LLM Pricing Explained: The Real Cost of Running AI

The price of artificial intelligence is collapsing faster than any technology in history. Research institute Epoch AI estimates that the cost of reaching a fixed level of AI performance has been falling about 47% every quarter — roughly 13 times a year — since 2023. In 2021, GPT-3-class quality cost $60 per million tokens; a few years later, equivalent open models were doing the same work for $0.20. And yet AI budgets keep rising. Both things are true at once, and understanding why is the difference between an AI bill you control and one that controls you.

What You Actually Buy: Input Tokens, Output Tokens, and the Invisible Ones

Every LLM provider bills you for two separate products that happen to share a name. Input tokens are everything you send: the system prompt, the tool definitions, the documents, the conversation history. Output tokens are everything the model generates in return — and, crucially, everything it generates invisibly while thinking. All the major providers confirm that internal reasoning or “thinking” tokens are billed as output tokens even though they never appear in the response.

As a rough feel for scale, one million tokens is around 750,000 English words — roughly ten novels. A single chat question might use 2,000 input tokens and produce 300 output tokens. An autonomous agent doing the “same” job can use a hundred times more, because it re-reads its entire growing context at every step. That multiplication, not the headline price, is where real bills come from. If the unit itself is new to you, our explainer on tokens and context windows, and why long conversations get expensive, fills in the mechanics.

The October 2026 Price Map: Flagship vs Economy

The table below is a snapshot of standard per-million-token rates, cross-checked across provider documentation and independent price trackers in late October 2026. Treat it as a map, not a contract — providers revise catalogues constantly, so check the provider’s own page before you budget anything real.

Two things jump out. First, the ratio of output to input is almost exactly 5 to 1 on every current model — that is not a coincidence, as the next section explains. Second, the flagship-to-economy spread is about 100x: GPT-6 Astra at $10/$50 against GPT-6 Luna at $0.10/$0.50. Nobody’s invoice is decided by finding a 10% discount. It is decided by which row of this table the workload actually needs.

Why Output Costs Five Times More Than Input

This ratio looks like price-gouging until you understand the machine. Reading your input happens in one parallel pass — the model looks at all of it at once. Writing the answer happens one token at a time, and every single token requires a full pass through the model’s weights, with the result of each pass feeding the next. Generation is the expensive direction by physics, not by marketing.

That is also why the “hidden” reasoning tokens sting: a model that thinks for 2,000 invisible tokens before writing a 300-token answer has just billed you for 2,300 output tokens. Reasoning models are brilliant at hard problems and brutal on invoices, because their thinking is priced at the output rate. Budget for the thinking, not just the answer.

The Three Levers That Move Your Bill More Than the Price List

1. Prompt caching. In a real application, most of the input never changes: the system prompt, the tool definitions, the reference documents. Providers let you cache that repeated input and bill re-uses at a fraction of the price — around 90% off on most models, 95% on Anthropic’s Opus 5.5 and Sonnet 5.5, 97.5% on Fable 5.1. The catch is that writing to the cache costs extra (for example, Sonnet 5.5 charges $2.50 per million for a five-minute cache write and $4 for a one-hour write), so caching pays off when the same context is reused many times, not once.

2. Batch processing. Work that can wait up to 24 hours — overnight reports, bulk classification, backfills — gets roughly 50% off both input and output across OpenAI, Anthropic, and Google. If your workload has no human staring at a loading spinner, there is no reason to pay the real-time price.

3. The agent multiplier — and this one works against you. A chatbot answer is one model call. An agent that fixes a bug, researches a market, or processes an invoice makes dozens of calls, re-reads its growing context at every step, and calls tools in between. One published walkthrough priced a single five-step support agent at 30,000 conversations a month: $3,690 on Sonnet 5.5, $7,380 on Opus 5.5, $18,450 on Fable 5.1 — and $184.50 on Haiku 5.5. Two details from that analysis are worth memorising: 88% of the bill was input tokens, not the answers, because every step re-sends the system prompt, the tools, and everything said so far; and prompt caching cut the same bill by 61–67%. The model you pick moves the bill 100x. How you engineer the context is what decides whether you pay the flagship price or the economy one. Our guides to context engineering and how agentic AI actually works cover that engineering side in depth.

A Worked Example: What a Support Chatbot Really Costs

Take a small business chatbot handling 10,000 conversations a month, each using about 2,000 input tokens (system prompt, tools, history) and 300 output tokens. Here is the arithmetic, step by step — our calculation, using the October 2026 rates above:

On Claude Sonnet 5.5 ($2 input / $10 output per million):
Input: 10,000 × 2,000 = 20 million tokens → $40.00
Output: 10,000 × 300 = 3 million tokens → $30.00
Monthly total: $70.00 — about $0.007 per conversation.

