The thinking tax is not in the multiple

Q: Do reasoning models cost more then?

Not because of the multiple. On the workload used here — 6M input and 2M output a month — the most expensive reasoning model costs $140 and the steepest-multiple one costs $3. What decides the bill is the input rate, which differs by 100× between them, plus how many output tokens the thinking actually generates.

Q: Which model has the highest output/input multiple?

qwen3-30b-a3b-think and qwen3-30b-a3b-thinking-2507 , at ($0.10 per 1M input, $1.20 per 1M output). Both are Alibaba. They are also two of the cheapest models in the index on a real workload, which is the point of this page.

In short: Reasoning models hold the steepest multiples in the index, but their median multiple is 4.00×, identical to 4.00× for everything else. The tax is not in the multiple. It is in the input rate and in how many output tokens the thinking generates: the steepest one in the set is one of the cheapest models you can run, while a model sitting exactly at the median multiple is among the most expensive. 3 of the 3 same-model thinking and non-thinking pairs are priced identically, so the non-reasoning twin is a behaviour switch, not a cheaper tier.

The standard story about reasoning models goes like this: a reasoning model thinks before it answers, the thinking is emitted as output tokens, and output is the expensive direction So reasoning models cost more.

I went into this expecting to measure exactly that. I had the rates for every model in the index and a clean way to split the reasoning ones out, and I predicted their output-to-input multiples would sit well above the rest.

They do not. On the median they are identical to everything else, and the number I built the article around turned out to be an artefact of how I was counting.

You need a key before the code below runs. Create an account, generate a key, and copy the base URL (https://aicomp.ai/v1). Create one free →

Check current rates → Free to sign up · $1 minimum top-up · No prepayment

What the two populations actually look like

The index holds 223 models, cleaned the same way every ranking on this site is cleaned. Of those, 21 carry a reasoning marker in their identifier, and 202 do not.

	Reasoning models	Everything else
Models	21	202
Median multiple	4.00×	4.00×
Mean multiple	4.62×	4.48×
Upper quartile (p75)	4.00×	6.00×
Steepest	12.00×	10.00×

The medians are the same number: 4.00× for reasoning models, 4.00× for everything else. Half of all reasoning models are priced at exactly the 4× that everyone quotes, and are not surcharged at all.

The means are close enough that the gap is not a finding: 4.62× against 4.48×. I want to be blunt about this because I published the opposite number first. An earlier version of this page said reasoning models averaged 5.52× against 4.31× for standard models. That was measured over the raw gateway catalogue, which carries dated snapshots of the same model at the same rate — qwen3-30b-a3b-thinking-2507 alongside qwen3-30b-a3b-think, and several more of the same shape. Those duplicates inflated the reasoning mean and pushed it in a direction the cleaned data does not support. The corrected number is a rounding error, and reporting a rounding error as a surcharge was wrong.

Then there is the row that runs the other way. The upper quartile of reasoning models sits at 4.00×, against 6.00× for standard models. Above the median, reasoning models are the flatter group. OpenAI's whole gpt-5 line sits at 8.00×, and none of those models are sold as reasoning models.

So the population comparison gives you nothing. Reasoning models are not priced with a steeper ratio. What they do have is the top of the range: the steepest multiple in the index, 12.00×, belongs to a reasoning model, against 10.00× for the steepest standard one.

The steepest multiple in the index costs three dollars

A day of work on this page's workload is 300k input and 100k output tokens — 6M input and 2M output over a 20-day month. Here is what that workload costs on five reasoning models, three chosen for the steepest multiples and two for the largest bills:

Model	Input / 1M	Output / 1M	Multiple	Monthly cost
`qwen3-30b-a3b-think`	$0.10	$1.20	12.00×	$3
`qwen3-30b-a3b-thinking-2507`	$0.10	$1.20	12.00×	$3
`qwen3-235b-a22b-thinking-2507`	$0.12	$1.15	10.00×	$2.99
`o3-pro`	$10.00	$40.00	4.00×	$140
`deepseek-r1-h`	$2.00	$8.00	4.00×	$28

Read the first and the last row together. qwen3-30b-a3b-think carries 12.00×, the steepest multiple anywhere in the index, and it costs $3 a month here. o3-pro carries 4.00×, precisely the market median, the most ordinary multiple there is, and it costs $140.

That is a 47× difference in bill between two models whose multiples differ by 3×, and it runs in the opposite direction from the multiples.

The reason is that a multiple is a ratio, and a ratio carries no information about scale. Alibaba charges $0.10 per million input tokens on qwen3-30b-a3b-think, so a 12× multiple lands output at $1.20. OpenAI charges $10.00 for input on o3-pro, a hundred times as much, so a 4× multiple lands output at $40.00. The steepest multiple in the index is attached to one of the cheapest models in it.

This is the same trap as comparing models on input price alone, wearing a different hat. Neither the input rate nor the multiple tells you what a workload costs; only the two together do.

Where the tax actually is

If the reasoning surcharge is not in the rate card, it is in the token count, and that is worth being precise about because it changes what you can do about it.

A reasoning model emits its chain of thought as output tokens before it answers. You are billed the output rate for the model talking to itself. On a workload that produces 2M output tokens a month, a model that spends 3,000 tokens thinking before each 500-token answer is generating several times more output than you ever see, and every one of those tokens is billed at the output rate.

That is a real cost and it is specific to reasoning models. But note what it is not: it is not a surcharge on the rate. It is the ordinary output rate applied to a larger volume. Two consequences follow.

