Per-token pricing Paying for AI by the amount of text it reads and writes.
Short answer. Per-token pricing charges for AI by the amount of text processed: input tokens the model reads and output tokens it writes, priced per million tokens, with output several times dearer. Anthropic estimates one token at about four characters, or 0.75 words, of English, varying by language and content. On vendor pricing pages checked in October 2026, OpenAI’s gpt-6-sol costs US$2 per million input tokens and US$10 per million output tokens, as does Anthropic’s Claude Sonnet 5.5, while Google’s Gemini 3.8 Flash costs US$0.75 and US$3.75 until 31 December 2026, doubling from 1 January 2027. At US$2 and US$10 per million tokens, summarising a 25,000-word contract once costs about US$0.08. Costs climb with agents that resend long context at every step, and with long documents asked about again and again.
01
Prices per million tokens, as of October 2026
Model
Input (US$)
Output (US$)
Notes
OpenAI gpt-6-sol
2.00
10.00
Short-context standard rate; the batch rate is half
OpenAI gpt-5.4-mini
0.75
4.50
Smaller, cheaper model
Anthropic Claude Sonnet 5.5
2.00
10.00
Anthropic says its newer tokenizer produces about 30% more tokens for the same text
Anthropic Claude Haiku 5.5
0.10
0.50
For prompts up to 100,000 tokens
Google Gemini 3.8 Flash
0.75
3.75
Until 31 December 2026; US$1.50 and US$7.50 from 1 January 2027
02
A worked example
A 25,000-word contract is roughly 33,000 tokens at 0.75 words per token. Sending it once to a model priced at US$2 per million input tokens costs about US$0.07, and a 1,500-token summary at US$10 per million output tokens adds about US$0.015.
Ask ten follow-up questions that each resend the contract and the input cost is ten times higher. An agent that takes seven tool steps and resends the same context each time multiplies it again. A tokenizer that produces 30% more tokens for the same text raises every figure by 30%.
03
What drives the bill
Output costs more than input: five to six times as much on the rates above.
Context is resent: chat history and attached documents count as input again on every turn.
Language and content: token counts vary, so measure with your own documents.
Discounts: cached input is billed at lower rates on OpenAI’s and Anthropic’s price lists, and OpenAI’s batch rates are half its standard rates.
Price changes: Google’s page already lists a doubling of Gemini 3.8 Flash rates from 1 January 2027.
Data use differs by tier: Google’s pricing page shows free-tier use is used to improve its products and paid-tier use is not.
04
Per-token versus owning the hardware
Per-token pricing is cheap for occasional use and hard to predict for heavy use, because cost follows usage. Hardware you own costs the same whatever the token count, and the data stays in the office. ANVE runs its models on the device; an administrator can add cloud models, whose answers are labelled and whose calls are logged, so per-token costs arise only where an office chooses them.
05
Questions people ask
How many tokens is a page of text?
At Anthropic’s rough estimate of 0.75 English words per token, a 500-word page is about 670 tokens. Counts vary by language and content, so measure your own documents.
Why does output cost more than input?
Vendors price output tokens higher; on the rates above, output costs five to six times as much as input. Asking for shorter answers keeps bills down.
Can per-token costs run away?
Yes, especially with agents and long documents. Set a monthly budget and alerts if your vendor’s console offers them, and limit how many steps an agent may take.
Does a local model have per-token costs?
No vendor bills you per token. The costs are the hardware, electricity and upkeep, which do not change with the number of tokens processed.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
