AI Token Cost by Language: What Five Vendors Actually Publish

Last updated: August 11, 2026

AI token cost by language is the change in billable token volume when the same content is written in a different language. Nine public pages from five vendors were read on 11 August 2026. All five publish a token figure for English, and four of those figures sit within a narrow band. Two publish a figure for Chinese. None publishes one for Arabic, and no page in the set supplies a figure that compares two languages against equivalent content.

Verdict in one line: the published figures converge on English and fragment past it, so count your own text against the exact model before you commit.

Where the five vendors agree

Read the five side by side and the first thing that stands out is how closely they agree, as long as the language is English.

Anthropic’s pricing page states that one token is approximately four characters or 0.75 words in English. OpenAI’s help center lists its numbers under the heading “Helpful rules of thumb for English”, starting with four characters and three quarters of a word. Google’s token documentation gives “a token is equivalent to about 4 characters”, then adds 100 tokens for roughly 60 to 80 English words. Tencent Cloud’s billing page puts English at about 0.75 words per token. DeepSeek states 0.3 tokens per English character, which works out near three and a half characters per token.

Five vendors, five separate documents, and one language covered by all of them in figures that sit within a narrow band.

Chinese has a smaller version of the same agreement. Tencent gives about 1.8 Chinese characters per token. DeepSeek gives about 0.6 tokens per Chinese character, which is roughly 1.7 characters per token. Two vendors, two figures, close together.

Arabic has none. Across the nine pages, no vendor publishes a characters-per-token or words-per-token figure for Arabic. The two Arabic numbers in the set measure something else. OpenAI’s 2024 announcement counts one sample sentence under two tokenizer generations. Anthropic’s multilingual page scores Arabic against an English performance baseline.

The published record therefore has three tiers. A buyer working in English can sanity-check an estimate against five vendors. A buyer working in Chinese can do it against two. A buyer working in Arabic has nothing published to check against.

Sources: Anthropic, accessed 11 August 2026 · OpenAI, accessed 11 August 2026 · Google, updated 30 July 2026 · Tencent Cloud, 11 August 2026 · DeepSeek, accessed 11 August 2026

What each page publishes

The table records what each figure measures rather than how detailed it looks. Denominators matter more than decimal places, because two figures in different denominators cannot be divided into each other.

Covers the nine pages read on 11 August 2026. Pages not read are outside this table and no result is claimed for them.

Two figures carry the most weight for a non-English buyer, and both come from Chinese vendors. Tencent’s billing page states “中文约 1.8 字符 ≈ 1 Token，英文约 0.75 单词 ≈ 1 Token”, which FSR renders as roughly 1.8 Chinese characters per token and roughly 0.75 English words per token. DeepSeek’s token page states “1 English character ≈ 0.3 token. 1 Chinese character ≈ 0.6 token.”

Sources: Tencent Cloud, 11 August 2026 · DeepSeek, accessed 11 August 2026 · Google pricing, accessed 11 August 2026 · Anthropic, accessed 11 August 2026 · OpenAI, 13 May 2024

Agreement is not comparison

Two vendors publishing the same kind of figure for the same language is useful. It is not the calculation a multilingual buyer needs, and the gap between the two is where budgets go wrong.

Start with Chinese. Tencent’s 1.8 characters per token and DeepSeek’s 1.7 sit close enough to place side by side. Neither tells a buyer what a document costs in Chinese against the same document in English, because characters are not units of content. An English sentence and its Chinese translation contain very different character counts, and neither page supplies the bridge between them.

Tencent’s own line makes the problem visible in a single sentence. Chinese is expressed in characters and English in words. Converting one into the other requires a words-to-characters ratio for equivalent content, and the page does not give one.

The English agreement has the opposite shape. Four vendors converge because they are all describing the same language, so the figures line up and say nothing about any other.

DeepSeek states the remaining limit on its own page: conversion ratios vary by model, and the number that counts is the one returned in the usage results. The underlying methodological point is not new. Peer-reviewed work since 2023 has held that cross-language cost comparison requires content-equivalent text rather than raw character counts, and this briefing does not add to that literature.

Sources: Tencent Cloud, 11 August 2026 · DeepSeek, accessed 11 August 2026 · Petrov et al., 2023

Precise about audio, approximate about text

Google’s Gemini Developer API pricing page shows how exact a vendor can be about token counts when the input is not text.

Audio on a text-to-speech model counts at 25 tokens per second. Video output on one model is calculated at 5,792 tokens per second at 720p. An output image at 1,024 by 1,024 pixels consumes 1,120 tokens, and image input on another model is fixed at 560 tokens per image. Google’s token documentation adds 258 tokens for an image at or below 384 pixels in both dimensions, 263 tokens per second of video, and 32 tokens per second of audio for the models it covers.

Text is priced on that same page, per million tokens, alongside everything else. What no Google page in this set supplies is a text-to-token conversion for any named language other than the English figures already quoted.

