Token Calculator
Calculate the cost of using AI language models. Estimate tokens from text and compare pricing across models like GPT-5, Claude, and Gemini.
Denne beregner er endnu ikke oversat til Dansk.
Vi viser foreløbig versionen på Engelsk, mens vi færdiggør oversættelsen.
$
$
$0.02
Input: $0.0050 · Output: $0.01
$1.75
100 requests per day
$52.50
30 days
$638.75
365 days
Estimate tokens from text
Characters
0
Words
0
Estimated tokens
~0
Token cost calculator. LLM API costs for GPT-5.6 Sol, Terra and Luna, Claude Opus 5, Sonnet 5, Gemini and DeepSeek.
This token calculator prices any LLM API call — GPT-5.6 Sol, Terra or Luna, Claude Opus 5 or Sonnet 5, Gemini, Grok or DeepSeek — from your token counts. It applies each model's published per-million rate to your input and output tokens, then shows the cost per request, per day, per month and per year.
What a token is, and why AI APIs bill by the token
A token is the sub-word chunk a language model actually reads and writes — not a whole word. In English, one token averages about 4 characters, or roughly 0.75 words, so 1,000 words works out to about 1,333 tokens. Every model in the GPT-5.6, Claude, Gemini, Grok and DeepSeek families breaks your text into these chunks before it does anything else, and the meter starts running on the very first one.
AI APIs bill per token, and they split the bill in two: input tokens (your system prompt, context, chat history and question) and output tokens (the model's reply). Providers quote both as a price per million tokens, and output is almost always the pricier side. The multiple is remarkably consistent inside each family: all three GPT-5.6 tiers charge exactly six times more for output than input, the whole Claude line charges exactly five times more, and only models like DeepSeek V4 Pro keep it to roughly double. A single call might cost a fraction of a cent, but multiply that by tens of thousands of requests a day and token spend becomes a real line item.
The exact token count for the same text changes from one model to the next, because each provider ships its own tokenizer. OpenAI's newer GPT models use the o200k_base encoding; Anthropic, Google and DeepSeek each use their own. What surprises most teams is that the count also changes between generations of the same family. Anthropic's pricing documentation states that Claude 4.7 and later models — which includes Claude Opus 5, Opus 4.8 and Sonnet 5 — use a newer tokenizer that produces approximately 30% more tokens for the same text, while Sonnet 4.6 and earlier use the previous one. The rate card did not move, so the same prompt on Sonnet 5 bills roughly 30% higher than on Sonnet 4.6 even though both are listed at $3 input and $15 output. Code, JSON, numbers, emoji and non-English text fragment hardest, which is why token-based pricing never maps cleanly onto a word count.
July 2026 reshaped the top of the market twice. On 9 July OpenAI made GPT-5.6 generally available as three separate models rather than one model with an adjustable reasoning effort — Sol as the flagship, Terra as the balanced everyday tier, and Luna as the fastest and cheapest. On 24 July Anthropic shipped Claude Opus 5 next to Claude Fable 5 and Claude Sonnet 5. The striking part is the price sheet: Sol costs the same as GPT-5.5 at $5 and $30, Terra the same as GPT-5.4 at $2.50 and $15, Opus 5 the same as Opus 4.8 at $5 and $25, and Sonnet 5 the same as Sonnet 4.6 at $3 and $15. A whole generation of capability arrived at flat prices, so the interesting question stopped being "how much did it go up" and became "which tier do I actually need". The one genuinely new price point is GPT-5.6 Luna at $1 and $6, which lands right beside Claude Haiku 4.5 at $1 and $5 — identical on input, a dollar apart on output.
