How LLM Token Pricing Works
Short answer: every rate card here is quoted per 1,000,000 tokens and metered on two counters — what you send and what you get back. One request with 3,000 input tokens and 600 output tokens costs $0.00132 to $0.0300 depending on the model, a 22.7x spread for byte-identical traffic. Published input rates span $0.20 to $5.00 per 1M and output rates $1.20 to $25.00 per 1M. Below: the unit, the formula, the three rate classes, and the five reasons your invoice never matches your estimate.
Four things every rate card assumes
Once these four are clear, every price on this site — and every line on your invoice — becomes arithmetic.
The same rate at three scales
Per-million rates are hard to feel at request size. Here is the identical price expressed per 1,000 tokens — the scale most real prompts actually sit at.
| Model | Input per 1,000 | Input per 1M | Output per 1,000 | Output per 1M |
|---|---|---|---|---|
GPT-5.6 Sol OpenAI · flagship (promo) | $0.00400 | $4.00 | $0.02000 | $20.00 |
GPT-5.6 Terra OpenAI · mid | $0.00200 | $2.00 | $0.01200 | $12.00 |
GPT-5.6 Luna OpenAI · budget | $0.00020 | $0.20 | $0.00120 | $1.20 |
Claude Opus 5 Anthropic · flagship | $0.00500 | $5.00 | $0.02500 | $25.00 |
Claude Sonnet 5 Anthropic · mid | $0.00200 | $2.00 | $0.01000 | $10.00 |
Claude Haiku 4.5 Anthropic · budget | $0.00100 | $1.00 | $0.00500 | $5.00 |
Gemini 3.6 Flash Google · mid (promo rate) | $0.00075 | $0.75 | $0.00375 | $3.75 |
Gemini 3.5 Flash-Lite Google · budget | $0.00030 | $0.30 | $0.00250 | $2.50 |
Per-1,000 figures are the published per-1M rates divided by 1,000, shown to five decimals so the cheap end stays visible: the spread from the cheapest line item to the dearest is 125x ($0.00020 input tokens on GPT-5.6 Luna against $0.02500 output tokens on Claude Opus 5). Rates verified 2026-08-23 (OpenAI) and 2026-08-09 to 2026-08-16 (Anthropic and Google). Long-context, caching, batch and taxes are excluded — this is the base rate card only.
Two multiplications, one addition
That is the whole model. Nothing else enters the calculation.
cost = (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate)
Take a realistic support request: a 3,000-token prompt — system instructions, a retrieved article and the customer's message — and a 600-token answer. On GPT-5.6 Sol that is 0.003 × $4.00 = $0.0120 for the prompt plus 0.0006 × $20.00 = $0.0120 for the answer: $0.0240 per request. The prompt is 83.3% of the tokens and exactly half the money.
Run the same shape on GPT-5.6 Luna and the arithmetic gives $0.00132. On Claude Opus 5 it gives $0.0300. Identical text, identical token counts, 22.7x apart — which is why model choice, not prompt polish, is usually the first lever. Price your own shape in the API Cost Calculator.
At 1,000 requests a day — 30,000 a month — that single request shape becomes $39.60 a month on Luna and $900.00 on Opus 5. Linear in tokens, linear in volume, no step change anywhere: there is no volume discount in the rate cards tracked here, and no minimum commit.
One request shape, eight bills
3,000 input tokens, 600 output tokens, 1,000 requests a day. The output share column is the part that surprises people: output is 16.7% of the tokens and half or more of the money.
