Quick answer: every AI and LLM row on the Devotel Orbit pricing page prices one token, and as of November 2 the unit column says exactly that. The amounts did not move; the label did. This post walks what the previous 'per 1K tokens' label understated, how an AI charge composes from your account's input and output token rates, what one minute of AI voice costs once the telephony leg is in, and how to verify any line against your own statement.
1. What changed, and why the old label understated per-line rates
Until November 2, the unit column on the AI and LLM rows read 'per 1K tokens' while the listed amount was always the price of a single token. Read literally, the old label divided every listed rate by a thousand: a reader doing that arithmetic arrived at a per-token cost one thousand times below the rate actually applied to their usage. The label understated every AI line on the page and in the downloadable rate CSV, in every language both surfaces ship in.
The correction changes the label, not the math. Every AI and LLM row on the pricing page and in the rate CSV now names its unit as 'per token', in every language, and the rates applied to your usage, the token metering, and your statements are exactly what they were the day before. Two nuances were always real and stay real: several models carry separate input and output token rates, and some accounts carry account-specific pricing, so the rows that apply to you are the ones on your pricing page, not a figure quoted in any post.
2. How an AI charge actually composes
An AI charge is two multiplications and a sum. The platform counts the input tokens a run consumes (your prompt, history, and retrieved context) and the output tokens it produces, then applies your account's rates: input tokens times the input rate, plus output tokens times the output rate. A model with a single blended rate collapses to one multiplication; a model with split rates keeps two lanes, and your billing records show the per-direction token counts and the applied rate for each.
Verification takes two surfaces that now agree. The billing records list token counts and applied rates per model and per direction, and the monthly statement CSV on the Statements page rolls those charges into the AI agents line. Since the October 25 fix, the statement CSV reports the full AI agents figure, including AI work billed outside the standard per-message shape, so the file matches what the AI usage page shows. If you are reconstructing a month, that parity is what makes the CSV a source you can audit against rather than a summary that quietly drops lines. The changelog carries both corrections, dated November 2 and October 25.
3. Worked math: one minute of AI voice, all-in
A minute of AI voice is more than the model. The per-minute all-in figure is the telephony leg, plus speech-to-text and text-to-speech where those are metered separately on your account, plus the model's input and output tokens at your rates, plus an optional second voice leg when live interpretation is on. The component-math post prices each stage separately, and the human-agent comparison sets the resulting per-minute figure against a seated agent's fully loaded minute.
A worked example shows the shape, with numbers that are illustrative on purpose: substitute your account's CSV rates. One conversational minute runs roughly 900 input tokens of prompt and history and 250 output tokens of spoken reply. At an illustrative $0.000003 input rate and $0.000012 output rate, the model leg is 900 times $0.000003, or $0.0027, plus 250 times $0.000012, or $0.0030, about $0.0057 for the minute. Turn on live interpretation and you add a second voice leg plus a second model pass, which is why interpretation prices as an option on top of the base minute rather than folded into it. The cost breakdown and the voice AI cost page carry the same composition against the current published ranges.
4. How to read the AI rate CSV
Each AI row in the rate CSV names the model, the direction where a model splits input and output, the amount, and the unit column, and since November 2 that unit column reads 'per token' on every AI row. Reading a row is one multiplication per direction: the amount times the token count your billing records show for that model and direction.
The unit column is also what makes the CSV comparable across vendors, as long as you normalize first. Model providers most often publish per million tokens, and some publish per thousand. To compare any published figure with a per-token row, convert everything to one unit: a per-token amount times 1,000,000 is the per-million figure, so a $0.000003 row is $3.00 per million input tokens. Compare direction by direction, input against input and output against output, and compare your account's applicable rate against a vendor's list rate, because contract pricing moves both sides. Per-token is also the honest unit for telephony math: a one-minute call consumes hundreds of tokens, and per-million framing hides the multiplication a per-minute budget actually needs.
5. The honest caveat
The November 2 entry states its own limit plainly: the unit label alone does not confirm an individual invoice amount. The label tells you how to read a row; your invoice applies your account's rates to your measured input and output counts, and account-specific pricing means two accounts can read different amounts for the same model. The verification loop is the point of this post: read the unit as per token, multiply by your per-direction counts in the billing records, and reconcile the sum with the AI agents line on the statement CSV. When a line still does not reconcile, that gap is a support question with a concrete shape, not a suspicion about the label.
Frequently asked questions
Did the November 2 correction change what I pay?
No. It changed the unit column on the pricing page and in the rate CSV, in every language. Applied rates, token metering, and statements already computed per-token amounts.
Why keep a CSV if the pricing page shows the rates?
The CSV is the offline artifact for procurement: rows a reviewer can sort, diff against last quarter, and normalize against another vendor's published units. Page and file carry the same rates and the same unit column.
Which rate row applies to my account?
The one on your pricing page and in your CSV. Accounts with account-specific pricing see their contracted rates there, and the billing records show the rate that was actually applied, per model and per direction.
How do I compare a per-token row to a per-million publication?
Multiply the per-token amount by 1,000,000, or divide the per-million figure by the same, and compare one direction at a time. Normalize first, then name the list-rate versus account-rate difference instead of hiding it.
The takeaway
The corrected unit is a small change doing a large job: every AI row on the pricing page and in the rate CSV now says, in every language, that the listed amount prices one token. Charges compose as your account's input and output rates times measured counts, the per-minute all-in figure adds the voice leg and optional interpretation, and the statement CSV now agrees with the AI usage page. Read the rows as per token, normalize before comparing vendors, and verify against your own billing records, because the label tells you how to read a rate and your records tell you what you paid.
Published 2 November 2026.