Inference economicsPublic price bands and where they stop being useful
An order-of-magnitude reference: what public GPU-hour and token prices look like, and why a project estimate cannot be derived from a model's unit price. Figures checked on 9 November 2026. This is not our price list and not a commercial offer.
What this page is, and what it is not
The figures below are public bands from official price lists and industry reviews, checked on 9 November 2026. They are an order-of-magnitude estimate for an early architecture decision, not a quote for a specific project.
- This is not our price list, not an offer and not a ranking of tariffs: each segment bundles different availability, support and jurisdiction, so rows cannot be subtracted from one another.
- Taxes, cross-border payments, SLA commitments and integration work are not included.
- Conversions use a fixed reference of 1 USD = 84 RUB for comparing magnitudes only. It is not a settlement rate and does not imply RUB settlement or Russian closing documents.
- We do not publish a deployment price list: an estimate is built per process and per architecture, not per token price. The per-token rate is published separately, on the prices page.
GPU-hour price bands
Three segments, three different bundles. The table is for scale, not for picking a row.
| Segment | Public band | What to keep in mind |
|---|---|---|
| International spot marketplaces, H100 | 1.65–2.89 USD per GPU-hour | Spot: availability and instance persistence are not guaranteed by the tariff. |
| Russian providers, H100 | 350–500 RUB per GPU-hour (≈ 4.17–5.95 USD at 84 RUB/USD) | Local infrastructure and local jurisdiction; the bundle differs from spot. |
| Cloud.ru official rate | 10.60 USD per GPU-hour | A managed service: the price covers more than the hardware. |
The spread reflects different provider obligations, not a markup. Comparing rows without describing the perimeter produces a wrong estimate.
Token price ≠ process price
Public token rates help estimate a single request. A project budget is driven by something else.
| Model | Public rate per 1K tokens | What it helps estimate |
|---|---|---|
| GigaChat Lite | 0.07 RUB | Order of magnitude for classification and simple bulk operations. |
| GigaChat Pro | 0.50 RUB | Order of magnitude for heavier requests. |
One business action includes input cleaning and masking, cache hits, retries, routing between models and context length. Multiplying the number of actions by a token price does not work: the result depends on gateway architecture. We therefore estimate per process, fixing the responsibility boundary and scope of work, and do not give comparisons against local tariffs in correspondence.
Three trade-offs to settle before the estimate
- On-premise. Data stays inside the perimeter and the responsibility boundary is unambiguous. The strategic price of that choice is capital cost, your own operations team and the pace of model updates.
- Public API. Fast start and elasticity. The price of that choice is a cross-border data boundary and single-vendor dependency that you have to govern yourself.
- Hybrid gateway. The customer decides which request, with what cleaning, goes down which route — and can audit it. The price of that choice is that someone on the customer side must own the rules, logs and quotas.
The responsibility boundary has to sit where the data boundary sits.
Frequently asked questions
Are these your prices?
No. These are public bands of third-party providers, collected for an order-of-magnitude view. We do not publish our own prices on this page.
Why is there no price list?
The estimate depends on the process: cleaning volume, caching, retries, routing and the required availability level. A token price table contains none of those parameters.
Is 84 RUB per USD a settlement rate?
No. It is a fixed reference used only to compare rows in the table. It is not a settlement rate and does not imply RUB settlement or Russian closing documents.
Are taxes and SLA included?
No. The bands exclude taxes, cross-border payments, SLA commitments and integration work. Those appear only in a draft contract for a specific perimeter.
How do I get an estimate for our perimeter?
Describe the process: which data is involved, where it must stay, the routing modes you need and the expected load. We reply with scope of work and a responsibility boundary, not a tariff comparison.
How to cite this page
Links and reuse of the tables with attribution are welcome. Suggested form: “AIO Keys, ‘Inference economics: GPU-hour and token price bands’, November 2026”, with the access date. Please keep the order-of-magnitude caveat when reusing the numbers.
An estimate for your process
Send a description of the process and the perimeter. We reply with scope of work and a draft contract — no price list, no comparisons against local tariffs, and no delivery promises before a technical call.