Stop paying flagship prices for easy questions.

Fugal is built on a simple idea: don't send every question to the most expensive model. A tiny 0.6B router reads your question, predicts the odds every candidate model would get it right, and calls the one that earns it — once. No retries, no second opinion, no hidden extra spend. You see the predicted odds and the real price of every choice.

How a question flows

One question, two steps. That's the whole pipeline.

The demo's default is Route mode: step 1 laid bare — the full ranked board, every model's predicted odds and estimated price — always free, no model called. Answer (sign-in) sends the question on to the routed model and streams the reply.

The decision rule, stated plainly

The router doesn't classify your question by topic and look up a model — there's no "maths goes to model A" table anywhere in it. For every model on the board it predicts a single number: the probability that model answers this question correctly. Then it scores every candidate on

utility = p_correct − λ × price

and calls the highest one. λ (lambda) is the exchange rate between accuracy and money — the one knob that decides how many dollars a point of expected accuracy is worth. Turn it up and the router gets stingy; turn it down and it buys certainty at almost any price. Everything else is measurement.

How can a model that small know that?

It's the fair objection, and it's worth answering directly: a 0.6B model can't answer a hard question — so how could it possibly know whether a frontier model will?

Because judging difficulty is a much easier job than solving. You don't have to be able to evaluate a contour integral to recognise one on sight, or to see that "what's the capital of France" and "prove this convergence" are not the same order of problem. The difficulty of a question is largely legible on its surface — in its structure, its vocabulary, how much has to be held at once, how many steps it implies — and all of that is readable long before anyone works out the answer. The router never attempts your question. It reads it and estimates how much work it takes, which is why one forward pass of a small model is enough, why it generates no text, and why it costs nothing to run.

That's the argument for why this is possible. Whether our particular router does it well is a separate question, and a mechanism being plausible is not evidence that it works. It was measured — the result is below.

You can't pick the model — that's deliberate

There's no model selector, and no way to pin one. Every other AI product puts a dropdown in front of you; this one takes the decision away, because the decision is the product. A picker would also quietly break the thing you came for: choosing a flagship for everything is exactly the overpaying this exists to stop, and choosing the cheapest is exactly the false economy. What you get instead is the full board on every question, free — every model's predicted odds and price, and which one was taken. You can always see the choice and why. You just don't make it.

Where λ is set today, honestly: λ is tuned to weight accuracy heavily, so on a question two models are both likely to get right, it will still often take the pricier one. That's a deliberate default — being wrong costs a user more than a fraction of a cent — but it does mean the cheap end of the board is rarely chosen, and finding the setting with the best accuracy-per-dollar is open tuning work, not a solved problem. You can watch the rule operate on any question in Route mode, free: the ranked board you get back is this calculation.

And λ is a choice, not a measurement. Everything else on this page is something we measured; λ is something we picked. Someone has to decide what a point of accuracy is worth in dollars, and today that someone is us — which is exactly the kind of decision a company should not get to make quietly on your behalf forever. Who sets λ, and how it changes without us, is the subject of the Network page.

The numbers — and what they measure

Every model is scored on the same 320 questions across five domains — GSM8K, MATH-500, AIME-25, HumanEval and MMLU-Pro — and charged at its real API price. That board measures the models, not us. It stays true no matter what we ship, which is why it's the evidence everything else here rests on.

And it says something specific: the flagship is not the top of this board. A model costing several times less matches it to within a handful of questions out of 320 — which is parity, not victory; that margin is noise. The cost difference is not noise. The flagship isn't even on the Pareto frontier here: something is both more accurate and cheaper than it.

Below that the floor drops further than most people expect. The cheapest model on the board still answers about nine questions in ten correctly, for a small fraction of a cent — roughly two orders of magnitude below flagship pricing, for about five points of accuracy. So paying flagship prices on every question buys headroom most questions never use, and paying the floor on every question loses the ones that needed it. Neither is the right default. The gap between them is the entire reason to route — and it is why a router has to be able to reach models from more than one vendor to be worth anything at all. The current names and figures are in the tiles below, and in full on the Frontier board.

Every per-model figure, the Pareto frontier and every domain breakdown are on the Frontier board.

And what Fugal itself scores

The board above measures the models. Measuring the router is a separate experiment, and it has now been run — 110 fresh questions, route once, call the chosen model once, nothing else. Against gpt-5.5, the flagship:

Parity is the honest word, and it cuts both ways. The 1.8-point gap sounds like a loss and isn't one: at n=110 the two systems disagree on only four questions, and an exact McNemar test puts p = 0.63 — far from any threshold at which you'd call them different. The cost difference is not in that grey zone. Same accuracy, within measurement error, for roughly a sixth of the price.

Stated as a limit rather than buried: 110 questions across three domains is a small sample with little statistical power, and this is a single snapshot. It tests how the router does today; it does not test the claim that matters most to us — that a router stays current while a fixed choice quietly rots. That needs a measurement across time, and nothing here has made it yet.

