ModelCensusopen-source ai reliability harness

You don't pay the meter rate. You pay the fare.

token cost vs. task cost

Blog26 Sept 2026tokenomicsindustrystrategy

The meter rate isn't the fare

Everyone compares price per token. You pay per task. On Artificial Analysis's own numbers, the gap between the two is where the budget goes.

This week someone dropped Artificial Analysis's intelligence-vs-cost chart into a trading group I'm in, with two green arrows pointing at the top right.

Everyone looks at the top right. I looked at the x-axis. It isn't price per token. It's cost per task, and those are different numbers.

Intelligence vs. cost per task0204060$0.05$0.1$0.2$0.5$1$2$5$10cost per task (log)intelligence indexMuse Glimmer (high)GPT-6 Luna (max)MiMo-V2.6-ProGPT-6 Sol (max)Opus 5.5 (max)
Redrawn from Artificial Analysis's published data (Intelligence Index v4.3.2, the 31 models in its default view that report a cost, read 26 Sep 2026). The dotted line and red dots are the best-value frontier. Look at the flat stretch: MiMo-V2.6-Pro scores 46 for 13 cents a task, and the next model up costs 8× more for one extra point.

Rate times distance

Think of a taxi. Price per token is the meter rate. Tokens per task is the route the driver takes. You pay the fare.

Meter rate$ per token× routetokens per taskThe fare$ per task
Token price vs. task price$0.05$0.1$0.2$0.5$1$2$5$10$0.2$0.5$1$2$5$10$20$50$100$ per 1M output tokens (log)$ per task (log)GPT-6 Luna (max)Qwen3.8 MaxFable 5.1 (max)GPT-6 Astra (xhigh)
Same data. The meter rate explains most of the fare (r² ≈ 0.7), but not all of it. GPT-6 Astra's tokens cost 8× Qwen3.8 Max's, and its tasks cost less than half as much, because it takes about a sixth of the tokens.

The dial you own

Same model, same price per token, four effort settings:

$ per task$1.34medium · 51$1.82high · 54$3.46xhigh · 56$5.98max · 58
Claude Opus 5.5 at each effort setting, with its intelligence score beside the setting. Tokens per task go from 26k to 119k. The price per token never moves.
medium → high$0.20 a pointhigh → xhigh$0.68 a pointxhigh → max$1.58 a point
What each extra index point costs per task as you turn the dial. The last notch costs 8× the first.

The price of a point

GPT-6 Luna (max)0.19¢MiMo-V2.6-Pro0.28¢Muse Glimmer (high)0.34¢Opus 5 (max)11.5¢Qwen3.8 Max11.9¢Fable 5.1 (max)14.3¢
Cost per task divided by intelligence score: the three cheapest points on the board and the three dearest. 76× apart.

So what do I do with this?

Budget in tasks, not tokens. Ask vendors for cost per task on your workload, then set the effort dial as low as the task will stand.

We learned this the hard way in our own census. The model that ate about a third of our bill didn't have the highest price per token. It just took the long way round.

One thing I can't settle: these are one lab's numbers, on its workload, read on one day, and it hasn't published the v4.3 weights. Your tasks will take a different route.

You don't pay the meter rate. You pay the fare.

When did you last see an AI budget line priced per task instead of per token?

Sources

Every figure here describes something measured and committed. See the measurements · read the method

Related

Open or closed is a label, not a risk rating
Open models trail the frontier by about four months and are showing up everywhere. I split my 20-model census by weights. The label barely predicted failure. The model did.
Every role on the chart has a next chapter
Gartner's new org chart merges twelve roles into six, and six into three teams. It reads like a list of cuts. It's a map: for each role, what AI takes, what stays human, and the first skill to learn.
Bad data used to cost accuracy. Now it costs tokens.
In traditional ML, a prediction had a fixed price and data quality moved accuracy. In GenAI it moves the bill too. One run showed both paradigms side by side.