This calculator ships with the published standard rates for nine providers: OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Amazon, Meta and Cohere. The spread is close to three orders of magnitude. Amazon Nova Micro sits at one end at $0.035 input and $0.14 output per million; GPT-5.5 Pro sits at the other at $30 and $180, an 857-fold gap on the input side alone. In between you get Claude Fable 5 at $10 and $50, GPT-5.6 Sol at $5 and $30, Claude Opus 5 at $5 and $25, Claude Sonnet 5 at $3 and $15, GPT-5.6 Terra at $2.50 and $15, Grok 4.5 at $2 and $6, Gemini 3.6 Flash at $1.50 and $7.50, GPT-5.6 Luna at $1 and $6, and DeepSeek V4 Pro at $0.435 and $0.87. Every figure is the standard on-demand rate, with no batch, cached-input or long-context tier applied, and Claude Sonnet 5 is listed at its standard rate rather than its launch promotion so a budget built here does not break when the promotion lapses. If your model isn't in the list, pick "Custom" and enter its per-thousand or per-million rate — the per-request, monthly and yearly totals compute the same way. All figures are in US dollars, the currency these providers bill in.
How to use the token cost calculator
Going from a model name to a monthly bill takes only a few steps.
1. Pick a model. The dropdown groups presets by provider — choose GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, Claude Opus 5, Claude Sonnet 5, DeepSeek V4 Pro or any other preset and its official rate fills in automatically. Running something that isn't listed, like an internal or brand-new model? Choose "Custom" and type the price yourself.
2. Confirm the price and unit. The 1K / 1M switch next to the price toggles between "per thousand tokens" and "per million tokens." Most providers publish per-million rates, but if you only have the per-thousand figure, flip the switch instead of dividing by hand.
3. Enter tokens and frequency. Fill in input tokens, output tokens and requests per day. Not sure how many tokens you're using? Run your text through the estimator further down the page first.
4. Read the four result cards. You'll see cost per request, per day, per month and per year, with the input and output costs broken out separately — so you can tell whether the problem is an overlong prompt or a rambling reply.
5. Estimate tokens from text. Paste any text and the tool shows characters, words and estimated tokens using the general heuristic max(⌈chars / 4⌉, ⌈words × 1.33⌉). It's enough for budgeting; for an exact count, use tiktoken or the usage field the API returns.
The cost of a single call follows one formula: cost per request equals input tokens ÷ 1,000,000 × the input price, plus output tokens ÷ 1,000,000 × the output price. In words: input tokens times the input rate, plus output tokens times the output rate, all scaled from per-million to per-token. Multiply that by requests per day for the daily figure, then by 30 for a monthly budget or by 365 for the year — the calculator uses a flat 30-day month so two projections built a month apart stay comparable.
The most useful habit is to run the same token counts through three tiers before committing. Price your workload on the flagship, on the balanced tier and on the cheap tier of the same family, then decide whether the quality gap is worth the difference in the yearly card. Everything runs in your browser — the text you paste is never uploaded to a server and never touches a model's quota.
What 1M tokens costs on each model (input vs output)
| Model | Input (USD / 1M tokens) | Output (USD / 1M tokens) | Output ÷ input |
|---|---|---|---|
| Amazon Nova Micro | $0.035 | $0.14 | 4× |
| DeepSeek V4 Flash | $0.14 | $0.28 | 2× |
| DeepSeek V4 Pro | $0.435 | $0.87 | 2× |
| GPT-5.4 Mini | $0.75 | $4.50 | 6× |
| Claude Haiku 4.5 | $1 | $5 | 5× |
| GPT-5.6 Luna | $1 | $6 | 6× |
| Gemini 3.6 Flash | $1.50 | $7.50 | 5× |
| Gemini 3.5 Flash | $1.50 | $9 | 6× |
| Grok 4.5 | $2 | $6 | 3× |
| GPT-5.6 Terra / GPT-5.4 | $2.50 | $15 | 6× |
| Claude Sonnet 5 / 4.6 | $3 | $15 | 5× |
| Claude Opus 5 / 4.8 | $5 | $25 | 5× |
| GPT-5.6 Sol / GPT-5.5 | $5 | $30 | 6× |
| Claude Fable 5 | $10 | $50 | 5× |
| GPT-5.5 Pro | $30 | $180 | 6× |
Cost examples with real numbers from the calculator engine
Support chatbot on GPT-5.6: 100 conversations a day
An e-commerce support bot handles 100 conversations a day. Each one runs about 2,000 input tokens (system prompt + chat history + the customer's question) and about 500 output tokens for the reply.
