DEV Community

Andrey Altrouter
Andrey Altrouter

Posted on

The cheapest LLM for reading is 72% more expensive for writing

You want the cheapest model for your side project. You open a price list, sort by the input price, and take the top row. For a lot of projects that's the right answer. For a project that mostly writes text, it can be the wrong one by 70%.

The reason is that every model has two prices, and which one matters depends on what your calls look like.

A few words first

A token is a chunk of text, about three quarters of an English word; APIs bill per million of them. Input tokens are everything you send: instructions, documents, chat history. Output tokens are what the model writes back. A list price is the vendor's official price per million tokens, with no discounts. RAG (retrieval-augmented generation) means pasting a few relevant documents into the prompt so the model can answer from them.

Two prices, one bill

The cost of one call is input tokens × input price + output tokens × output price. If model A is cheaper on both prices, A wins every workload, and that's what you see for most pairs. But sometimes A is cheaper on input and B is cheaper on output. Then the winner depends on the shape of your call: how many output tokens you get back for each input token.

I took the official list prices of the 55 chat models in the catalog of the LLM gateway I work on (snapshot of November 19, 2026) and compared every pair. 45 of the 1,485 pairs flip like this. That's only 3%, but the flips cluster in the places beginners shop: among the 11 models with input at $0.30 per million or less, 3 of the 55 pairs flip, and one of them is the top of the list.

The top of the list changes hands

Sorted by input price, the cheapest model is claude-haiku-5.5 at $0.10 input and $0.50 output. Second is mimo-v2.6-flash at $0.14 input and $0.28 output: 40% more expensive to read, but 44% cheaper to write. Here is what that does to four typical call shapes, at list prices per million calls:

Workload (input / output tokens) claude-haiku-5.5 mimo-v2.6-flash The other one costs
Classify a ticket (600 / 5) $62.50 $85.40 +37%
Answer from documents, RAG (3,000 / 250) $425 $490 +15%
One chat turn (1,500 / 300) $300 $294 +2%
Write a blog post (200 / 1,500) $770 $448 +72%

For classification and RAG, the model at the top of the list really is the cheapest. Most of the call is reading, and haiku-5.5 reads cheaper. For writing it's the other way round. A post generator on haiku-5.5 pays 72% more than it would on the second row of the same list. A chat turn is close to the line, so there the two cost almost the same.

I also checked the whole catalog: at every shape I tried, the cheapest model of all 55 was one of these two.

Where the line is

You don't need a table for every pair. Two models cost the same when the extra you pay for input on one is equal to the extra you pay for output on the other:

def crossover(a_in, a_out, b_in, b_out):
    """Prices in $ per 1M tokens. A is cheaper on input, B on output.
    Returns input tokens per output token at which both cost the same."""
    return (a_out - b_out) / (b_in - a_in)

print(crossover(0.10, 0.50, 0.14, 0.28))  # 5.5   haiku-5.5 vs mimo-v2.6-flash
print(crossover(2.00, 12.0, 2.50, 10.0))  # 4.0   gpt-5.6 vs gpt-4o
Enter fullscreen mode Exit fullscreen mode

If your calls send more than 5.5 input tokens for each output token, haiku-5.5 is cheaper. If fewer, mimo-v2.6-flash is. The same thing happens higher up the price list. gpt-5.6 has a cheaper input price than gpt-4o ($2.00 against $2.50) but a pricier output ($12 against $10). It wins RAG by 11% and loses blog posts by 19%, and the line is at 4 input tokens per output token. gpt-5 against grok-4.7 flips at 5.3, and the loser on writing pays 62% more.

To find your own ratio, don't guess. Every API response includes a usage field with the input and output token counts. Log it for a day of real traffic, add up both columns, and divide. A chatbot with a long system prompt and history can be at 10:1. A tool that turns bullet points into emails can be at 1:5. That one number tells you which column of the price list to sort by.

What this doesn't solve

This is a price comparison, not a model comparison. Two models that cost the same per call don't give the same answers, and a cheaper model that needs a second try costs double. Check quality on your own examples first, then use the crossover to choose between the models that pass.

Output length also isn't fixed per model. If one model writes 30% longer answers for the same request, its effective output price is 30% higher, and the crossover moves. Measure the output tokens of each candidate, don't assume they match.

Finally, these are list prices from one snapshot. Prompt caching, batch discounts and price changes shift both columns, sometimes unevenly, so recompute the crossover when any of them applies to you. And for 97% of pairs none of this matters: one model is simply cheaper on both prices, and sorting by either column gives the same answer.

Top comments (1)

Collapse
 
timmyzinin profile image
Tim Zinin •

The crossover calculation becomes more useful when it is paired with a quality threshold, since a retry changes the effective price. Have you considered comparing cost per accepted answer for the same task set?