The 4× rule in model pricing

Everyone in this industry repeats one number: output tokens cost about four times what input tokens do. It appears in pricing docs, in cost calculators, in every "how to cut your LLM bill" post. It is repeated so often that it stops being a claim and becomes background noise.

It is also, as far as I can tell, true. But the reason it is true turns out to be different from what the repetition implies, and the difference is worth real money.

We track rates for 846 models. Of those, 738 publish both an input and an output rate, and 712 survive a filter for text models. I ran the multiple on all of them. The median is exactly 4.00×.

Then I sorted those 712 models by input price and split them into quartiles, expecting the expensive ones to carry a steeper output surcharge. Generation is the harder half of the job, so a premium model should charge more for it.

Identical, to two decimal places, across a price range spanning three and a half orders of magnitude.

That is the finding this article is about.

What the identical medians rule out

A multiple that tracks generation cost should move with the model. Serving a token out of a large model uses more of the expensive resource than serving one out of a small model, and if the multiple were following that, the top quartile would sit noticeably higher than the bottom one.

It does not move at all. So the 4× is not measuring anything about the models.

The most likely explanation is simpler and less flattering: four is a number pricing teams converge on. Once a large share of the market quotes 4×, a vendor quoting 7× looks expensive on the comparison tables that buyers actually use, and one quoting 2× leaves money on the table for no competitive gain. The multiple is a coordination equilibrium, not a cost signal.

This matters because of what follows from it. If the multiple were a cost signal you would expect it to be roughly uniform and there would be nothing to shop for. Being convention instead, it is set vendor by vendor, and vendors do not agree.

The multiple is decided by the vendor, not the model

Here is the same calculation grouped by vendor, text models only, models with fewer than three entries excluded:

The spread between the top and the bottom of that table is more than fivefold. Meta's median sits just above one, meaning output and input are priced almost the same across its range. OpenAI's sits at six.

These are not small vendors at the margins. The two largest model labs in the table differ by 20% on the multiple before you compare a single rate.

Where the outliers are, and what they tell you

A median hides the shape. The full distribution of the 712 text models:

 <1x    2   #
  1x   52   ####
  2x   45   ###
  3x  125   #########
  4x  212   ################
  5x   92   #######
6-7x   78   ######
  8+  106   ########

Two things stand out. The mode is a hard spike at 4× (212 models, nearly 30% of the set), which is what a coordination equilibrium looks like in data. But there is a second, smaller cluster out past 8× with 106 models in it, and that cluster is not noise.

Nearly all of it is reasoning models. qwen3-30b-a3b-think, qwen3-30b-a3b-thinking-2507, and qwen3-vl-30b-a3b-thinking all sit at 12.00×, the highest in the text set, at $0.10 in and $1.20 out.

The reason is structural rather than commercial. A reasoning model emits its chain of thought as output tokens before it answers. You pay the output rate for the model talking to itself. So the effective output volume on those models is not the length of the answer you receive, it is the length of the answer plus everything the model generated on the way there.

That is a case where the multiple does track real cost, and it is the exception that makes the rest of the table legible: where the 4× is convention, reasoning models are the place where convention gives way to something measurable.

At the other end, 53 models (7.4%) price output at or below input. Two are genuinely inverted: llama-3.2-1b-instruct at 0.25× and llama-3.2-3b-instruct at 0.50×, where generating is cheaper than reading.

What it costs you in practice

A day of work on this page's workload is 500k input and 167k output tokens — 10.0M input and 3.34M output over a 20-day month. The multiple decides which half of your bill you should be optimising. At 4×, a workload with a 3:1 input-to-output ratio spends 57% of its cost on output. At OpenAI's 6×, the same workload spends 67%.

That gap is the whole point. Take a concrete case: 10M input tokens and 3.34M output tokens in a month, priced at each vendor's median input rate and median multiple.

At most vendors output is the majority of the bill, and usually a comfortable one: 57% everywhere the multiple is 4× or steeper, 63% at Anthropic's and xAI's 5×, 67% at OpenAI's 6×. The familiar advice to stop resending your context is sound, but on a workload shaped like this it addresses the smaller half.

Meta is the single exception, and it is worth stopping on. At 1.14×, output falls to 28% of the bill, the only vendor in the set where input is the larger half. Note what that is not: Meta's median input rate here is $1.00, higher than DeepSeek's $0.50 and far above Google's $0.15. It is not a cheap vendor. The reason its bill is shaped differently is entirely the multiple.

The second is what happens when you compare two models on input price alone. Google's median input rate here is one tenth of Anthropic's, which would predict a 10× difference in cost. The actual difference on this workload is 11×, because the steeper multiple widens the gap rather than narrowing it. A comparison that ignores the multiple is not slightly wrong, it is wrong in a consistent direction 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.

On reasoning models, budget for the thinking, not the answer. At 12×, qwen3-30b-a3b-think charges $1.20 per million output tokens against $0.10 input. A one-paragraph answer can carry several thousand tokens of reasoning you never see. If a reasoning model looks cheap on input, that is not the number to judge it on. See what actually drives agent cost for the same effect in a coding-agent context.

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. Meta at 1.14× and DeepSeek at 3.00× are not cheaper because they are worse models, they are cheaper because of a pricing decision.

How this was measured, and what it does not cover

Rates come from our own price index, read on 2026-10-19, covering 846 models. The filter keeps models publishing both an input and an output rate in USD per million tokens (738 of them), then removes 26 non-text models (TTS, image generation, embeddings, rerankers, moderation).

That last filter matters more than it sounds. Left in, those 26 models drag the mean from 4.49× to 4.88× and push the maximum from 12× to 60×. tts-1-hd sits at 60×, and gemini-3-pro-image-preview-l at 60×, but neither is comparable to a text model: the "output" being priced is synthesized audio or generated pixels, and the cost structure is unrelated. Including them would produce a more dramatic number and a less true one.

Three limits worth stating. These are gateway rates, not official list prices — 24 of the models in the index carry a rate verified against the vendor's own published pricing, the rest do not publish a comparable USD list. The quartile result is a median comparison, so it says the central tendency is identical across price bands, not that every model is. And the vendor table reflects each vendor's current catalogue, which is a product decision as much as a pricing one: Meta's 27 models being mostly small open-weight models is what puts its median at 1.14×, and that is a fact about what Meta ships, not about how Meta prices generation.

FAQ

FAQ

Related

• 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
• Cheapest models by workload — why the cheapest model depends on the shape of what you send
• Cost calculator — put in your own input-to-output split and get a number that is not a guess
• Price index 2026 — the full 846-model dataset behind this article

Get API access