On GPT-5.6 Sol at $5 input and $30 output per million: $0.025 per conversation, $2.50 a day, $75.00 a month and $912.50 a year.
On GPT-5.6 Terra at $2.50 and $15: $0.0125 per conversation, $1.25 a day, $37.50 a month and $456.25 a year — exactly half of Sol on every line, because Terra is priced at exactly half on both sides.
On GPT-5.6 Luna at $1 and $6: $0.005 per conversation, $0.50 a day, $15.00 a month and $182.50 a year.
On Claude Opus 5 at $5 and $25: $0.0225 per conversation, $2.25 a day, $67.50 a month and $821.25 a year.
On Claude Sonnet 5 at $3 and $15: $0.0135 per conversation, $1.35 a day, $40.50 a month and $492.75 a year.
On Claude Haiku 4.5 at $1 and $5: $0.0045 per conversation, $0.45 a day, $13.50 a month and $164.25 a year.
On DeepSeek V4 Pro at $0.435 and $0.87: $0.001305 per conversation, $0.1305 a day, $3.92 a month and $47.63 a year.
Same traffic, and the monthly spread runs from $75.00 at the top to $3.92 at the bottom — roughly nineteen times. You don't even have to change provider to move the needle: staying inside the GPT-5.6 family and dropping from Sol to Luna cuts the bill five-fold, $75.00 down to $15.00 a month, or $730.00 saved over a year. Notice too that Luna and Claude Haiku 4.5 share the same $1 input rate, and the $1 gap in their output rate is worth exactly $1.50 a month at this volume. Model choice is almost always the biggest lever on the bill.
GPT-5.6 Sol vs Claude Opus 5: coding agent at 200 calls a day
A coding agent is the workload where the two new flagships actually compete. Say each call carries 20,000 input tokens (the repo context, the diff and the instructions) and returns 3,000 output tokens, and the team fires 200 calls a day.
On GPT-5.6 Sol at $5 and $30: $0.10 input plus $0.09 output is $0.19 per call, $38.00 a day, $1,140.00 a month and $13,870.00 a year.
On Claude Opus 5 at $5 and $25: $0.10 plus $0.075 is $0.175 per call, $35.00 a day, $1,050.00 a month and $12,775.00 a year.
The two flagships charge the identical $5 input rate, so the entire gap comes from output: $90.00 a month, or $1,095.00 a year — about 8.6% in Opus 5's favour. That is a real number but a small one, and it flips the moment your workload becomes output-heavy or input-heavy, which is exactly why you should price your own token mix rather than trust a headline.
The tier below is where the money is. On GPT-5.6 Terra at $2.50 and $15: $0.095 per call, $19.00 a day, $570.00 a month and $6,935.00 a year — half of Sol. On Claude Sonnet 5 at $3 and $15: $0.105 per call, $21.00 a day, $630.00 a month and $7,665.00 a year. On GPT-5.6 Luna at $1 and $6: $0.038 per call, $7.60 a day, $228.00 a month and $2,774.00 a year.
And the ceiling: Claude Fable 5 at $10 and $50 costs $0.35 per call, $70.00 a day, $2,100.00 a month and $25,550.00 a year — exactly double Opus 5, which is what Anthropic means when it positions Opus 5 as near-frontier capability at half the price of its top model. Between Luna and Fable 5 on the same 200 calls a day there is $22,776.00 a year.
