Prices

What each vendor's rows leave out

akamai, read 2026-11-11

Rows are the public Akamai Cloud (Linode) types API (https://api.linode.com/v4/linode/types), read 2026-11-11; prices are the USD hourly list prices for the default region. Nanode and Standard plans are shared vCPU; Dedicated and Premium plans are dedicated vCPU. Each row prices one whole instance per hour; `covers: instance` and `cpu.count` divide it to a per-vCPU figure. Monthly list prices are published by the API but not stored as rows.

akash, read 2026-11-11

Rows are Akash marketplace floors read from console-api.akash.network/v1/gpu-prices. Each row is a GPU-hour price, normalised through the GPU registry: `gpu.type` is the console API's model, ram and interface joined by spaces, and the row's `note` keeps the listing's availability. Fetched by scripts/providers/refresh.mjs from akash.

atlas-cloud, read 2026-11-11

Seeded 2026-11-11 from https://api.atlascloud.ai/v1/models with the scripts/providers/sources/atlas-cloud.mjs adapter. Atlas Cloud's public /v1/models feed; default condition and conditional bands as separate rows. Fetched by scripts/providers/refresh.mjs from atlas-cloud.

aws, read 2026-11-11

GPU rows are fetched by scripts/providers/refresh.mjs. CPU rows were seeded by hand on 2026-11-11 from AWS's own on-demand metered unit map for US East (N. Virginia) Linux (the JSON behind the EC2 pricing page): one row per general-purpose instance family at its smallest size of at least 2 vCPUs. No AWS product here is sandbox-shaped: every row prices an EC2 instance (`product: instance`), and `QUOTED_PRODUCTS` names no AWS product, so none of these rows is quoted against the sandbox shape. Fetched by scripts/providers/refresh.mjs from aws-ec2.

azure, read 2026-11-11

Rows are Azure Retail Prices API rates in eastus: the ND96isr H100 v5 GPU node and the Linux Dsv5 general-purpose CPU family (D2s through D96s). The GPU VM is a node of eight H100 NVL 94GB GPUs: `covers: node` and `gpu.count: 8` divide the node-hour figure down to one GPU-hour. CPU rows price a whole VM per hour; `covers: instance` and `cpu.count` divide it to a per-vCPU figure. Spot, Low Priority and Windows meters are not stored. Fetched by scripts/providers/refresh.mjs from azure-retail.

beam, read 2026-11-11

Rows are Beam on-demand per-machine prices from beam.cloud/pricing; each machine carries one GPU. `gpu.type` is the machine name; `covers: instance` with `gpu.count: 1` leaves the machine-hour figure as one GPU-hour.

chutes, read 2026-11-11

Seeded 2026-11-11 from https://llm.chutes.ai/v1/models with the scripts/providers/sources/chutes.mjs adapter. Chutes' public llm.chutes.ai/v1/models feed; figures are USD per 1M tokens per Chutes' pricing page. Fetched by scripts/providers/refresh.mjs from chutes.

Cloudflare, read 2026-11-11

Sandbox SDK pricing is the Containers pricing it runs on (developers.cloudflare.com/sandbox/platform/pricing/). Billed per 10 ms while running, plus the $5/month Workers Paid plan, Workers and Durable Objects usage; those are not modelled. A shape is priced on the smallest instance type that fits it.

coreweave, read 2026-11-11

Rows are CoreWeave node-hour prices from coreweave.com/pricing; `gpu.count` is the GPUs per node the page lists. `gpu.type` is the node's GPU name; `covers: node` divides the node-hour figure down to one GPU-hour.

Daytona, read 2026-11-11

All billing is calculated per second. Rates as listed; $200 of free compute and volume discounts are not modelled.

deepinfra, read 2026-11-11

Seeded 2026-11-11 from https://api.deepinfra.com/v1/openai/models with the scripts/providers/sources/deepinfra.mjs adapter. DeepInfra's public /v1/openai/models feed; per-1M-token list prices with the feed's promotional discount applied. Fetched by scripts/providers/refresh.mjs from deepinfra.

digitalocean, read 2026-11-11

Rows are GPU Droplet per-GPU-hour prices from digitalocean.com/pricing/gpu-droplets, effective 2026-08-01. Each row is a GPU-hour price; `gpu.type` is the GPU Droplet's GPU name and `gpu.tier` its billing tier. CPU rows are DigitalOcean's own Droplet price list (Basic and General Purpose), read 2026-11-11; hourly rates for Linux Droplets, billed per second with a 60-second minimum.

E2B, read 2026-11-11

Running sandboxes are billed per second for provisioned CPU and RAM; disk is included with every plan. Plan fees (Pro $150/month, concurrency add-ons) are not usage rates and have no row.

