AI model prices
Model and provider identifiers, tariff variants, and available input, output or cache charges. Every component keeps its billing unit.
MODEL CONTEXT PROTOCOL
MCP Server for AI & GPU Prices
Give your AI assistant a direct connection to the latest archived AI model and cloud GPU prices. Public, read-only and ready for compatible MCP clients.
YOUR CONNECTION POINT Public · Read-only
One link. A clearer view of AI pricing.
Add the xpu live MCP server to a compatible assistant to explore the latest public model and GPU prices in our archive.
https://xpu.live/mcp-serverStreamable HTTPNo API key requiredLatest archived observations
Connection instructions
BUILT FOR USEFUL ANSWERS
What your assistant receives
Public pricing data with the context needed to make a meaningful comparison.
Cloud GPU prices
GPU rental tariffs, providers, rental types, currencies and billing conditions from the latest successful provider collections.
Sources & conditions
Source links, exact tariff identifiers, units and conditions travel with each price. Missing observations are never treated as free.
Freshness & collection status
Observation timestamps, last successful collection times, latest attempt status and stale flags help your assistant judge the data.
FOUR FOCUSED TOOLS
Find prices. Keep the context.
Use these MCP server tools individually or combine them in an assistant workflow.
search_latest_prices
Search the latest public model or GPU tariffs. Filter by provider, resource, currency, rental type or freshness.
View parameters →get_latest_price
Retrieve an exact tariff from a search result, including price components, source, conditions and observation time.
View parameters →list_price_providers
Discover provider IDs, public price counts and collection status. New providers flow through the shared archive automatically.
View parameters →get_price_freshness
Check stale observations and provider collection timestamps before using an archived price in an estimate.
View parameters →GET CONNECTED
Set up the xpu live MCP server
You need an assistant or MCP client that supports remote Streamable HTTP servers.
- Add a remote MCP server
Open your assistant’s connector settings and name the server
xpu live Archive. - Paste the connection URL
Use
https://xpu.live/mcp-server. Select Streamable HTTP and no authentication if prompted. - Discover, then ask
Inspect the available tools. Start with
list_price_providers, then search for a model or GPU.
Client support and settings vary. Opening the endpoint in a browser is not a connection test, and linking to this page alone does not install a connector.
TRY ASKING YOUR ASSISTANT
A better price question starts with context.
Find the latest archived H100 SXM rental tariffs. Include the provider, billing unit, conditions, source and observation time. Flag any stale prices.
This is an example request, not a claim that a matching offer is always available.
MCP server reference
Explore tool parameters, response fields and the rules behind each price observation.
search_latest_prices
Search public prices from each provider’s latest successful archived collection. Results are ordered by provider, tariff ID and series ID; this is not a cheapest-price ranking.
kind: optionalmodelorgpu.provider: optional exact ID fromlist_price_providers.resource: optional exact provider resource ID or catalog ID.q: optional text search across IDs, resource, catalog ID and tariff variant.rental: optional exact rental label, such ason-demand,spot,reservedorserverless; availability varies.currency: optional uppercase currency code.include_stale: boolean; defaults totrue. Set false to exclude observations beyond their freshness window.limit: 1–100; default 20.offset: 0–20,000; default 0.
{"kind":"gpu","resource":"h100-sxm","include_stale":false,"limit":10}get_latest_price
Fetch the exact tariff id returned by a search. An optional provider narrows the match. An empty result means that tariff is not in the selected latest public archive snapshot. Do not interpret it as a zero price.
list_price_providers
Discover provider IDs, public price counts, last successful collection time and latest attempt status. Optional kind accepts model or gpu. A provider without public tariffs may be shown in the unfiltered registry. New providers enter through the shared archive automatically.
get_price_freshness
Get public price and stale counts, archive read time and provider collection timestamps. Optional provider selects one provider. It does not expose administrative error logs.
Usage resource
Clients that support MCP resources can read xpu://archive/usage for the comparison and freshness rules.
What “latest” means
For each provider, the server selects the latest successful collection committed to the archive. It returns only the public observations belonging to that collection. It does not merge the last known row of every historical tariff: an offer missing from the selected successful collection stays absent. A successful empty collection returns no offers.
If a newer collection failed, the last successful collection remains the source and the latest attempt status is reported separately. Observations retain their original observed_at time and stale flag. Initial migration baselines and legacy change events are not used as current collections.
The server caches its archive snapshot in memory for up to 30 seconds. generated_at is when that snapshot was read; observed_at is when a price was observed; collection_checked_at is the provider collection time. All timestamps use UTC.
Price response fields
Each price includes id, series_id, run_id, kind, provider, resource, nullable catalog_id, variant, currency, price_type, components, conditions, source, observation and collection timestamps, stale and stale_after_seconds.
Component amounts are decimal strings, preserving stored precision. Each component has its own name and unit. Keep input, output, cache and GPU compute charges separate. Preserve tariff conditions and source links when displaying a result. Page responses include total, limit, offset and next_offset; use the latter until it is null.
Example questions
- Find the latest archived H100 SXM rental tariffs. Show the provider, billing unit, conditions, source and observation time.
- Which public price providers have fresh GPU observations?
- Find the archived tariffs matching a model ID, keeping input and output charges together for the same offer.
- Check whether a listed tariff is stale before using it in a cost estimate.
Limits and responsible comparisons
- The service is read-only. It cannot edit the site, create records, collect new prices, or access private account tariffs.
- Only the latest successful archived collections are available. Historical charts and administrative logs are outside this service.
- Rate limit: 60 requests per minute per client IP, with gateway burst controls. HTTP 429 means slow down and retry after the indicated delay.
- Requests are limited to 32 KiB. Invalid arguments are rejected. Temporary archive failures return an error rather than invented or silently substituted prices.
- Published tariffs are observations, not a guarantee of capacity, account eligibility, current availability or your final invoice.
- Never compare different units, currencies, editions or tariff conditions as if they were equivalent. Unknown or missing prices are not zero.
- No prompts or tool results are stored by the MCP service. Standard infrastructure access logs may record request metadata such as IP, path, status and time.
Read our comparison methodology and the existing pricing API documentation. For corrections, email [email protected] or report a catalog error.
Documentation reviewed 18 October 2026. Server version 1.0.0.