Long document or RAG: reading 100,000 tokens at once
A long-document or RAG job feeds one big report — about 100,000 input tokens, roughly a short book — into the model and asks for a 2,000-token summary, all in a single request.
On GPT-5.6 Sol: $0.50 for the input, $0.06 for the output, $0.56 in total.
On Claude Opus 5 at $5 and $25: $0.50 plus $0.05, or $0.55 per request — one cent apart from Sol, because at this token mix the shared $5 input rate accounts for about 90% of the bill and the output rate barely gets a vote.
On Claude Sonnet 5 at $3 and $15: $0.30 plus $0.03, or $0.33.
On GPT-5.6 Terra at $2.50 and $15: $0.25 plus $0.03, or $0.28.
On Gemini 3.6 Flash at $1.50 and $7.50: $0.15 plus $0.015, or $0.165.
On GPT-5.6 Luna at $1 and $6: $0.10 plus $0.012, or $0.112.
On DeepSeek V4 Pro: $0.0435 plus $0.00174, or $0.0452 per request — about a twelfth of the Sol figure, $0.56 against $0.0452. The lesson is the mirror image of the chatbot case: with a long context the input side swallows almost the whole cost, so the input rate, not the output rate, is the deciding factor for documents and RAG. It also explains why Sol and Opus 5 look interchangeable here and diverge on chat traffic.
The same-price trap: Claude Sonnet 5 against Sonnet 4.6
Here is the case a rate card will never show you. Claude Sonnet 5 and Claude Sonnet 4.6 are both listed at $3 input and $15 output per million, so on paper an upgrade is free. But Anthropic documents that Claude 4.7 and later models use a newer tokenizer producing approximately 30% more tokens for the same text, and Sonnet 4.6 is on the older one.
Take the support bot from the first example: 2,000 input and 500 output tokens per conversation, 100 conversations a day. On Sonnet 4.6 that is $0.0135 per conversation, $40.50 a month and $492.75 a year.
Feed the identical text to Sonnet 5 and the same prompt tokenizes to roughly 2,600 input and 650 output tokens. At the identical $3 and $15 rates: $0.01755 per conversation, $1.755 a day, $52.65 a month and $640.58 a year.
Same text, same rate card, $12.15 more a month and $147.83 more a year — a 30% increase that never appears on any pricing page. On prose the inflation is at the lower end of that range; on code, JSON and structured payloads it runs higher.
How to use the calculator here: don't reuse the token counts you measured on an older model. Paste a real sample through the estimator or, better, read the usage field the API returns on a handful of live calls against the new model, then enter those counts. A migration that looks cost-neutral on the rate card can still land 30% over budget.
Estimating tokens from text: how many tokens is this?
The text estimator converts with max(⌈chars / 4⌉, ⌈words × 1.33⌉).
A short English sentence — "The quick brown fox jumps over the lazy dog." — is 44 characters across 9 words: by characters ⌈44 / 4⌉ = 11, by words ⌈9 × 1.33⌉ = 12, and the tool keeps the larger figure, 12 tokens.
A 500-word document of 3,999 characters: by characters ⌈3,999 / 4⌉ = 1,000, by words ⌈500 × 1.33⌉ = 665, so it lands on 1,000 tokens — the character side wins.
With ordinary spaced text, the characters-over-4 side almost always dominates. Three caveats: code, JSON and long unbroken strings fragment into more tokens; non-English text is usually denser, with Spanish or German running 15-25% more tokens than the same English; and the newest Claude models tokenize the same text about 30% more heavily than their predecessors. The heuristic is a generic BPE approximation, not a per-model count. Good enough for a budget; for an invoice-level count, use tiktoken or the API's usage field.
Common mistakes when estimating token cost
- Mixing up the input price with the output price. Output is almost always dearer, and the multiple is a family trait: every GPT-5.6 tier charges six times more for output than input — Sol at $5 against $30, Terra at $2.50 against $15, Luna at $1 against $6 — the whole Claude line charges exactly five times more, and only models like DeepSeek V4 Pro keep it to roughly double, $0.435 against $0.87. When you budget, look at the output side first: capping reply length with max_tokens usually saves more than trimming the prompt.