The first is that the multiple gives you no warning. qwen3-30b-a3b-think has the steepest multiple in the index and the smallest bill in this table, because the thing the multiple multiplies is a very small input rate.

The second is that on a reasoning model, output optimisation is worth more than anywhere else. Cutting the thinking tokens is worth 12× the input rate on that Alibaba model and 4× it on o3-pro, and on either one it is worth more than trimming the prompt.

You may be paying for thinking you turned off

The most actionable thing in the data is not a multiple. It is a pair of rows.

Three vendors' models are published in two variants, one that reasons and one that does not, told apart by a reasoning or non-reasoning suffix. In the index there are 3 such pairs, and 3 are priced identically on both input and output:

grok-4-1-fast-non-reasoning            $ 0.1 in / $ 0.25 out
grok-4-1-fast-reasoning                $ 0.1 in / $ 0.25 out

grok-4-20-non-reasoning                $ 0.625 in / $ 1.25 out
grok-4-20-reasoning                    $ 0.625 in / $ 1.25 out

grok-4-fast-non-reasoning              $ 0.1 in / $ 0.25 out
grok-4-fast-reasoning                  $ 0.1 in / $ 0.25 out

Two identifiers, two different behaviours, one rate card. All 3 pairs belong to xAI.

There is a methodological point hidden in here that I had to fix mid-analysis. If you count grok-4-fast-reasoning and grok-4-fast-non-reasoning as two reasoning models, then identical pricing between the two groups becomes true by construction You put the same two rows in both columns. The reasoning set here excludes the explicitly non-reasoning twins, so the 21 models counted above are ones sold as reasoning models on their own terms.

The practical consequence is narrow but worth money. The non-reasoning variant is not a cheaper tier. It is a behaviour switch with the same price tag. If you were routing traffic to it expecting a saving on the rate, that saving does not exist. What it does save is the thinking tokens, and on short mechanical work that is the part you were overpaying for.

What to do with this

Stop ranking models by output multiple. The 12.00× model costs $3 a month and the 4.00× model costs $140 on the same workload. The multiple is a property of the rate card, not of the bill. Rank by the blended rate your own token mix produces — the calculator takes the split and does the arithmetic.

On a reasoning model, cut the thinking, not the model. Where you can cap reasoning effort, shorten the prompt that triggers it, or skip it on work that never needed it, that is where the money is. Switching to a different reasoning model with a friendlier multiple will mostly move you to a different input rate.

Do not expect a discount from the non-thinking twin. All 3 same-model pairs in the index are priced identically. Routing to the non-reasoning variant saves the thinking tokens and nothing on the rate.

If a workload never needed chain-of-thought, use a cheap standard model. Bulk classification, extraction over long documents, anything where the question is mechanical and the answer is short — those are cheaper on an inexpensive standard model than on either variant of a reasoning one.

Method and limits

Rates come from our own price index, read on 2026-10-22. The catalogue is cleaned before any of these numbers are computed, test entries, dated snapshots of the same model, namespaced and wildcard identifiers, and non-chat models are all removed first. That is the same cleaned set behind every ranking on this site, and it is 223 models.

Reasoning models are identified by thinking, reasoning, reasoner, r1, o1, o3 or o4 in the identifier, with explicitly non-reasoning variants removed.

That step is the weakest part of the method and worth naming: it is a naming convention, not a capability test. A vendor that reasons without saying so in the identifier is counted as a standard model, and there is no way to catch that from rate data alone. The 21 is a floor, not a census.

Three further limits. These are gateway rates rather than official list prices, and few of the models in the index publish a comparable USD list price of their own. The monthly figures are for one specific workload shape (6M input and 2M output) and a different token mix would move all of them; a reasoning-heavy mix would move them most. And the twin pairs are whatever vendors currently publish, which for this particular trick is one vendor.

FAQ

Do reasoning models have a higher output multiple?

On the median, no. Reasoning models sit at and everything else at , the same number. They hold the extremes: the steepest multiple in the index is on a reasoning model, against for the steepest standard one.

Do reasoning models cost more then?

Not because of the multiple. On the workload used here (6M input and 2M output a month) the most expensive reasoning model costs $140 and the steepest-multiple one costs $3. What decides the bill is the input rate, which differs by 100× between them, plus how many output tokens the thinking actually generates.

Does turning thinking off save money?

It saves tokens, not rate. All of the same-model reasoning and non-reasoning pairs in the index are priced identically on both input and output, so the non-reasoning variant is a behaviour switch, not a cheaper tier. The saving comes from not generating the thinking tokens in the first place.

Which model has the highest output/input multiple?

qwen3-30b-a3b-think and qwen3-30b-a3b-thinking-2507, at ($0.10 per 1M input, $1.20 per 1M output). Both are Alibaba. They are also two of the cheapest models in the index on a real workload, which is the point of this page.

Should I avoid reasoning models to cut costs?

Not for reasons of the multiple. If a workload does not need chain-of-thought, a cheap standard model beats both variants of a reasoning one. If it does need it, the lever is the volume of thinking, not the rate ratio.

The 4× rule in model pricing, why the median multiple is 4.00× across every price quartile
Why output dominates the bill — how the two directions are billed
Official vs gateway prices — the two price columns on this site and which models are verified
Cheapest models by workload (no cheapest model), only a cheapest token mix
Cost calculator — put in your own split and get the blended rate
Price index 2026 — the full dataset behind this article

Get API access