Input vs output tokens and why AI output costs more

The pair of input vs output tokens shapes every AI price list. Input, what you send, costs little. Output, what the model writes, costs two to six times more on today's official price lists and sometimes even more depending on the model. The reason isn't commercial. It's physical and holds at every provider, because reading happens in one block while writing happens token by token.

The gap is easy to explain without equations and easy to put in figures, provider by provider. Its practical consequences matter most. Depending on whether your work produces text or analyzes it, you need to watch a different line of the bill and different habits pay off, as the worked example further down shows in black and white.

01Key takeaways

• Input covers everything you send, instructions and documents alike. Output covers everything the model generates.
• Output costs two to six times the input on today's official price lists.
• The cause is technical. Input is read in one pass, while output is generated sequentially, fragment by fragment.
• For creators, shortening the output pays more than shortening the input. For analysis work it's the other way round.
• Four simple habits cut the output bill without touching quality.

02The short answer for readers in a hurry

Output costs more than input because it takes more computing. The model produces its text one token after another, each fragment requiring a full pass of the machine, whereas it takes in your request in a single movement. Price lists reflect that imbalance with two columns and two prices, output always the higher one, with no exception on the market.

Everything that follows details and quantifies that one idea. It fits into the wider logic of AI without a subscription, which is to understand what you pay for so you pay the right price. If the very notion of a token is still fuzzy, the full explainer on tokens is the ten-minute detour to take first.

03Input and output, who is what in a call

Input covers everything that goes to the model. That means your instruction, the brief, the examples, attached documents and, in a conversation, the whole history of earlier exchanges sent again at every turn. Users almost always underestimate it, since the brief often weighs as much as the expected answer and sometimes far more in document analysis tasks.

Output covers everything the model writes. That's the visible answer and, on reasoning models, the internal thinking that gets billed even when it isn't shown. Two separate meters carry two separate prices. The provider's statement splits them line by line, which makes the gap easy to observe on your own usage.

04The technical reason, without equations

The price gap exists because the two operations differ in nature. When reading, the model processes all the tokens of your request in parallel, in one big computing pass. That's the fast step. When writing, it works in a loop. It produces a token, rereads it with everything else, then produces the next one. A thousand output tokens mean a thousand turns of the loop, each one using the whole machine.

You can even see this sequential generation with your own eyes. The text that appears word after word in your tool is the loop running live, each fragment taking a few milliseconds of machine time, with latency included in the show. The price follows the work. The output column of the price lists pays for those turns of the loop, while the input column pays for the single pass. There's no commercial mystery here, only the physics of computing, the same at every provider because model architecture is the same too.

05The gap in figures, provider by provider

Prices per million tokens and output-to-input ratio, official price lists recorded on October 16, 2026

Model	Input	Output	Ratio
xAI grok-4.3	$1.25	$2.50	2×
DeepSeek V4.1-Flash (peak hours)	$0.30	$1.20	4×
Mistral Small 4	$0.15	$0.60	4×
Claude Haiku 4.5	$1	$5	5×
Claude Sonnet 4.6	$3	$15	5×
GPT-5.4 / GPT-5.4-mini	$2.50 / $0.75	$15 / $4.50	6×

The ratio ranges from two to six depending on the company. Remember the order of magnitude rather than the exact figure. Output dominates the bill for creative work everywhere, whatever provider and tier you choose. The full side-by-side of providers, flagship models included, sits in the detailed price list comparison.

06What the gap changes for your kind of use

If you create content, your output far exceeds your input, since a 500-token brief produces a 2,500-token article. The length of the generated texts therefore drives your bill. Asking for three variants just to see triples the main line. Knowing that is often enough to keep it in check. Still, don't give up useful iterations. Even tripled, the output of an article stays at a few cents, so the gap is no reason to cap quality, only a reason to measure it out knowingly.

For document work such as summaries, analysis or news monitoring, the ratio flips, because you send twenty pages to get one back. Input dominates and the useful habits change. Here the lever that pays is to shorten the input without losing anything, with caching as backup. Long-running conversation is the last case, where history sent again swells the cumulative input at every turn, a mechanism detailed in the calculation method.

07The gap on your statement, a worked example

Put the gap to work on the typical article, with 800 input tokens for the brief and 2,500 output tokens for the text. On Claude Sonnet 4.6, the input costs $0.0024 and the output $0.0375. Output makes up 94% of the call's bill even though it weighs only about three times more tokens. That's the price ratio at work, a silent multiplier on every generation you run.