- Assuming a same-priced successor costs the same. Claude Sonnet 5 and Sonnet 4.6 both list at $3 and $15, and Claude Opus 5 and Opus 4.8 both list at $5 and $25, but Anthropic documents that Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text. Identical rates on more billable tokens is a higher bill. Re-measure your real token counts against the new model before you sign off a migration as cost-neutral.
- Treating a token as a word and ignoring the tokenizer. English averages about 4 characters per token, but code and non-English text run denser — Spanish or German can use 15-25% more tokens for the same meaning. This page's estimator uses the general four-characters-per-token heuristic; for a model's exact count, reach for tiktoken (GPT) or the usage field the API returns.
- Reading the per-request cost and forgetting to multiply by scale. A single call at $0.0135 looks like nothing, but at 100 requests a day that's $40.50 a month and $492.75 a year on Claude Sonnet 5 (see the first example). What sets the budget is price times frequency times days — always fill in the requests-per-day field.
- Forgetting that every conversation turn resends the entire history. The API is stateless: on turn 10 you resend the previous nine, so input tokens grow almost quadratically and a moderately long chat becomes tens of thousands of tokens per call. Use a sliding window of the most recent turns, or summarize the history every so often.
- Defaulting to the flagship for the whole workload. Since GPT-5.6 shipped as three separate models, the cheapest and priciest tier of one family are five times apart on both sides — Luna at $1 and $6 against Sol at $5 and $30 — and Anthropic's own line runs from Haiku 4.5 at $1 and $5 up to Fable 5 at $10 and $50. Classification, extraction, routing and rewriting are fine on Luna, Haiku 4.5, GPT-5.4 Mini or DeepSeek V4 Pro; save Sol, Opus 5 or Fable 5 for genuine reasoning, hard code or long-form writing. A "small model first, escalate only when needed" router often cuts the bill in half.
- Overlooking prompt caching and the batch discount. Both are documented, first-party discounts, not tricks. A cache hit bills at 10% of the standard input rate on OpenAI and Anthropic alike — cached input on GPT-5.6 Sol is $0.50 against its $5 standard rate — and Anthropic charges 1.25 times the input rate to write a five-minute cache, so caching pays for itself after a single read. Work that isn't time-sensitive can go through the Batch API for 50% off both input and output. Put the fixed text at the very front of the prompt to maximize cache hits.
- Budgeting on a launch promotion. New models often arrive at a temporary introductory rate that reverts on a published date, and aggregator pricing tables frequently list the promotional figure without flagging it. A budget built on a promotion breaks the day it lapses. This calculator and this page use each model's standard on-demand rate for exactly that reason — if you are quoting a lower figure you saw elsewhere, check whether it has an expiry attached before you put it in a forecast.
Frequently asked questions about tokens and AI API pricing
How much does GPT-5.6 Sol cost per 1M tokens?
GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens on standard on-demand pricing. That is the same rate as GPT-5.5, so the flagship tier did not get more expensive this generation.
What is the price difference between GPT-5.6 Sol, Terra and Luna?
GPT-5.6 shipped as three separate models on 9 July 2026, and they are exactly five times apart end to end. Sol, the flagship, is $5 input and $30 output per million. Terra, the balanced tier, is $2.50 and $15 — precisely half of Sol on both sides. Luna, the fastest and cheapest, is $1 and $6. All three keep the same six-to-one output-to-input ratio, so switching tier scales your whole bill up or down without changing its shape: the support chatbot in the examples above costs $75.00 a month on Sol, $37.50 on Terra and $15.00 on Luna.
How much does Claude Opus 5 cost per 1M tokens?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, the same rate as Claude Opus 4.8. That is half the price of Claude Fable 5 at $10 and $50.