Fly.io, read 2026-11-11

Fly Sprites. sprites.dev/pricing redirects to fly.io/pricing, where these rates are published. Compute prices are for Ashburn (iad); other regions vary.

Modal, read 2026-11-11

Sandboxes are billed at non-preemptible rates of 3x list price; the sandbox rows are those rates as listed. Region selection multiplies base prices by 1.15-1.75x; not modelled.

namespace, read 2026-11-11

Rows are Namespace's own compute shape table (https://namespace.so/pricing, also served as /pricing.md), read 2026-11-11. A compute unit is 1 vCPU and 2 GB for one minute times the platform multiplier; the two variants are the plan's prepaid rate and the overage rate for minutes beyond the plan. Billing is per minute with a one-minute minimum; the next 15 seconds are rounded down. There is no monthly or per-second list price. Devboxes are the sandbox-shaped product: S/M/L/XL priced per minute by shape, not per vCPU; `cpu.count` is the shape's boost ceiling.

Northflank, read 2026-11-11

Usage rates on Northflank Cloud and BYOC; compute plans (nf-compute-*) are bundles of these rates and have no row.

novita, read 2026-11-11

Seeded 2026-11-11 from https://api.novita.ai/v3/openai/models with the scripts/providers/sources/novita.mjs adapter. Novita's public /v3/openai/models feed; `price_per_m_decimal` only, USD per 1M tokens. Fetched by scripts/providers/refresh.mjs from novita.

openrouter, read 2026-11-11

Seeded 2026-11-11 from https://openrouter.ai/api/v1/models with the scripts/providers/sources/openrouter.mjs adapter. OpenRouter's public /api/v1/models feed; every row is OpenRouter's own routed price, never relabelled to the upstream lab. Time-of-day price bands (`utc_start`/`utc_end` overrides with no `min_prompt_tokens`) are not yet in this seed: the adapter keys each band by its UTC window (`:utc-<utc_start>-<utc_end>`), so the next provider refresh lands the 12 time-of-day rows without duplicate SKUs. Fetched by scripts/providers/refresh.mjs from openrouter.

oracle, read 2026-11-11

Rows are OCI pay-as-you-go GPU per-hour prices read from the Oracle Cloud price list. Each row is a GPU-hour price; `gpu.type` is the price list's own display name for the part. Fetched by scripts/providers/refresh.mjs from oracle.

RunPod, read 2026-11-11

RunPod's GPU catalog: on-demand pod prices per GPU-hour. A GPU is listed in a cloud tier (secure or community) only where RunPod offers it there; serverless prices appear with the pricing type "serverless". Fetched by scripts/providers/refresh.mjs from runpod.

sambanova, read 2026-11-11

Seeded 2026-11-11 from https://api.sambanova.ai/v1/models with the scripts/providers/sources/sambanova.mjs adapter. SambaNova's public /v1/models feed; per-token strings converted to USD per 1M tokens. Fetched by scripts/providers/refresh.mjs from sambanova.

verda, read 2026-11-11

Rows are Verda instance-hour prices read from api.datacrunch.io/v1/instance-types. A GPU row prices an instance of `gpu.count` GPUs and `covers: instance` divides it down to one GPU-hour; `gpu.type` is the instance type's GPU model. CPU-only instance types stay `other` rows. Fetched by scripts/providers/refresh.mjs from verda.

Vercel, read 2026-11-11

Vercel Sandbox Pro rates in the default iad1 region; Active CPU, Provisioned Memory, data transfer and drive rates vary by region. Each vCPU includes 2 GB of memory; vCPUs are 1 or an even number. Pro allows up to 8 vCPUs and 16GB per sandbox; memory follows the vCPU count, so only the vCPU cap is applied to a shape. Quotes bill Vercel memory as Vercel does: 2 GB per vCPU (memory_per_vcpu), so a 2 vCPU sandbox is 4 GB whatever memory the shape asks for. The page's example: "an idle sandbox with 2 vCPUs ... Its 4 GB of provisioned memory counts as 4 GB-hours" (https://vercel.com/docs/sandbox/pricing, read 2026-11-11).

vultr, read 2026-11-11

Rows are Vultr's own Cloud Compute and Optimized Cloud Compute hourly prices, read 2026-11-11. Cloud Compute sizes are shared vCPU; Optimized Cloud Compute General Purpose sizes are dedicated. Monthly prices are published as a 672-hour cap and are not stored as rows. Each row prices one whole instance per hour; `covers: instance` and `cpu.count` divide it to a per-vCPU figure.