Understand Compute token pricing and billing units
What Compute charges per token, how text, code, documents and media are counted, and how your quote sets the most you can be charged.
Start with the tariff and one estimate
Compute has one customer tariff: USD 0.05 per 1,000,000 input tokens and USD 0.10 per 1,000,000 output tokens. The rates apply to the normalized token counts in the applicable metering recipe. Normal platform retries and hedges, default verification, delivery and repair are included. There are no page, image, character, minute, hour, base, minimum, surge, urgency, region or model price lines.
Lithi measures text, code, documents and media in tokens. The metering profile defines how admitted input and accepted output are counted.
For a code-review illustration, assume the Lithi metering recipe records 800,000 input tokens and
100,000 output tokens. The processing estimate is 0.8 × USD 0.05 + 0.1 × USD 0.10 = USD 0.05.
This is planning arithmetic, not a quote or a charge; Get a free estimate is
non-binding and grants no access or execution authority.
The example uses assumed Lithi-metered quantities. A third-party provider's token count is not interchangeable with a Compute count. The current signed quote remains the authority for the exact work, metering profile, accepted quantities and maximum charge. A message, page, image, minute or file is a workload quantity; it is not automatically a Compute billing unit.
Turn your work into an estimate
Tell your agent the result you want, the file type, the duration or length, the item count, and the languages involved. The agent counts permitted input locally, chooses supported operations, links one stage's output to the next, and uses a published assumption when no measurement exists. The result is an input and output allowance for planning until the work is produced; it is not a quote.
| Work you describe | What informs the estimate |
|---|---|
| Prepared text or code | Your agent measures the complete prepared input once with the released tokenizer and chat template. Character and file-size ratios do not replace that count. |
| Documents | Extracted text is counted; page visuals are counted only when the selected operation reads them. |
| Audio | Decoded duration informs the media allowance. |
| Images and video | Admitted dimensions, sampling and processing passes inform the allowance. |
| Archives and repositories | Sum only bounded, admitted archive entries or included repository files; the container and excluded files add no tokens. |
You do not need to guess a token quantity. Give the result, file type, duration or length, count, and languages so the agent can identify what is missing and keep each operation's assumptions visible. The released tokenizer and metering profile remain the source of the estimate's exact token count when the prepared input is available.
Private files are measured locally by your agent on the machine authorized to read them. A chat can accept a bounded count or content descriptor that you provide, but it cannot inspect an uploaded or private file itself. The complete prompt is counted once, including its chat-template framing; an embedded value is not counted again as a separate part. Canonical metering then sums each admitted content descriptor under its family. For a document, native page text and page visuals are separate derivatives, so a page image contributes only when the operation reads it. For an archive, the container contributes zero and its admitted entries are counted under their own families. For a repository, each included file contributes its path and text; excluded files contribute zero.
Read the unit before comparing the price
A price is meaningful only with its unit, denominator, currency and included work. A page, minute, task and token are not interchangeable. Read the signed quote for the operation you intend to use; an old example is not evidence for the Compute charge.
| Detail | What to check | Why it matters |
|---|---|---|
| Work and operation | The quoted workflow, input bounds and requested outputs match your intent. | Similar task names can cover different work. |
| Billing unit | The accepted input and output token quantities produced by the quoted metering recipe. | File count, messages, media duration and billing quantity can differ. |
| Denominator | The fixed 1,000,000-token denominator for each rate. | A provider's denominator or tokenizer is not a Compute denominator. |
| Currency and amount scale | USD dollars, cents or millionths of a dollar, as explicitly labeled. | Misreading an integer scale changes the amount by orders of magnitude. |
| Included processing | Preparation, generation, verification and delivery included in that quote. | Do not add a second charge for something already included. |
| Precision and rounding | The metering recipe, integer field scale and money rounding rules in the quote. | Multiplication alone may not give the final quoted amount. |
| Expected and maximum charge | The estimate, the approved ceiling and any current account limit. | An estimate does not authorize work beyond the approved ceiling. |
| Validity | The quote's expiry and any change to the work it covers. | A price preview or expired quote does not authorize admission. |
Keep units and amounts separate
The metering definition describes how measured content becomes billable token quantities. The rate and metering definition travel with the quote. A capability in a reference catalog does not mean every operation or profile is available to your account.