The practical difference is worth stating plainly. A team budgeting an audio workload can convert duration into tokens from published numbers before signing anything. A team budgeting an Arabic text workload has no published conversion to work from.

Sources: Google pricing, accessed 11 August 2026 · Google tokens, updated 30 July 2026

When a tokenizer change moves the bill

Language is not the only variable that moves token volume while the rate card holds still.

Anthropic’s pricing page states that Claude 4.7 and later models use a newer tokenizer, and that “this tokenizer produces approximately 30% more tokens for the same text”. The page qualifies the figure by content and workload shape and notes that Sonnet 4.6 and earlier models use the previous tokenizer. The token counting documentation repeats the number, adds that usage and billing reflect the newer counts, and tells developers to recount rather than reuse older measurements.

The same pricing page lists a second set of numbers that differ between generations. Tool use adds a fixed system prompt to every request: 497 to 589 tokens on Opus 4.6, and 675 to 804 on Opus 4.7. The page reports the counts per model without attributing the difference to a single cause.

Migration is therefore a budget event in its own right. A workload moved between generations without recounting can consume more billable tokens on identical input while the published rate per million stays where it was.

Length thresholds behave differently across the set. Anthropic’s long context section states that a 900,000-token request is billed at the same per-token rate as a 9,000-token request on current models. Google prices some models in bands: Gemini 3.1 Pro Preview lists $2.00 per million input tokens for prompts at or below 200,000 tokens and $4.00 above that line, and Gemini 2.5 Pro carries the same threshold. Tencent’s billing page notes that some models use segmented pricing by input length. Where a band exists, the number of tokens a given source produces decides which side of it a request lands on.

Sources: Anthropic pricing, accessed 11 August 2026 · Anthropic token counting, accessed 11 August 2026 · Google, accessed 11 August 2026 · Tencent Cloud, 11 August 2026

The unit on the invoice

A price table denominated in tokens does not guarantee an invoice denominated in tokens.

Anthropic’s pricing page documents this for two routes. Claude Platform on AWS bills through AWS Marketplace in Claude Consumption Units: token usage is rated in USD at standard rates, discounts are applied, the result converts to CCUs at one cent each, and the AWS bill shows a single CCU line item. Claude in Microsoft Foundry follows the same structure through the Azure Marketplace. The page states that 100 CCU represents one US dollar of fees.

Google’s pricing page runs several units at once. Text and multimodal input bill per million tokens. Veo video bills per second of output. Imagen bills per image. Lyria bills per song. Search grounding bills per thousand requests. Context cache storage bills per million tokens per hour. The page does not address whether language affects any of them.

Tencent’s billing page does address it, three times over. The language estimate covers text models and not vision models. The voice section states that the language model token estimate does not apply there. The video section says the same, and adds that some model families bill per second or per credit per second rather than per token.

Compare quotes on the route you will buy on. The unit used to calculate a charge and the unit printed on an invoice can differ, and only one vendor in this set tells a reader where its language estimate stops applying.

Sources: Anthropic, accessed 11 August 2026 · Google, accessed 11 August 2026 · Tencent Cloud, 11 August 2026

Arabic in the reviewed set

No characters-per-token or words-per-token figure for Arabic appeared in the nine pages. Chinese conversion estimates appear on two of them, both from vendors whose documentation is published in Chinese alongside English.

Two Arabic numbers do sit in the set, and neither answers a budgeting question. OpenAI’s 2024 announcement carries a table of 20 languages with two counts on each row, and the Arabic row reads “Arabic 2.0x fewer tokens (from 53 to 26)”. Those counts describe one sample sentence under an earlier tokenizer and under the one introduced with that model, and the English row on the same table shows 27 and 24. The comparison runs within each row rather than down the column. Anthropic’s multilingual page places Arabic at 97.2 percent of its English performance baseline on Claude Sonnet 4.5 and 92.5 percent on Claude Haiku 4.5, which is a capability score.

For an Arabic workload, that leaves counting as the first route to a number. Two vendors publish counting that needs no account: OpenAI offers an interactive tokenizer and the open-source tiktoken library, and DeepSeek offers a downloadable tokenizer package. On the pages reviewed, Anthropic and Google document counting through an API endpoint.

FSR records the placement and stops there. Explaining why a vendor names one language rather than another would need evidence this review does not hold.

Sources: OpenAI, 13 May 2024 · OpenAI help center, accessed 11 August 2026 · Anthropic, accessed 11 August 2026 · DeepSeek, accessed 11 August 2026

How to build the budget instead

Four of the five vendors document a counting route, and two of them tell readers in plain words to use it rather than a rule of thumb. The steps below are FSR’s, built from what those pages establish.

Fix the purchase first: exact model ID or snapshot, API route, region or inference geography, and pricing tier. Each of these appears in at least one reviewed page as something that changes either the count or the rate.

Assemble a real sample of the workload rather than sample prose. Anthropic’s pricing page shows why: the tool system prompt alone adds several hundred tokens per request, before any of your own content is counted.