| Model | Input cost | Output cost | Per request | Output share of spend | Monthly (30,000) |
|---|---|---|---|---|---|
GPT-5.6 Sol OpenAI · flagship (promo) | $0.0120 | $0.0120 | $0.0240 | 50.0% | $720.00 |
GPT-5.6 Terra OpenAI · mid | $0.00600 | $0.00720 | $0.0132 | 54.5% | $396.00 |
GPT-5.6 LunaLowest OpenAI · budget | $0.00060 | $0.00072 | $0.00132 | 54.5% | $39.60 |
Claude Opus 5 Anthropic · flagship | $0.0150 | $0.0150 | $0.0300 | 50.0% | $900.00 |
Claude Sonnet 5 Anthropic · mid | $0.00600 | $0.00600 | $0.0120 | 50.0% | $360.00 |
Claude Haiku 4.5 Anthropic · budget | $0.00300 | $0.00300 | $0.00600 | 50.0% | $180.00 |
Gemini 3.6 Flash Google · mid (promo rate) | $0.00225 | $0.00225 | $0.00450 | 50.0% | $135.00 |
Gemini 3.5 Flash-Lite Google · budget | $0.00090 | $0.00150 | $0.00240 | 62.5% | $72.00 |
Output is 16.7% of tokens on this shape and 50.0% – 62.5% of spend — the 8.33x premium on Gemini 3.5 Flash-Lite pushes it furthest. Monthly figures assume 30 days at 1,000 requests a day. Excludes caching, batch, tool use and taxes; OpenAI long-context rates do not apply at 3,000 input tokens.
Standard, promotional, long-context
Three kinds of price appear on the rate cards tracked here. Only the first is permanent.
Standard rates are the default: six of the eight tracked models sit here, from $0.20/1M input on GPT-5.6 Luna to $5.00/1M on Claude Opus 5.
Promotional rates have an expiry. GPT-5.6 Sol is $4.00/$20.00 through at least Nov 21, 2026 — it was $5.00/$30.00 before Aug 21, 2026. Gemini 3.6 Flash is $0.75/$3.75 through Dec 31, 2026 and returns to $1.50/$7.50 on Jan 1, 2027. When a promo lapses, nothing about your traffic changes and the bill still moves: the 3,000/600 request priced above goes from $0.0240 to $0.0330 on Sol and from $0.00450 to $0.00900 on Gemini 3.6 Flash. Every change we confirm is logged in the pricing changelog.
Long-context rates are a threshold, not a tier you choose. OpenAI applies a second rate above 272,000 input tokens in a single request: input doubles and output rises 1.5x. Claude and Gemini charge one flat rate across the whole window. The practical consequence is that a 1,050,000-token request on Sol costs $8.40 in input while the same tokens in 100K chunks cost $4.20 — see What Is a Context Window? for the full comparison.
Two rate classes are deliberately absent. Cached and batch requests are cheaper on most platforms, but we cannot verify those rates from first-party sources, so we publish none. Every figure on this site is standard on-demand pricing — which makes our numbers conservative, not inflated.
What happens when the promo ends
Both promotional rates have published end dates. This is the same 3,000-in / 600-out request before and after.
| Model | Rate today | Rate after the promo | Change | This request today | This request after | Monthly after |
|---|---|---|---|---|---|---|
GPT-5.6 Sol OpenAI · through at least Nov 21, 2026 | $4.00 / $20.00 | $5.00 / $30.00 | input +25.0%, output +50.0% | $0.0240 | $0.0330 | $990.00 |
Gemini 3.6 Flash Google · through Dec 31, 2026 | $0.75 / $3.75 | $1.50 / $7.50 | both +100.0% | $0.00450 | $0.00900 | $270.00 |
Post-promotional rates are the rates those cards carried before the current promotion (Sol, before Aug 21, 2026) or the rates the provider has published for the following period (Gemini 3.6 Flash, from Jan 1, 2027). Monthly figures assume 30,000 requests. Providers extend promotions as often as they end them — confirm against the changelog before you budget on either column.
Price any token count at both rates
Set a token count and see what it costs as input and as output on all eight models — including where OpenAI's long-context rate switches on.
| Model | Input rate / 1M | Output rate / 1M | These tokens as input | These tokens as output | Rate class |
|---|
Capacity, ceilings and idle months
Four things people assume are charged. None of them are.
An unused context window. A 1,050,000-token window on GPT-5.6 Sol is free until you fill it. Sending 3,000 tokens into it costs $0.0120; filling it completely costs $8.40 at long-context rates. The window sizes the request you are allowed to make, not the invoice.
A max_tokens ceiling. Setting a 4,000-token cap does not cost 4,000 tokens — you are billed for tokens actually generated. The cap is insurance, not a charge: a single maxed-out reply costs up to $3.20 on Claude Opus 5, and the ceiling is what stops an unbounded loop from finding that number repeatedly.
Streaming. Delivering tokens as they are produced changes latency and nothing else. The same token count is metered either way.
