Seven times cheaper for roughly three-quarters of the frontier's intelligence index: cost-per-task is now the deciding metric

Kimi K3.1 Pro scores 46 on the Artificial Analysis Intelligence Index against 65 for Claude Fable 5.2 — at $3.68 per task versus $26.54. When the gap between good enough and the frontier costs seven times as much, routing by request difficulty beats standardising on one model.

Two numbers do most of the work here, and they come from two different documents, so it is worth keeping them straight from the first line.

On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro — the strongest open-weight model in that comparison — scores 46. Claude Fable 5.2 scores 65. That is a gap, not a victory: the open-weight model does not win the index. The more consequential number is the second one. In the figures Hugging Face published on 2026-11-16 alongside its OpenRouter integration announcement, cost per task is $3.68 for Kimi K3.1 Pro against $26.54 for Claude Fable 5.2 — about one seventh (https://huggingface.co/blog/openrouter).

Put those together and you get the trade: roughly 71% of the frontier’s intelligence-index score, at roughly one-seventh of the price per task. The index scores themselves are Artificial Analysis’s, as cited by that announcement; the cost-per-task figures are quoted by Hugging Face (https://artificialanalysis.ai).

The two “cheaper” ratios are not the same ratio

This is worth naming explicitly, because the two ratios from the same announcement get quoted side by side and mean different things.

  • Cost per task: $3.68 versus $26.54 — roughly 7x. This is the ratio in the headline.
  • Average price per million tokens, open-weight against closed, from the same announcement: $0.28 per million input tokens and $1.62 per million output tokens for open-weight, against $1.27 and $6.36 for closed models — roughly 4x on input and roughly 4x on output.

Same source, same date, different measurements. Cost per task folds together the price of tokens, how many tokens the model needs, how often it has to be retried, and how much of the answer you actually use. Per-million-token pricing tells you what a single token costs and says nothing about how many you will need. A model that is 4x cheaper per token can still be the worse deal if it takes three times as many tokens to finish the same job. When a task-level number and a token-level number sit next to each other, they are not comparable, and most of the confusion in this market comes from mixing them.

The traffic already voted, and it voted cheap

None of this is theoretical, because demand already moved.

OpenRouter’s own published figures for that week show the most-used model on the platform was the open-weight DeepSeek V4.1 Flash 1017, at 59.2 trillion tokens (https://openrouter.ai). One model, one week, 59.2 trillion tokens — and the most-requested model is not the one sitting at the top of the intelligence index. The volume is not piling up at the frontier. It is piling up at the cheap end of the curve, and the OpenRouter announcement carries the same traffic data alongside its cost figures (https://huggingface.co/blog/openrouter).

That is a distribution, not a ranking. It says nothing about which model is better at any particular task, and it certainly does not mean the interesting work happens at the cheap end. It means the requests people actually send in bulk are mostly the requests that never needed the frontier — which is exactly the condition that makes routing worth the effort.

What that implies for routing

Here I want to be careful about the line between what a source says and what I am reading into it.

What the sources establish: a measurable capability gap (46 against 65 on Artificial Analysis’s index), a measurable price gap ($3.68 against $26.54 per task), and a traffic distribution already concentrated on an open-weight model.

My reading of it: once the gap between “good enough” and “the frontier” is that wide, standardising on a single model for every request stops being a sensible default. That is the analysis, not a claim from the announcement — the sources do not tell you to route by difficulty, they just make it worth doing.

The practical shape of it is simple. Sort requests by how hard they actually are, send the routine ones to the cheap end, and reserve the frontier for the ones where the extra 19 index points buy something. A team that sends every request to the frontier model is, on the arithmetic above, paying frontier prices on the large majority of requests that never needed frontier intelligence. That decision is also cheaper to make than a wholesale migration: it is a policy in front of the model, not a replacement of the model.

Which requests belong where is not a question a leaderboard can answer for you. It is a question you answer by instrumenting your own traffic: take a sample of recent requests, mark the ones a weaker model would have handled just as well, and look at what share of spend they represent. If that share is large — and the 59.2 trillion token figure suggests the industry average is large — then the routing policy pays for itself before you optimise anything else.

One more figure, attributed to OpenRouter rather than to me: the same 2026-11-16 announcement reports that its rebuilt routing cut capacity errors by 46%, improved time to first token by 16.9% and improved throughput by 16.3%. Those are OpenRouter’s reported numbers for its own system. They are not the argument here, but they are a reminder that per-request economics are shaped by the routing layer as well as by the model underneath it.

What would change the conclusion

Some caveats, stated plainly.

An index score is a summary. A gap of 46 against 65 is a real capability difference on the tasks that index measures, and “about 71% of the frontier” is not the frontier. On the genuinely hard request — the one where the answer is the whole job — the cheaper model may simply be wrong, and a wrong answer costs more than any token saving.

Cost per task is also a measurement, and how it is measured matters: the task set, the prompt, the retry policy and the success criteria all move it. Treat the 7x as a shape rather than a constant, and re-measure it against your own workload before you reorganise anything around it.

There is a second-order risk in leaning on the cheap end. The cheapest-per-task model this quarter may be the one most likely to be repriced, deprecated or capacity-constrained next quarter. A routing layer is also a hedge; single-model standardisation is not.

And the provenance is 2026-11-16. These are the figures in the announcement published that day and the Artificial Analysis index it cites. Prices, indexes and traffic shares all move; this is a snapshot of one specific week, not a standing fact.

None of this is an argument against frontier models. It is an argument against using one model for every request.

The decision, if you only take one thing

The old question was “which model is best?”, which is genuinely hard to answer and expensive to guess at. The cost question is easier, and at these ratios it is more decisive: measure what a task costs you, not what a token costs you, and route on the answer.

The 7x figure is the reason to go looking. The 59.2 trillion tokens are the reason to believe people have already gone looking.