Skip to content
AI Compute Radar

Developers & agents

Open data, open door.

Everything the radar measures is available as JSON, and as tools an AI assistant can call directly. Read-only, no keys, CORS-open. Every response carries its own provenance — whether a number was measured or is a fixture, and when it was collected — so what you build on it can be as honest as the site. Since 31 October 2026 the pages answer in words; the API carries every number behind them, and the translation table says which field each word stands for.

Quick start

Top of the board, with the signals behind each Heat Score:

curl https://aicomputeradar.dev/api/v1/models?limit=5

Which tracked models fit an RTX 4090 at 16K context:

curl "https://aicomputeradar.dev/api/v1/fit?hardware=rtx-4090&context=16384"

Current GPU rental medians:

curl https://aicomputeradar.dev/api/v1/prices

Thirty days of the RTX 4090 at every source, as CSV for a spreadsheet:

curl "https://aicomputeradar.dev/api/v1/history?gpu=rtx-4090&days=30&format=csv"

Endpoints

PathReturnsParameters
GET /api/v1/modelsTracked models ranked by Heat Score: score, confidence, direction, flags, raw Hugging Face and OpenRouter signals, measured GGUF size, links.slug, limit (1–100)
GET /api/v1/fitFit verdicts for one machine: measured weights + computed context cache + overhead versus usable memory, with plain-language reasons and a best recommendation.hardware (required), model, context (512–1048576, default 8192), kv (f16 · q8_0 · q4_0)
GET /api/v1/hardwareCurated GPU and Mac profiles: ids for /fit, memory, usable memory after margins, bandwidth.—
GET /api/v1/pricesMedian verified on-demand rental price per GPU class (Vast.ai), USD/hour, with min, p75 and offer counts — and per class the Rent Index: this week's median against last week and against the first week collected, trend words, excluded glitch days, RunPod's lowest posted on-demand price per class as a labelled list price, Clore.ai's median for the same class (a second marketplace, never blended), and the AWS, Azure and Oracle Cloud list prices per GPU-hour, each with the instance the price sits in. Described, never forecast. Day-over-day moves of 10 % or more also as RSS at /alerts/feed.xml.—
GET /api/v1/historyDaily price history per GPU class and source, up to 365 days (the series starts 26 August 2026): the rows the Rent Index and the 30-day range are computed from. Marketplace days are medians over the day's counted hourly medians with min, p75 and offer count; list-price days carry the posted price; every row says which kind it is. JSON, or CSV with format=csv.gpu (class slug), source (key), days (1–365, default 30), format (json · csv)
GET /api/v1/pickThe pick of the week: the model a published rule selected, its snapshot frozen at selection time (heat, signals, fit ladder at 8K and 32K), the rule outcome and the written report with caveats. Null before the first issue; the archive lists every published week. Also as RSS at /pick/feed.xml.week (ISO week, e.g. 2026-w37)
GET /api/v1/openapi.jsonOpenAPI 3.1 description of all of the above — feed it to a tool-using agent.—

Use the data

The numbers are the same whoever reads them, and the licence does not ask what for. Three readers come here: a person with a graphics card who wants a fit verdict, an assistant that calls the MCP server, and an analyst who pulls the daily series into a notebook. For the third, since 27 October 2026, /history returns up to a year per class and source, as JSON or CSV, with the kind of each number named; the series starts 26 August 2026, and since 29 October 2026 every class page and the price board link the CSV behind them directly. Everything is CC BY 4.0: use it, redistribute it, build on it, and name the source. For redistribution beyond what CC BY 4.0 covers, a licence can be arranged once the project's legal notice is complete; until then there is nothing to buy here.

Reading a response

Every payload is { meta, data }. The meta block is the part that keeps you honest:

{
  "meta": {
    "source": "live",            // "fallback" = demo fixtures, a source was unreachable
    "collectedAt": "2026-…",     // oldest contributing reading; null for curated data
    "license": { "license": "CC BY 4.0", "attribution": "AI Compute Radar — https://aicomputeradar.dev" },
    "citeAs": "AI Compute Radar (2026). GPU rental prices and open-model signals, measured per source. https://aicomputeradar.dev. Retrieved 2026-10-29. Data CC BY 4.0."
  },
  "data": [ … ]
}

Field names say what a number is: downloads30d is Hugging Face's rolling 30-day counter, not an all-time total and not unique users; heat is null while a model's history is still collecting — never a low score; fit estimates carry the same margins as the site and say UNKNOWN when a model's attention shape is unpublished. Since 23 October 2026 every fit carries a basis block naming where each part of the estimate comes from: the weights are measured from the GGUF file listing, the context cache is computed from the model's config, the overhead is a declared assumption. Nothing in any response is a speed figure, a benchmark or a forecast. The full rules are on the methodology page.

MCP server

The same data as tools for AI assistants, over the Model Context Protocol (Streamable HTTP, stateless, no auth):

https://aicomputeradar.dev/api/mcp

Tools: trending_models, find_fit, gpu_prices, list_hardware, weekly_pick. Results are the JSON payloads above, provenance included, so an assistant can quote a collection time with every number.

Claude Desktop, Cursor and other Streamable-HTTP clients

{
  "mcpServers": {
    "ai-compute-radar": { "url": "https://aicomputeradar.dev/api/mcp" }
  }
}

stdio-only clients

The npm package ai-compute-radar is a dependency-free bridge to the same endpoint: it holds no data, keeps no cache and sends no telemetry.

{
  "mcpServers": {
    "ai-compute-radar": {
      "command": "npx",
      "args": ["-y", "ai-compute-radar"]
    }
  }
}

Listed in the Official MCP Registry as dev.aicomputeradar/ai-compute-radar — clients and directories that read the registry find it there automatically.

Prefer plain text? /llms.txt maps the site for language models.

Reading the API

As of today's readingsOn November 2, 2026 the API serves 49 models, 37 hardware profiles, and rental prices for 8 GPU classes from 10 sources. Everything is licensed CC BY 4.0; the attribution text is part of every response.

Everything on the site is available as JSON through a read-only API without keys, and as tools an AI assistant can call through the MCP server. The endpoints return the model board with the signals behind every Heat Score, the fit verdicts of every model against a chosen machine at a chosen context, the hardware profiles, the GPU rental prices per class and source, and the current pick of the week. Every response carries its own provenance: whether a number was measured or is a fixture, and when it was collected.

The API returns what the pages show and nothing the pages do not. There are no benchmark results, no speed figures and no forecasts in any field, because the site publishes none; a consumer that needs those must measure them elsewhere. Where a reading is missing the field is null, never a guess, and where a median stands on few offers the offer count travels with it so a downstream chart can mark it as the site does.

The data is licensed for reuse with attribution, and the attribution text is part of every response's envelope. The limits are the ones a shared public endpoint needs: a modest rate per client and cached responses that refresh on the same cadence as the pages. For a bulk history of hourly rental readings, which the API does not expose, the methodology page says what the site is waiting for before it publishes one.

Terms

Building something with it? We would like to see it — the contact address is in the legal notice.