Token Bench Index

What everyday AI work costs, model by model. The same four ordinary tasks, priced every month on each major model using real token counts and verified list prices, so changes in cost are changes in the market.

Finding of the month: Gemini 4 reshuffles the language tax

Gemini 4 Argon uses a new tokenizer. On the English tasks it counts within one token of Gemini 3. On Hindi it counts more, and in this basket it costs more than Gemini 3 in every Indian language measured.

A four-task basket is a small sample, so we checked the finding on 1,012 professionally translated sentences per language, counted on both models on 31 October 2026.

Argon needs more tokens for ten of the thirteen Indian languages measured, including the five most widely spoken. It needs fewer for Odia and Punjabi, and about the same for Nepali. Argon’s price doubles to $4 / $20 per million tokens on 1 December 2026, so the extra tokens will cost twice as much from next month. Full results are in our Indian languages study.

Sentences from the FLORES+ benchmark, used under the CC BY-SA 4.0 licence. Only token counts are reported here, not the sentences.

Indian-language sub-index

What the customer-support task costs in each language, as a multiple of the same task in English. Exact counts only.

Estimated only

These providers publish no tokenizer, so their counts can only be estimated from the number of characters. Estimates are least reliable for non-English text, so these models are kept out of the headline figures and the sub-index.

Price notes

• Gemini 4 Argon: price rises to $4 / $20 on 1 December 2026.
• Gemini 3.8 Flash: price rises to $1.50 / $7.50 on 1 January 2027.
• DeepSeek V4 Pro and V4.1 Flash: the off-peak rate is shown; peak hours cost double. The Search and Fetch tool fees DeepSeek introduced on 1 November 2026 are not token prices and don’t affect the index.

Method

The basket is four everyday tasks: an English customer-support reply, a Python code review, an invoice extraction to JSON, and a Hindi translation of the customer-support task. Each has a prompt and a reference answer, so the index covers what a model reads and what it writes. The basket stays fixed, so later editions are directly comparable with this one. Its full text is in the PDF edition.

The index value for each model is the cost of running all four tasks 1,000 times: 1,000 × Σ (prompt tokens × input price + answer tokens × output price) ÷ 1,000,000.

Token counts come from each provider’s own tokenizer: OpenAI’s o200k tokenizer, run in-house, and the counting services of Anthropic and Google, minus the fixed overhead each adds to every request.

Prices are standard list prices for real-time use, without caching, checked against each provider’s official pricing page in the week before each edition. Where a price change has been announced, the price in force on the edition date is used. Secondary sources are never used for prices.

What the index doesn’t measure. Quality: a cheaper model may do a task worse, or need more attempts. Hidden reasoning: reasoning models also bill their internal thinking as output, so real bills from reasoning models will be higher.

Past editions

This is the first edition. A new edition is published on the 1st of every month.