Months you do not call. There is no minimum commit and no seat fee in the rate cards tracked here. Zero requests is a zero bill — which is the structural difference from a flat subscription, priced in the API vs Subscription Break-even Calculator.
Why your invoice is never your estimate
The formula is exact. The inputs are not. Five things move the number, and only one of them is under your control.
1. Tokenizers differ. The same text meters several percent differently across providers, because each model uses its own vocabulary. Our counters use a documented heuristic — see how we verify — which is accurate enough for planning and never exact to the token.
2. Output length is unknowable in advance. You can predict your prompt; you cannot predict the reply. This is the single largest source of variance and the reason a cap matters. Budget at the ceiling you set, not at the average you hope for.
3. Some input is added for you. Platform-injected instructions, tool and function schemas, and safety preambles are metered as your input even though you did not type them. Measure what your prompt actually contains with the Prompt Weight Analyzer.
4. Chat re-sends history. Every turn pays for the entire transcript again. On a 20-turn support conversation, 96.6% of input tokens are re-sent history — the compounding is modelled in the Conversation Cost Growth Calculator.
5. Rates change mid-cycle. Two of the eight tracked models are on promotional pricing with published end dates. A rate change inside a billing month shows up as a blended number you will not be able to reproduce from a single rate card.
Three numbers you actually need
Every cost forecast on this site reduces to the same three inputs. Get them right and the arithmetic takes care of itself.
Token pricing questions
How is LLM API pricing calculated?
Cost equals input tokens divided by one million times the input rate, plus output tokens divided by one million times the output rate. On a request with 3,000 input tokens and 600 output tokens that is $0.0240 on GPT-5.6 Sol, $0.0300 on Claude Opus 5 and $0.00132 on GPT-5.6 Luna. No other term appears in the published rate cards tracked here.
How much does one AI token cost?
Per 1,000 tokens, input ranges from $0.00020 on GPT-5.6 Luna to $0.00500 on Claude Opus 5, and output from $0.00120 on GPT-5.6 Luna to $0.02500 on Claude Opus 5 — a 125x spread between the cheapest and dearest line item. Per million tokens the same rates read $0.20 to $5.00 for input and $1.20 to $25.00 for output.
Is LLM pricing per token, per request or per month?
Per token, quoted per 1,000,000 tokens, counted separately for input and output. The published rate cards tracked here carry no per-request fee, no seat fee and no platform fee — tokens are the only metered unit. What you pay in a month is that unit rate multiplied by how many tokens you actually sent and received.
Do I pay for the whole context window?
No. A context window is a capacity, not a charge. Sending a 3,000-token prompt into GPT-5.6 Sol's 1,050,000-token window costs $0.0120, not the $8.40 it would cost to fill the window completely at long-context rates. You are billed for tokens actually sent and generated, which is why an oversized window is free until you use it.
Why is my bill higher than my estimate?
Five things move the number: tokenizers differ by provider so the same text meters several percent differently; output length cannot be known in advance; platform-injected instructions and tool schemas are metered as your input; taxes and currency conversion sit outside the rate card; and rates themselves change on promotional timelines. Cap your output and measure real request shapes to close the gap.
Do LLM API prices change?
Yes. Two of the eight models tracked here are on promotional rates with published end dates: GPT-5.6 Sol at $4.00/$20.00 through at least Nov 21, 2026, previously $5.00/$30.00, and Gemini 3.6 Flash at $0.75/$3.75 through Dec 31, 2026, moving to $1.50/$7.50 on Jan 1, 2027. When those lapse, the 3,000-in/600-out request priced here rises 37.5% on Sol and doubles on Gemini 3.6 Flash.
Where these numbers come from.
Rates are published per-1M input and output prices, re-verified weekly against official provider documentation: OpenAI 2026-08-23, Anthropic and Google 2026-08-09 to 2026-08-16. Per-1,000 figures are those rates divided by 1,000; per-request and monthly figures are the formula above applied to an illustrative 3,000-in / 600-out shape at 1,000 requests a day. OpenAI long-context rules are applied automatically above 272,000 input tokens, where input doubles and output rises 1.5x. Caching, batch, tool use and taxes are excluded throughout — always confirm against your own invoice before making purchasing decisions.
Official sources: OpenAI model docs, Anthropic pricing, Google Gemini API pricing.