Build semantically aligned versions in every deployment language, then count the same request against each target model’s current tokenizer or counting interface. Apply input, output, cache write, cache read, and any regional or fast-mode modifier separately rather than as one blended rate.

Recount before any model migration. Anthropic’s published 30 percent figure is the clearest documented case in this set of a count moving while a rate did not.

Three questions belong in the procurement record, because the reviewed pages answer them only in part. Which tokenizer applies to the contracted model. Whether that tokenizer can change without a rate card change, and what notice applies. Which usage field or converted unit controls the invoice on the route being bought.

Sources: OpenAI, accessed 11 August 2026 · DeepSeek, accessed 11 August 2026 · Anthropic, accessed 11 August 2026 · Google, updated 30 July 2026

Limits of this review

FSR measured nothing and publishes no multiplier for any language pair. How many tokens Arabic or Chinese consumes relative to English has been studied in peer-reviewed work since 2023, including research finding that parallel text can differ substantially in tokenized length, with consequences for cost, latency, and context capacity.

Nine pages at five vendors were read. OpenAI’s developer API pricing page, its consumer and business plan pages, DeepSeek’s pricing page, terms of service, rate limit documentation, and vendors not named were not part of this set, and no result is claimed for them. Logged-in checkout, console screens, enterprise order forms, and negotiated contracts sit outside what a document review can reach and are excluded rather than reported as empty.

The figures quoted here are vendor statements. FSR did not independently validate any of them, and each vendor page attaches its own qualifiers. A fixed ratio per language is not a stable object either: DeepSeek states that ratios vary by model, and Anthropic qualifies its 30 percent figure by content and workload shape.

Sources: DeepSeek, accessed 11 August 2026 · Anthropic, accessed 11 August 2026 · Petrov et al., 2023

FAQ

Methodology

Tier C declaration. No paid account. No purchase. No invoice, dashboard, console, or checkout screen. No product test. No token measurement. Statements about a vendor describe page content on the access date. Reasoning, buyer recommendations, and the classification scheme are FSR’s own.

Pages read, with dates. Tencent Cloud TokenHub billing methods, page last updated 11 August 2026. Anthropic pricing, token counting, and multilingual support, no page dates visible, accessed 11 August 2026. OpenAI GPT-4o announcement, page dated 13 May 2024. OpenAI help center token article, page states it was updated 13 days before access. Google Gemini Developer API pricing, accessed 11 August 2026. Google token documentation, page last updated 30 July 2026. DeepSeek Token and Token Usage, no page date visible, accessed 11 August 2026.

Classification. Pages were sorted by which languages each figure names and which denominator it uses. The scheme was written for this review after the source set was assembled, and it is stated here so a reader can apply it to a different set.

Pages located but excluded. DeepSeek’s Models and Pricing page was retrieved during preparation. A later independent attempt to open the same address returned a different page in the same documentation set, so this briefing builds no claim on it. OpenAI’s business and enterprise plan pricing was reviewed during preparation and is excluded because a stable public address for the version read was not recorded.

Out of scope, declared rather than reported as empty. Logged-in checkout, console screens, enterprise order forms, individual quotations, negotiated contracts, terms of service, rate limit documentation, and OpenAI’s developer API pricing page.

Vendors not reviewed. Other vendors publish language-related token guidance and were not part of this set. Their absence from the table is a scope boundary rather than a result.

Translation. The Tencent Cloud page is published in Chinese, and the English rendering of the quoted line is FSR’s own. Figures read in a localized version of any page were checked against the English version of the same page.

Prior research. Cross-language token disparity and its cost consequences have been established in peer-reviewed literature since 2023. This briefing does not restate those measurements as its own findings.

Deliberately excluded. No measurement, no multiplier for any language pair, no consumer subscription cap analysis, and no characterization of any vendor’s disclosure practice as a compliance matter.

Affiliate posture. FSR holds no affiliate relationship with any vendor named and received no compensation connected to this briefing.

Sources: Tencent Cloud, 11 August 2026 · Anthropic pricing · Anthropic token counting · Anthropic multilingual · OpenAI, 13 May 2024 · OpenAI help center · Google pricing · Google tokens, 30 July 2026 · DeepSeek tokens

Verdict

Do not pick a vendor from a published language conversion rule. Across nine pages read on 11 August 2026, all five vendors publish a token figure for English, and four of those figures land in the same narrow range. Two publish a figure for Chinese that also broadly agrees. Nobody publishes one for Arabic, and no page in the set connects any two languages through equivalent content, which is the calculation a multilingual budget actually needs. Google’s pricing page will convert a second of audio into tokens exactly, and offers nothing equivalent for a paragraph of text in any language. Use the published figures to find where your uncertainty sits, then count a real sample of your own workload against the exact model, route, and tier you intend to buy, and count it again before any migration.

Sources: all nine pages listed under Methodology, read 11 August 2026.