The identical workload on GPT-6 Astra ($10/$50): input $200, output $150 → $350 a month. Same conversations, same answers in most cases, five times the bill. Now add caching: if 80% of the input is cacheable on Sonnet 5.5 (cache reads at $0.10 per million), input drops to $9.60 and the monthly total to $39.60 — before the small cache-write fees.

This is the whole game in one paragraph. The price list never told you any of this. The token counts did.

The Shopkeeper Rule: Never Price a Project From the Brochure

There is an old buying rule worth keeping next to every pricing page: never take advice from the shopkeeper. The appliance seller who pushes the brand with the best margin for him is not exactly lying — but his recommendation was purchased before you walked in. Get your buying advice from someone with no financial stake in your choice.

AI pricing pages are brochures written by the shopkeeper. Benchmarks quoted by vendors are run on the vendor's favourite test. "Cheapest" claims quietly assume the workload the vendor wants you to imagine. The only honest number in the building is your own: run your real workload on two models, count the tokens, and compare cost per completed task. That measurement takes an afternoon and it beats every comparison article on the internet — including this one. (On a related note, pricing is frequently designed to be misread — AI price lists are no exception.)

Why Your Bill Grows While Prices Collapse

Here is the paradox the whole industry is living inside. The price of a fixed unit of AI capability falls about 13x a year. But the tokens consumed per useful outcome have risen one to two orders of magnitude, because agents and reasoning models do far more work — and far more re-reading — per answer. Cheaper resource, heavier use, bigger total bill. Economists have a name for it: the Jevons paradox. When steam engines got more efficient, coal consumption went up, not down, because efficiency unlocked new uses. AI is running the same script.

The honest way to read the trend: last year's intelligence gets cheap; this year's intelligence stays expensive. Epoch AI's own example makes it concrete — GPT-3 scored 43.9 on the MMLU benchmark while charging $60 per million tokens; later, Meta's open-weight Llama 2-7B scored 45.3 at $0.20 per million. The price of that fixed capability fell roughly 31x a year. But nobody's production system runs last year's capability on last year's architecture. They run this year's agents on this year's frontier models, and that is priced like a frontier.

The stance to take: the winners are not the teams that find the cheapest token. They are the teams that budget cost per task and re-measure it every quarter, because the ground moves that fast.

Who Should Not Pay Per Token

Pay-per-token is the right model when your usage is spiky, experimental, or growing. It is the wrong model in three common cases. First, if your workload is steady and simple — the same classification or extraction job, millions of times — a flat subscription or a committed-use discount usually beats the meter. Second, if your tasks are small and repetitive, routing them to Haiku-class or Luna-class models cuts the bill 10–100x with no visible quality loss; save the flagship for the 5% of tasks that actually need it. Third, if your data can never leave your building, small on-device models turn a variable API bill into a fixed hardware cost — no tokens, no meter, no surprises.

The decision rule is one sentence: match the pricing model to the shape of your usage, not to the hype of the model. For a grounded look at where AI spending actually pays off for a small operation, see our piece on AI in trading and business, and if agents are on your roadmap, start with what AI agents really are before you price them.

Frequently Asked Questions

How much does AI cost per month?

There is no single answer — it ranges from $0 for light chatbot use to five figures for agentic production workloads. The worked example above puts a 10,000-conversation support bot at roughly $40–$70 a month on a mid-tier model. Your number comes from three inputs only: tokens per request, requests per month, and model choice.

How much does AI cost per token?

Per-token is the wrong unit to think in; providers price per million tokens. At the economy end, $0.10 per million input tokens means a single 2,000-token request costs two-hundredths of a cent in input. The bill only becomes real at volume — or when agents multiply the token count per task.

Why is AI so expensive to run?

Two reasons. Training a frontier model costs hundreds of millions of dollars, and that has to be recovered. But the bigger driver of your bill is inference: every output token requires a full pass through the model, reasoning tokens are billed as output, and agents re-read their context at every step. Most of what you pay for is computation you never see.

Is AI getting cheaper?

Per unit of capability, dramatically — about 13x a year by Epoch AI's measurement. Per project, often not, because usage grows faster than unit prices fall. Ask "is my cost per task falling?" instead of "is AI getting cheaper?" — the first question has an answer you can act on.

Which is cheaper, OpenAI or Anthropic?

The question has no answer at brand level. Anthropic's Haiku 5.5 ($0.10/$0.50) and OpenAI's GPT-6 Luna ($0.10/$0.50) are priced identically; their flagships, Fable 5.1 and GPT-6 Astra, are both $10/$50. Compare specific models against your measured workload, not logos against logos.

The Takeaway

LLM pricing looks complicated because the price lists are long. It is actually simple: tokens in, tokens out, output priced about five times input, with caching and batch discounts for the patient and an agent multiplier for the ambitious. Measure your own usage, budget per task, and re-check quarterly — in a market where prices fall 13x a year, last quarter's math is already stale.