The 4× rule in model pricing

Everyone in this industry repeats one number: output tokens cost about four times what input tokens do. It turns up in pricing docs, in cost calculators, in every post about cutting an LLM bill. Repeat something often enough and it stops being a claim and becomes weather.

It is also true. The median output-to-input multiple across 231 models in our index is 4.00×.

The interesting part is not that the number is right. It is why it is right, because the reason tells you something you can act on, and it is not the reason the repetition implies.

The multiple does not move with the price

If the output multiple were tracking the cost of generation, it would move with the model. Serving a token out of a large model consumes more of the expensive resource than serving one out of a small one, so the expensive end of the market should charge a steeper multiple.

I sorted the 231 models by input price and split them into quartiles:

The three cheapest quartiles sit on the same number. The most expensive quartile lifts to 5.00×, which I will come back to, but a gap of one turn of the multiple across a price range from $0.01 to $15 per million input tokens is not a cost curve. A cost curve moves. This does not.

Two spikes, and they are 4× apart

Look at the whole distribution instead of its middle:

<1x   0    #
1x    20   #####
2x    27   #######
3x    24   ######
4x    61   ###############
5x    45   ###########
6-7x  23   ######
8+    31   ########

The histogram has a hard spike at 4× — 57 models sitting at exactly 4.00×, not near it. Then the tail does something it should not do if it were measuring anything physical: there is a second spike out at 24 models sitting at exactly 8.00×.

That is the part that convinced me. A number arrived at by measuring compute lands somewhere like 4.37× or 7.81×, because the underlying quantities are not round. 57 models at exactly 4.00× and 24 more at exactly twice that is not what measurement produces. It is what happens when pricing teams pick a number.

Four is the one everyone landed on first, and eight is the one you go to when you want to charge more without looking like you are charging more, it is still a round number, and it is still legible as "output is expensive."

If it were a cost signal, vendors would not disagree

The strongest evidence is that vendors do not agree, and disagree by a lot. Same calculation, grouped by vendor, excluding anyone with fewer than three models in the index:

OpenAI at 6.00× at the top and ByteDance at 1.80× at the bottom is a 3.3× spread. These are not marginal vendors. The two largest model labs in the table differ before you compare a single rate.

What a coordination equilibrium looks like from the inside: once most of the market quotes 4×, a vendor quoting 7× looks expensive on every comparison table a buyer opens, and one quoting 2× leaves money on the table for no competitive gain. So everyone quotes 4×, or a clean multiple of it, and the number stops carrying information about the models.

This is worth money precisely because of what follows from it. If the multiple were measuring the cost of generation, there would be nothing to shop for It would be the same everywhere, like the speed of light. Because it is a convention set vendor by vendor, it is a lever, and it is a different lever from the model you pick.

Where the 8× cluster actually is

A median hides the shape, and the shape above 8× has a story attached to it that I want to correct. An earlier version of this page said the steep-multiple cluster was mostly reasoning models. It is not.

There are 31 models at 8× or steeper, and 16 of them are OpenAI's gpt-5 line, every one at exactly 8.00×. Only three are reasoning models. The steepest multiples in the index belong to Alibaba's thinking variants, which reach 12.00×, but they are a small part of the cluster, and the cluster itself is mostly one vendor's flagship pricing.

That is worth separating, because the two things call for opposite responses. A steep multiple on a cheap model is a rounding error on a small bill. A steep multiple on a flagship input rate is where the money goes. The thinking tax page works that one through properly; the short version is that the steepest multiple in the index costs $3 a month on a real workload, and the most ordinary one costs $140.

What it costs you in practice

A day of work on this page's workload is 500k input and 167k output tokens — 10M input and 3.34M output over a 20-day month. The multiple decides which half of that bill you should be attacking.

At 4× a workload with this shape spends 57% of its cost on output. At 6× the same workload spends 67%. At 1.8× it spends 38%, and input becomes the half worth attacking instead.

Here is that workload priced at each vendor's median input rate and median multiple, sorted by what it costs:

⚠️ The median input rate and the median multiple in that table are computed separately, so they do not necessarily describe one specific model. What the table measures is the shape of a bill at that vendor, not a quote.

Two things fall out of it.

The first is that output is the majority of the bill almost everywhere, and it is the majority by a wide margin at the vendors with steep multiples. The familiar advice to stop resending your context is sound, but on a workload shaped like this it addresses the smaller half.

The second is what happens when you compare two models on input price alone. Google's median input rate here is about a seventeenth of Anthropic's, which predicts a 17× difference in cost. The actual difference is 21×, because the steeper multiple at the expensive end widens the gap rather than narrowing it. A comparison that ignores the multiple is not slightly wrong It is wrong in a consistent direction, and it will always understate the gap between a high-multiple vendor and a low-multiple one.

Three things this changes

Compare on blended rate, not input rate. The input column is the one everyone quotes and the one that matters least. What you want is the input rate plus the output rate weighted by your own actual split, and if you do not know your split, measure it before comparing anything. The calculator asks for the ratio first for exactly this reason.

Vendor choice is a lever that model choice is not. Picking a cheaper model lowers the bill. Picking a vendor with a shallower multiple lowers a different part of it, and the two are independent. Remember the quartile table: the multiple does not follow the price, so a cheap model from a steep-multiple vendor can cost more than an expensive model from a flat one.

If your output is long, the multiple is the first thing to look at. At 6× versus 1.8×, the same tokens cost you a third more or a third less, and no amount of prompt trimming will recover it. Where output dominates, see why output pricing dominates for the mechanics.

How this was measured, and what it does not cover

Rates come from our own price index, read on 2026-11-14. Before any number on this page is computed, the catalogue is cleaned: test entries, dated snapshots of the same model, wildcard and namespaced identifiers, misattributed vendors, and non-chat models are all removed. That leaves 231 models, and it is the same cleaned set behind every ranking on this site.

That cleaning step matters more than it sounds, and it is the reason the numbers on this page changed from a previous version. Computed over the raw catalogue instead, the quartiles came out identical at 4.00× in all four bands and the mean sat at 4.88×. Neither survived cleaning. The raw catalogue carries six identically-priced gpt-5.6-sol-* variants, dated snapshots of the same model at the same rate, and a gpt-5-test entry at a rate of 100. The duplicates pull the middle toward whichever entries a vendor happens to ship most of. Including them produces a cleaner-looking result and a less true one.

Three limits worth stating. These are gateway rates rather than official list prices: only a small number of models in the index carry a rate verified against the vendor's own published pricing. The quartile result is a median comparison, so it says the central tendency is flat across price bands, not that every model is (the most expensive quartile does lift to 5).00×, and that is a real signal I have not explained. And the vendor table reflects what each vendor currently ships, which is a product decision as much as a pricing one.

FAQ

FAQ

Related

• The thinking tax — the steepest multiple in the index costs $3 a month
• Why output pricing dominates — how the two directions are billed and where the multiple comes from
• Official vs gateway prices, the two price columns on this site, and which models carry a verified official rate
• What actually drives agent cost (the same effect in a coding-agent workload), where context resend dominates
• Cost calculator — put in your own input-to-output split and get a number that is not a guess
• Price index 2026 — the full dataset behind this article

Get API access