Token
In one sentence
A token is the unit a language model reads and writes, usually a fragment of a word rather than a whole word, and cost, context window capacity and latency are all measured in it.
Definition
A token is a chunk of text, usually part of a word rather than a whole one. Everything a language model does is priced and limited by the token.
The practical consequence is that long instructions and long retrieved passages cost money on every single exchange, not once.
Tokenization converts text into a sequence of integers the model can process, using a vocabulary learned from its training data. Common words are usually a single token, uncommon words split into multiple fragments, and whitespace and punctuation are typically bundled in with the adjacent text. For English, a rough working average is around four characters per token, though this varies substantially by content.
Why the fragment behavior matters commercially
- Text in languages underrepresented in the training vocabulary tokenizes less efficiently, sometimes requiring several times more tokens for equivalent content.
- This means the same conversation can cost materially more in one language than another, a real and under-discussed equity issue in multilingual deployment.
- Domain jargon, product codes and proper nouns tokenize poorly, inflating both cost and latency.
What tokens govern
- Cost. Providers price per token, input and output separately.
- Context window capacity, which is measured in tokens.
- Latency. Prefill time scales with input tokens, and generation time scales with output tokens.
Audio tokens, which are different
- Speech-to-speech and realtime models tokenize audio directly, using schemes that differ from text tokenization.
- Audio token pricing is not directly comparable across providers because the tokenization differs, which makes rate card comparison misleading.
- Cost modeling for audio-native pipelines requires measurement on representative traffic.
The practical discipline
- System prompt size is paid on every single turn. A bloated prompt is a recurring, invisible cost.
- Retrieved context size is the other major controllable input. Fewer, better passages reduce cost and latency and improve answer quality at the same time.
Common misconception
That a token is a word. Token counts run higher than word counts, and the ratio varies by language and content type, which makes word-based cost estimates unreliable.
Why it matters commercially
Tokens are the unit of cost in a voice pipeline that already has real variable costs. Prompt and retrieval discipline is a margin lever available without any quality sacrifice.
In voice specifically
Speech-to-speech and realtime voice models tokenize audio directly, with schemes that differ from text tokenization and from each other. That is why a published audio rate card cannot be compared across providers the way a text rate card can, and why an audio-native pipeline has to be costed by measuring representative calls rather than by reading a price per token.
Where AsqVox fits
Retrieved passage size directly affects token count, which affects both cost and time to first token. In AsqVox, retrieval efficiency is therefore a latency and margin lever, not only a quality one.
Visual
What is paid on every turn
The system prompt is paid on every single turn. A bloated prompt is a recurring invisible cost.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
For English text a working approximation is around four characters per token, though this varies substantially by content type, and uncommon words, proper nouns and domain-specific codes split into multiple tokens.
about 4 characters per tokenindustry rangeTokenizer behavior, engineering convention, 2026 - A rule of thumb, not a constant. The ratio moves with the tokenizer and the content, so use it to estimate, not to bill.
All-in production voice cost lands around USD 0.11 to USD 0.33 per minute, with model inference priced per token as a major component.
USD 0.11 to 0.33 per minuteindustry rangeIndustry-reported production voice cost range, 2026 - The per-token line is the one prompt and retrieval discipline can move without touching quality.
Platform base rates run roughly Vapi around USD 0.05 per minute, Retell around USD 0.07, and Bland around USD 0.09.
USD 0.05 to 0.09 per minutevendor claimVapi, Retell and Bland published base rates, 2026 - Vendor rate cards change frequently. Date any figure and check it against the current pricing page before quoting it.
Audio token pricing is not directly comparable across providers because tokenization schemes differ. Cost comparison requires modeling on representative traffic rather than reading rate cards.
-no reliable figureThe units are not the same across providers, so a lower audio price per token can still cost more per minute. Only representative traffic tells you which is cheaper.
Provider token pricing changes frequently. Any published rate should be dated and linked to the current pricing page.
-no reliable figureThis is why the platform rates above carry a year and a caveat rather than being treated as durable facts.
There is no published comparison of tokenization efficiency across languages for current models, despite the cost equity implications.
-no reliable figureThe multilingual cost penalty is real and widely acknowledged but not captured in a citable cross-language benchmark. Measure it on your own content.
Examples
In practice
A team reviews per-customer margin and finds one account unprofitable at the standard rate. Its system prompt is roughly three times the platform average, and that prompt is processed on every turn of every conversation. Trimming it restores margin without touching the price or the model, and reduces time to first token as a side effect.
The everyday version
A token is a chunk of text, usually part of a word rather than a whole one. Everything is charged by the token. The practical consequence is that long instructions and long retrieved passages cost money on every single exchange, not once.
Usage
Who says it
- AI engineers, constantly, in cost and context discussions.
- Finance teams, in cost modeling.
- Every provider pricing page.
Where it turns up
- On a spec sheet, next to pricing, context window size, rate limits and cost per interaction.
Common misuse
- Estimating cost from word counts.
- Comparing audio token pricing across providers as though the units match.
- Ignoring system prompt size as a recurring cost driver.
Questions people ask
Is a token the same as a word?
No. Token counts run higher than word counts. A rough English average is about four characters per token, but the ratio varies by language and content type, so estimating cost from word counts is unreliable. Uncommon words and proper nouns split into several tokens each.
Why does the same conversation cost more in another language?
Because languages underrepresented in the training vocabulary tokenize less efficiently, sometimes needing several times more tokens for equivalent content. Since cost is charged per token, the same conversation can be materially more expensive in one language than another. It is a real and under-discussed equity issue in multilingual deployment.
What makes a voice agent expensive per turn?
Two controllable inputs. The system prompt is processed on every single turn, so a bloated one is a recurring invisible cost. Retrieved context is the other, and passing too many passages inflates it. Trimming both cuts cost and latency at once and can improve answer quality rather than hurt it.
Can I compare audio token prices across providers?
No, not directly. Speech-to-speech and realtime models tokenize audio with schemes that differ from provider to provider, so the units are not the same and a lower price per token can still cost more per minute. Cost an audio-native pipeline by measuring representative traffic, not by reading rate cards.
Last reviewed 3 August 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.