What an LLM API Actually Costs Per Month (Real Numbers, Not List Prices)

I search "llm api cost per month" about once a quarter. Every time I get per-million-token prices, which tell me nothing about what I will actually pay at the end of the month.

So I stopped looking at unit prices and worked backwards from my own usage.

Short version: the same workload runs about $20 a month at list price, and about $1 a month on the platform I use now.

Start with the usage, not the price

I run roughly 30 rounds of conversation a day. Refactoring a function, reading a stack trace, asking it to explain code someone else wrote. Each round sends about 8,000 tokens in, and gets about 1,500 tokens back.

The first number is the one people underestimate. It is not just the few lines I type. It is also the model's previous answer, the files I attached, and the whole conversation history. That part is the expensive one, and it gets resent every single round.

Over 22 working days:

The list price bill

Take a flagship coding model at $2.00 per million input tokens and $10.00 per million output tokens.

5.28 × 2.00 = $10.56 0.99 × 10.00 = $9.90 

About $20 a month. Around $0.93 a working day.

That is a light workload. People who run coding agents against a whole repository can push 50,000 to 100,000 tokens into a single round, and the bill goes up tenfold without much effort.

The point of this calculation is not the exact number. It is that knowing your daily token volume is more useful than knowing the unit price.

The same workload, priced differently

On LONG AI, the same model tier bills at $0.0986 per million input tokens and $0.494 per million output tokens.

5.28 × 0.0986 = $0.52 0.99 × 0.494 = $0.49 

A dollar and one cent. From $20 to $1.01, a 95% drop.

The prices are not a promotion. LONG AI is a marketplace rather than a single reseller: sellers list spare capacity they already hold, set their own price, and buyers pick a model rather than a supplier. The platform handles matching, routing and settlement, and takes 0% commission right now. You can check the live price table — it puts the official price, the platform price and the discount on the same line, so you can see exactly how the number is derived instead of taking a marketing page's word for it.

The model plaza shows something that a single-supplier API cannot: multiple sellers quoting on the same model at once. The scheduler picks between them based on price, current load and success rate. If one seller's capacity gets throttled, traffic moves. That is the actual product, and the lower price is a side effect of it.

Why the discount range is so wide

The monthly report publishes the distribution, which is more informative than a single "up to 95% off" claim.

Of the models with at least one quote, 22 sit at the 0.1–0.5 discount band and 19 sit at the 5 discount band. The middle, around 3–4, is empty. Nobody parks there.

Two examples from the deep end of that distribution: GLM 5.3 Flash lists at $0.0012 per million input tokens against an official $0.12, and DeepSeek V4 Flash sits at $0.0030 against an official $0.30. A mid-range option like Qwen3.8 Max runs $0.018 against an official $1.79. The spread between models is wider than the spread between suppliers on the same model, which is worth knowing before you standardise on one. The 5 band is the platform's hard ceiling — a seller may not quote above 50% of the official price — and 0.5 is where most sellers settle. Anything in between is neither cheap enough to win a request nor profitable enough to bother with.

That gap is a useful signal: if you see a model sitting at the ceiling, the supply for it is thin. If it is at 0.1, several sellers are competing.

Browse the full model directory if you want to check availability before committing to anything. Each model page lists the billing unit, the official price, the platform price and the resolution or context tiers.

Not every model bills per token

This is where most cost estimates go wrong. Text bills per million tokens. Images bill per image. Video bills per second. Mixing them into a single per-token estimate gives you a number that is wrong by an order of magnitude.

Image generation also splits: some models bill per image, others per million tokens for the same task. Comparing a per-image quote against a per-token quote is meaningless.

Long-context work has its own trap. Kimi K3 lists at $2.99 per million input tokens officially and $0.030 here; MiniMax M3 is $0.31 officially and $0.0031 here. Both charge for the full context window they process, not just the new text. A 100,000-token conversation costs the same whether you typed one line or a thousand.

In the current snapshot, the cheapest text model is $0.0012 per million input tokens, the cheapest image model is $0.0013 per image, and the cheapest video model is $0.002 per second. If you estimate video at token prices, you will be off by a lot.

What this setup is not good for

Three honest limits.

Cheap and reliable are different things. Many listings carry a "no calls yet" label, meaning a seller listed capacity that nobody has actually exercised. For production traffic, filter for certified sellers. Some popular models only have two or three suppliers, and that is a real constraint rather than a temporary one.

Seller capacity is finite. This is the physical limit of the business: upstream quota can be reclaimed or rate limited. It works well for batch jobs, model comparisons and prompt validation. It is not a replacement for a contracted provider with an SLA — the terms of service say so plainly.

Retries are on you. Failed requests are not billed, but the default SDK retry logic will hammer the same seller three times. Retry across suppliers instead, or you are just repeating the same failure.

Channel health is public but not guaranteed. There is a status page showing per-channel availability and latency. Use it, but do not read it as an SLA. It reflects the last few minutes, not a contract.

Who this actually fits

People who want to test new models without prepaying $200 to see how they perform. Individual developers spending tens to low hundreds a month on API calls. Teams that need five or six model families behind one key and do not want to maintain five sets of credentials.

It does not fit teams that need contractual uptime, invoicing workflows, or a single accountable upstream.

If you want to check whether your own stack is listed, start with the price table and search for the model you already use. The homepage has the current listing count and the deepest discount, both timestamped. If there is no quote for it, do not top up yet. That advice costs me nothing and saves you the surprise.

Integration is two lines. The docs cover the base URL and the key format; the API is OpenAI-compatible, so existing clients work without code changes. Common questions about invoicing, data retention and withdrawals are in the FAQ.

Prices, model counts and discount figures reflect the live pages on 6 November 2026. The price table refreshes every 30 minutes and seller quotes change without notice.