AI token pricing is the half of the AI price list most wealth managers never read. Our companion snapshot on what AI costs independent wealth managers covers seat plans and industry systems; this piece covers what the same intelligence costs when it is metered by the token through the vendors’ APIs — the billing behind every portfolio management system, CRM or workflow that calls a model in the background. For the capabilities of the tools themselves, see our comparison of AI tools for wealth managers.
As before, every figure below is a published list price, checked on 17 August 2026; in other words, nothing here is an estimate.
What is a token, anyway?
Language models read and write text in fragments called tokens — a short word, part of a longer word, a punctuation mark. As a rule of thumb, a thousand tokens correspond to roughly 750 English words; German runs slightly heavier. Vendors meter both directions separately: the prompt and any documents you attach count as input, while the model’s answer counts as output — and output is, without exception in the table below, the expensive direction. Consequently, workloads that read much and write little — summarising, extraction, review — are cheap, while workloads that write at length are not.
AI token pricing at a glance
Prices are per million tokens, in US dollars, for the vendors’ current flagship and mid-range models. Notably, the two Chinese providers sit in the same table as everyone else — at the token level they are full competitors, not a footnote.
| Model (API) | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| OpenAI gpt-5.6-sol | US$2.50 | US$15.00 |
| OpenAI gpt-5.6-terra | US$1.00 | US$6.00 |
| Anthropic Claude Fable 5 | US$10.00 | US$50.00 |
| Anthropic Claude Opus 5 | US$5.00 | US$25.00 |
| Anthropic Claude Sonnet 5 | US$2.00 | US$10.00 |
| Google Gemini 3.1 Pro (preview) | US$2.00 | US$12.00 |
| Google Gemini 3.7 Flash | US$0.75 | US$3.75 |
| Mistral Large | US$0.50 | US$1.50 |
| DeepSeek deepseek-v4-pro | US$0.66 off-peak, US$1.32 peak | US$1.98 off-peak, US$3.96 peak |
| Alibaba Qwen3.5 Plus | US$0.40 | US$2.40 |
Sources: the API price pages of OpenAI, Anthropic, Google, Mistral, DeepSeek and Alibaba Cloud, all checked on 17 August 2026.
What a token buys
In effect, a million tokens is roughly 750,000 English words. For example, summarising a 10,000-token meeting transcript into a one-page note on a mid-range model costs about three US cents; doing it every working day for a year costs less than one month of a single seat subscription. Consequently, the token view explains the economics of the whole AI stack: seat plans and industry tools are packaged tokens plus interface, access controls, logging and support. That packaging is usually worth paying for — but it helps to know that the raw material is priced in cents, not francs.
Scale that to a firm and the numbers stay small, whatever the AI token pricing looks like per million. Suppose a ten-person firm runs its meeting summaries, e-mail drafts and a handful of long document reviews through an API each month — generously, some 15 million input and 3 million output tokens. On Claude Sonnet 5, that is about US$60 a month; on Gemini 3.7 Flash, closer to US$23; even on the most expensive row of the table it is about US$300. In other words, at realistic volumes the API bill is a rounding error next to the seat plans and systems in our companion snapshot — governance, not metering, is where the money and the risk sit.
The fine print: discounts and surcharges
In AI token pricing the list price is only the opening line, however. OpenAI, for example, halves prices for batch processing and discounts cached input by 90 per cent. In the other direction, it charges a 10 per cent uplift for data-residency endpoints on its newer models — a line item any Swiss firm weighing where its data is processed should read twice. Meanwhile, long prompts cost extra: Gemini 3.1 Pro moves from US$2.00 to US$4.00 per million input tokens above 200,000 tokens, and Qwen3.5 Plus rises similarly above 256,000. DeepSeek goes furthest and prices by the clock, charging double during Chinese peak hours. Prices move in both directions, though: Anthropic announced Sonnet 5’s rate as introductory and later made it the standard price.
The Chinese providers: cheapest tokens, hardest questions
DeepSeek and Alibaba’s Qwen undercut every Western vendor in the table on AI token pricing, and it is not close. For a Swiss wealth manager, however, the cheap rows are where the other columns of any due diligence do the work: neither vendor publishes enterprise seat plans, training guarantees or EU and Swiss residency options for its own hosted API. In regulated firms, therefore, these models tend to arrive — if at all — through Western cloud platforms with their own contracts and hosting, rather than through direct subscriptions.
AI token pricing for a Swiss wealth manager
Supervision does not care how the tokens are billed. FINMA’s Guidance 08/2024 expects an inventory, risk classification and clear responsibilities whether AI enters the firm through a seat plan or through an API buried in the portfolio system; likewise, the outsourcing file is due either way. Professional secrecy and Swiss data protection law follow the data, not the invoice — which is why the residency surcharges above matter more than a cent per million tokens. The wider Swiss rulebook is mapped in our overview of regulations in Swiss wealth management.
What token prices do not tell you
Finally, cheap tokens are not a strategy, and AI token pricing is not a plan. The arithmetic above says nothing about rollout, staff training, human review or audit trails — the costs that actually decide whether AI earns its keep in a wealth management firm. Equally, it says nothing about which tool fits which task. We have looked at the efficiency question in Maximising Efficiency with AI in Wealth Management and at the broader Swiss wealthtech landscape separately.
FAQ
Is AI token pricing relevant to a firm that only buys chat subscriptions?
Yes, AI token pricing matters indirectly but materially: every seat plan and every AI feature inside an industry system is a wrapper around these token prices. Knowing the raw cost therefore sharpens every vendor negotiation, because it shows what the packaging premium actually is.
Do cheap tokens change what may be sent to a model?
No. In Switzerland, professional secrecy, data protection law and FINMA’s governance expectations apply regardless of whether a token costs four cents or forty. In short, the price column never overrides the data column.
Which model should a wealth management firm pick?
The table alone cannot answer that, and it should not. Fit for the task, contract terms, hosting and controls decide; the token price then tells you what the decision costs. Above all, resist choosing a model the way one chooses the cheapest custodian — the savings are cents, and the risks are not.
How stable is AI token pricing?
Not very. Vendors have halved list prices within months, added surcharges and turned introductory rates into standard ones. Hence the date in this article: the snapshot is 17 August 2026, and anything older than a quarter deserves re-checking.
Published list prices, checked on 17 August 2026. This article reflects personal views, simplifies where necessary and is not investment, legal or tax advice. Verify current prices and terms with the vendors before deciding anything.