Input and output token quantities are separate measures. The applicable recipe determines what input framing is counted and which output is accepted. Do not price visible text alone when the quoted input includes instructions or formatting. Media uses its own recipe to produce token quantities; do not substitute a provider's counter or convert messages, pages, images or minutes by assumption.
The recipe applies to the quoted work and its accepted quantities. The quote covers the full instructions, context, output schema and evidence requirements used to determine metering. A customer agent may prepare or review those details in a task brief, but the brief does not grant access or approve execution. A historical usage statement can explain what a past job consumed; it does not become the current account balance, a current quote or permission to spend. Read current usage and the active quote separately when deciding what to run.
For API readers, the unit is part of the field contract. For example, maximum_charge_micro_usd and gross_micro_usd are amounts in millionths of a US dollar; amount_cents is in cents. These interfaces use bounded integer values, not a universal decimal-string format. Preserve the field's declared representation and do not round intermediate calculations with binary floating point.
Keep estimates, quotes and charges separate
The USD 0.05 illustration above is synthetic arithmetic, not a transaction. A free estimate helps you assess a workload and creates no access, reservation or execution authority. A signed quote sets the exact scope and maximum charge; accepted usage is settled from accepted logical work under that quote. The same amount may appear as dollars, cents or micro-USD, which are representations of one amount, not three charges. Never sum them together. Likewise, a reserved maximum is not usage already consumed. Anthropic's chat operating cost is separate from the customer Compute tariff.
Understand partial work and retries
Count accepted work under the quote, not every physical attempt. Normal platform retries, hedges, default verification, delivery and repair are included and do not create another customer charge. Distinct requested candidates and comparisons remain distinct work when accepted under their quoted policy. An abstention, failure, cancellation or partial result needs the applicable policy; none supports a blanket promise that every unsuccessful run is free.
If the metering recipe or inclusion rules are unclear, stop before paid approval. Keep the quote reference and ask for clarification rather than estimating from a different service's price. After execution, reconcile the recorded token quantities and rates with the quote and investigate differences before importing an invoice. Read Booking and priority when timing or reserved capacity affects the quote, then use Usage and statements to follow accepted work through the billing records.
Compare published prices for the same work
Lithi charges per input and output token. The table below applies each provider's published batch rate to the same token counts for one labelling job, so you can see the same work side by side. Prices were checked on the date shown, and each links to the provider's own page.
A third-party rate is not the Compute charge. Your own decision still uses a signed quote and your own measured costs; read Cost fit for that method.
Published list prices checked 5 October 2026.
| Provider | Product | Quantity | Total | Note |
|---|---|---|---|---|
| Lithi | Lithi Compute | 9,600,000,000 input + 1,200,000,000 output tokens | $600.00 | |
| OpenAI | gpt-6-luna (Batch) | 9,600,000,000 input + 1,200,000,000 output tokens | $780.00 | |
| Gemini 3.1 Flash-Lite (Batch) | 9,600,000,000 input + 1,200,000,000 output tokens | $2,100.00 | ||
| Anthropic | Claude Haiku 4.5 (Batch) | 9,600,000,000 input + 1,200,000,000 output tokens | $7,800.00 | |
| AWS | Amazon SageMaker Ground Truth automated labelling | 1,000,000 objects | $42,000.00 | An automated labelling service with its own workflow, not an AI-model call. |
| Google Cloud | Google Cloud data labeling | Not applicable | No current price | Google Cloud publishes no current price for an automated labelling service. |
Each provider counts tokens with its own tokenizer, so the same text can be a different number of tokens. Rows without a note apply each published rate to the same token counts. Each service bills in its own unit. These tables compare published prices for the same work, not output quality.
The same comparison for other work lives with each workflow:
- Dataset evaluation: compare published prices
- Translation: compare published prices
- Transcription, scenes, keyframes and clips: compare published prices
- Image analysis: compare published prices