Now reverse the scenario, with twenty pages sent for a one-page summary. That's 30,000 input tokens against 1,700 output tokens. The input costs $0.09 and the output $0.026, so the roles have fully swapped and input now makes up 78% of the total. Two uses give two bill profiles. Your provider's statement tells you at a glance which one you belong to. Most users mix both depending on the day, which makes the exercise all the more instructive.

08Four habits to pay for less output

1. Ask for the right length. An instruction like "answer in 150 words" cuts the output tenfold compared with the free answer of a wordy model, with nothing lost when the question is simple and the answer doesn't need room.
2. Cap generation. The max_tokens setting, present in every API and serious tool, puts a hard ceiling on the maximum output of each call. It's a technical safety net, never a style instruction.
3. Prefer structured formats. A list or a table requested explicitly avoids the padding paragraphs that cost money without informing anyone.
4. Cut the pleasantries. Writing "no introduction and no conclusion" in your templates spares the three ceremonial sentences that models serve by default on every call.

These habits target output because output is expensive. Their cousins on the input side, lighter templates and caching, complete the setup. A review of every lever, priced one by one, is in the full tour of savings.

09Correcting the belief that input is free

A well-understood gap sometimes produces its own excess. Since output dominates, some people stop watching input. That's a mistake in three specific cases. The first is the long conversation, where history sent again makes cumulative input the biggest line without anything flagging it. The second is attached documents, since twenty pages sent just in case weigh 30,000 tokens even if the answer fits in three lines.

The third is bloated templates. A two-page instruction copied into every call ends up costing more than the texts it produces, especially without caching and on small repeated generations. The balancing rule fits in one sentence. Watch the line that dominates YOUR statement, never the one from article examples. The price gap never excuses you from looking at volumes, since the two columns multiply with each other in the formula.

10Two special cases, caching and reasoning

Caching is the good news. Input seen recently, such as templates and fixed instructions, gets billed again at a fraction of the price, up to ten times less on today's price lists. The gap grows even wider as a result, because repeated input becomes almost free over successive calls. The full mechanism is in repeated input at a cut price.

Reasoning is where you need to stay alert. Models that think before answering bill that thinking as output tokens, often invisible in the answer. On a complex task this hidden line can multiply the cost several times. Keep these models for problems that justify them. Everyday writing isn't one of them and classic models handle it with room to spare.

11From theory to your statement

With the gap understood, two moves put it to work. First, price your typical calls with the two columns kept apart, using the step-by-step method in applying the cost formula. Then reread your statement and look at the ratio between your input and your output. It describes your usage profile better than any questionnaire and points by itself to the lever to pull first.

OpenTahr's tools frame output length through templates and brevity instructions. They work with your key at cost price, so the input vs output balance becomes a setting instead of a surprise. The account is free and you can start with your own key in five minutes, with the provider of your choice.

12Frequently asked questions

Why do output tokens cost more than input tokens?

Generation is sequential. Each token produced requires a full pass of the model, whereas input is read in a single parallel pass. The price reflects that extra work. Today's price lists show ratios of two to six depending on the model.

What exactly is the difference between input and output?

Input is everything you send to the model, including instructions, documents and conversation history. Output is everything it generates, including the internal thinking of reasoning models. Two meters carry two prices, kept separate on every price list and every statement.

How many times more does output cost?

Output costs two to six times more on the official price lists recorded for the comparison. The exact ratio is on your provider's price list. The order of magnitude holds everywhere.

How can I reduce the cost of output tokens?

Four habits do it. Ask for a target length in the instruction, cap generation with the dedicated setting, require structured formats and ban ceremonial sentences through your templates. Combined, they commonly cut the output line by two to five times with no loss of quality. The saving shows on the very next statement.

Do cached tokens change the gap?

They widen it in favor of input. Fragments seen recently are billed again at a fraction of the input price, up to ten times less. Your fixed templates become almost free from the second call, while output is paid at full price on every generation, since no cache exists for text that hasn't been written yet.

Transparency note. Prices and ratios were recorded on October 16, 2026 from the official price lists cited in the sources. The latency orders of magnitude are indicative and vary with the infrastructure. None of the providers named here pays us.

13Sources

1. OpenAI, official API price list (input, output, cache)
2. Anthropic, official Claude API price list
3. Mistral, API price list per model (Small 4 at $0.15 and $0.60)
4. DeepSeek, official API price list (peak and off-peak rates)