Is GPT-5.6 Sol or Claude Opus 5 cheaper to run?
It depends entirely on your output share, because the two flagships charge the identical $5 per million for input and differ only on output — $30 for Sol against $25 for Opus 5. On a long-document job of 100,000 input and 2,000 output tokens they are one cent apart, $0.56 against $0.55. On an output-heavy coding agent at 20,000 input and 3,000 output tokens, 200 calls a day, Opus 5 comes out $90.00 a month and $1,095.00 a year cheaper — about 8.6%. Enter your own token mix rather than trusting either headline, because the ranking is decided by that ratio and nothing else.
Does a new model at the same price really cost the same?
Not necessarily, and this is the trap of the current generation. Anthropic's pricing documentation states that Claude 4.7 and later models — Opus 5, Opus 4.8 and Sonnet 5 among them — use a newer tokenizer that turns the same text into approximately 30% more tokens, while Sonnet 4.6 and earlier use the previous one. Because the rate card is unchanged at $3 and $15 for both Sonnet generations, an identical workload bills roughly 30% higher on Sonnet 5 than on Sonnet 4.6: $52.65 a month instead of $40.50 on the support-bot example. Rate parity is not cost parity. Re-measure your token counts on the new model before you commit.
Why does this calculator use Claude Sonnet 5's standard rate instead of its launch promotion?
Because a budget built on a promotion breaks when the promotion ends. Anthropic launched Sonnet 5 with a temporary introductory rate that reverts to the standard $3 input and $15 output per million on a published date, and many pricing tables quote the promotional figure without saying so. This calculator uses the standard rate — which is also exactly what Sonnet 4.6 charges — so the monthly and yearly numbers you get here stay valid after the promotion lapses. If you want to model the promotional window, choose "Custom" and enter the lower rate by hand.
What is the cheapest frontier model API?
Among frontier-class models, the cheapest entries are Grok 4.5 at $2 input and $6 output per million and Gemini 3.1 Pro at $2 and $12, with GPT-5.6 Terra just above at $2.50 and $15. Step down one tier and GPT-5.6 Luna at $1 and $6 or Claude Haiku 4.5 at $1 and $5 roughly halve that again. If raw price matters more than frontier capability, DeepSeek V4 Pro runs $0.435 and $0.87, Meta's Llama 4 Scout $0.11 and $0.34, and Amazon Nova Micro just $0.035 and $0.14. Two caveats: Gemini 3.1 Pro and Grok bill at a higher rate above a 200,000-token prompt, and the cheapest capable model for your task is decided by your evaluation set, not by the rate card.
How many tokens are in 1,000 words of English text?
About 1,333 tokens. The rule of thumb is 1 token per 0.75 words, or roughly 1 token per 4 characters of English — so a 750-word post is near 1,000 tokens. It's an approximation: code, numbers and non-English text use more tokens per word.
How much does 1M tokens cost across GPT, Claude and DeepSeek?
It depends on the model and on input versus output. A million input tokens runs from $0.035 on Amazon Nova Micro and $0.435 on DeepSeek V4 Pro up to $5 on GPT-5.6 Sol and Claude Opus 5, and $30 on GPT-5.5 Pro; a million output tokens, from $0.14 up to $180. The frontier band is tighter than the full range suggests — flagship-class models sit between $2 and $5 for input, with output anywhere from $6 on Grok 4.5 to $30 on GPT-5.6 Sol. The full breakdown is in the table above.
Why do output tokens cost more than input tokens?
Two to six times more, depending on the model. Generating text is sequential — the model emits one token, appends it, and computes the next — while input is read in a single parallel pass. So all three GPT-5.6 tiers charge six times as much for output as input, the Claude family exactly five times, Grok 4.5 three times, and DeepSeek V4 Pro roughly double. When you optimize, trimming output saves the most.
How much does an AI chatbot cost per month?