Why this can't come from a model lab

When a model lab picks a model for you, it picks from its own catalogue. That isn't a gap they've been slow to close — it's structural. No lab is going to route you to a competitor, however obviously better that competitor is on your question, because their business is selling you inference on their own models. A router built inside a lab can only ever answer "which of ours?"

The board above says that's the wrong question. On these 320 questions the most accurate model is Qwen's and the cheapest is DeepSeek's — and the Pareto frontier itself spans two vendors, not one. Any router that can't cross vendor lines is choosing from a deliberately narrowed menu, and the narrower the menu, the less a router is worth. Neutrality isn't a feature we added; it's the only position from which the question can be answered honestly, and it's not one a lab can occupy.

Prior art — and where we differ

This is not a hunch we had alone. In June 2026 Sakana AI published Fugu, a family of orchestrator models trained to marshal a pool of frontier LLM agents, and reported state-of-the-art results against every publicly accessible model across coding, reasoning and agentic benchmarks. The name here is partly a nod to it — and partly to the older word it already was: fugal, in the manner of a fugue, many independent voices coordinated into one piece. That is the whole idea in one adjective — and it sits one letter away from frugal, which is the other half of the point. The core finding is one we're building on rather than disputing: a system that picks among models can beat any single model in the pool. A lab with far more resources than us arrived at that independently, which is better evidence for the direction than anything we could say about it.

Two differences, both deliberate. The first is which axis. Fugu is aimed at the ceiling — squeezing the last points of capability out of the strongest agents available, with Fugu-Ultra spending extra latency and extra calls to get them. Fugal is aimed at the floor: the same mechanism, pointed at accuracy per dollar, over a pool that runs from flagships down to models costing a hundredth as much. Those are different products from one insight, and the cheap end is where the gap on our board actually lives.

The second is who owns the router, and it matters more. Fugu is a closed model from a single lab: that lab decides which models enter the pool, when the thing is retrained, and — eventually — how its own future models are weighed against everyone else's. That is the same structural problem as the section above, one level up. It isn't a criticism of their engineering; it's a claim about what no lab can be trusted to arbitrate. Our answer is to not own the router either, which is the whole subject of the Network page.

Routers rot — and a promise is a weak fix

Model prices and capabilities shift almost monthly. A router trained on last month's prices slowly makes worse choices — it rots. Today Fugal answers that by re-measuring on a cadence: re-price the models, score the new ones, recalibrate the router, publish a dated snapshot. That works — but be clear about what it is. It's a promise, and only ever as good as our diligence. A router nobody is paid to keep fresh eventually isn't.

Honest by design

Every number here traces to an archived, re-gradeable artifact — gate_matrix_2026-07.json, re-scored offline by figs/emit_frontier.py — so any figure on this page can be recomputed from the record rather than taken on trust. Numbers that described an earlier research prototype have been removed rather than re-labelled: a measurement of a system you can't use is not evidence for one you can. Ties and failures are reported plainly; corrections are logged rather than overwritten. Answers come from third-party models via OpenRouter, sign-in is optional (guest chats live only in your browser; an account adds cloud sync), and every account is rate-limited.

FAQ

What does it cost?

Routing is always free — the router runs one local forward pass, no model is called, and there's no daily cap on it. Answers are free for your first 10 a day (or $0.02 of model spend, whichever binds first), which needs an account but no card. Past that you spend prepaid credit, in $10 / $25 / $50 packs. There's no subscription, no monthly minimum, and credit doesn't expire. Full breakdown on the API page.

How is a paid answer priced?

A paid answer is priced from the model the router selects for your question. Since that choice is made per question, the per-answer price tracks the question rather than a flat rate, and every routing decision shows its estimated price before you spend anything.

What's the difference between Route and Answer mode?

Route shows step 1 only: the full ranked board — every model's predicted odds of getting your question right, and its estimated price. Answer actually sends the question on to the winning model and streams the reply, with its real cost shown.

Why do I have to sign in for Answer mode?

Answering calls a real model, which costs real money — so it has to be attached to an account that can carry a free allowance and a credit balance. Route mode never needs an account.

Can I use it from my own tools?

Yes. Mint an API key in Settings and point any OpenAI-compatible client — the OpenAI SDKs, Cursor, Aider, Claude Code — at https://fugalai.com/v1. Both the OpenAI and Anthropic wire formats work, and the same free tier and credit apply. See the API docs.

Where are my chats stored?

Signed out: in this browser only — nothing leaves it. Signed in: synced to your account under row-level security, so only you can read them. Export or delete everything at any time in Settings.

Are my chats used for training?

No — not unless you explicitly opt in. The toggle is off by default and lives in Settings; details in the Privacy Policy.

Which model actually answers me?

Whichever model the router predicts is most likely to get your question right — reaching for a flagship only when the question demands it. The board always shows the pick, its predicted odds, and its price — nothing about the choice is hidden. You can't override it, in the app or over the API: there's no model parameter to set and no way to pin one. Routing is the product, and the reply always reports the model that actually answered, so your logs stay truthful.