet but I am a current YC founder and I can confirm they exist in the internal directory.
You're right to be skeptical; I saw a totally fake case of this just yesterday.
But in the present case, they're just a startup in the current batch.