It depends on the model and the volume. For a typical bot at 2,000 input and 500 output tokens per call, 100 calls a day: GPT-5.6 Sol comes to $75.00 a month, or $912.50 a year; GPT-5.6 Terra $37.50 a month; Claude Opus 5 $67.50; Claude Sonnet 5 $40.50; GPT-5.6 Luna $15.00; Claude Haiku 4.5 $13.50; and DeepSeek V4 Pro just $3.92 a month, or $47.63 a year. Change the tokens and frequency in the calculator to match your own case — and if your monthly figure lands above roughly $200, the routing and caching levers below are worth a day of engineering time.
How do I count tokens exactly before calling the API?
Three ways. The most exact for OpenAI models is the tiktoken library (newer GPT models use the o200k_base encoding). The quickest is to paste your text into the estimator on this page — token count and per-model cost, no code. The third is the four-characters-per-token rule of thumb. Note that Claude, Gemini and Llama use their own tokenizers that aren't public, and that Claude's newest generation counts about 30% higher than the previous one, so in production the API's usage field is the reliable source.
What is the context window and how does it affect cost?
The context window is the largest number of tokens a model can handle in one request, input and output combined. GPT-5.6 Sol carries a 1,050,000-token window with up to 128,000 tokens of output, and Anthropic includes the full million-token window at standard pricing on Claude 4.6 and later, so a 900,000-token request bills at the same per-token rate as a 9,000-token one. Google's Gemini Pro tiers and xAI's Grok, by contrast, move to a higher rate above a 200,000-token prompt. Either way a big window doesn't make using all of it cheap: a long chat resends the whole history as input tokens on every single call.
Why does a short conversation burn so many tokens?
Because the API is stateless — each turn resends the whole prior conversation. Turn 10's input contains all nine earlier turns, so tokens grow almost quadratically and even a short chat lands in the tens of thousands of input tokens. To contain it: enable prompt caching, keep a sliding window of recent turns, or summarize the history periodically.
Is this token calculator free and private?
Yes — it's free and needs no signup. Text you paste into the estimator is processed in your own browser: nothing is sent to a server and no model quota is spent, so there's no risk of leaking your prompt. Prices come from the public rate cards of OpenAI, Anthropic, Google, DeepSeek and the rest, and the result is a budgeting estimate — the final number is always whatever each API console reports.
How accurate is this token cost estimate?
The arithmetic is exact: the calculator multiplies your token counts by the published per-million rates, so given the right inputs the per-request figure matches your invoice line. Three things move the real bill. The token counts themselves are approximate unless you take them from the API's usage field. The monthly projection uses a flat 30-day month, so a 31-day month runs about 3% higher and February about 7% lower. And discounts are not applied — prompt caching, the Batch API and enterprise agreements all bill below the standard rate. Treat the output as a well-grounded planning figure, not a quote.
Do images and files consume tokens?
Yes. Multimodal models slice each image into tiles and turn those into tokens — a single image can run from a few hundred to over a thousand tokens depending on resolution. PDFs, Word and Excel are usually converted to text first and billed as text. Vision-heavy apps such as OCR and screenshot analysis blow up the bill fastest, so measure real token use on a test set before you trust a budget.
I know my monthly cost — how do I bring the bill down?
Five levers, biggest first. One: route by difficulty — send simple tasks to GPT-5.6 Luna, Claude Haiku 4.5, GPT-5.4 Mini or DeepSeek V4 Pro and reserve GPT-5.6 Sol or Claude Opus 5 for the hard ones. Dropping from Sol to Luna alone is a five-fold cut. Two: turn on prompt caching for repeated prefixes, which bills cache hits at 10% of the input rate. Three: push non-urgent jobs through the Batch API for 50% off. Four: cap output with max_tokens, which is where the six-to-one output multiple hurts most. Five: tighten your system prompt. Re-enter the new numbers in the calculator to see the monthly and yearly savings.