Glossary

Terms & definitions

Plain-language definitions for the metrics and categories used across the directory.

Inference

Running a trained model to produce output (tokens) from a prompt. Inference providers sell access to this compute, usually priced per token or per GPU-second.

Token

The unit a model reads and writes — roughly 0.75 words in English. Pricing and throughput are quoted per token or per million tokens (1M).

Blended $/1M

A single comparison price: the mean of input and output cost per million tokens for a provider's reference model. A rough anchor, not a bill estimate.

Input vs. output price

Most providers charge separately for prompt (input) tokens and generated (output) tokens; output is usually more expensive.

TTFT (time-to-first-token)

Milliseconds from sending a request to receiving the first output token — the felt "responsiveness." Lower is better.

Throughput (tok/s)

Sustained output tokens generated per second once streaming begins. Higher means faster completions. Independent of TTFT.

Uptime %

Share of time the API is available. Reference figures here are indicative, not contractual SLA numbers.

OpenAI-compatible API

The provider implements the OpenAI Chat Completions request/response shape, so code written for the OpenAI SDK works by changing the base URL and key.

Open weights

Models whose parameters are publicly downloadable (e.g. Llama, Mistral, Qwen). Many providers host these; some also let you self-host.

Proprietary model

A closed model available only through its owner's API (e.g. GPT, Claude, Gemini). No self-hosting path.

Reputation index

Tokenmeter's editorial 0–100 composite of public signals about a provider. Not a first-party metric and not audited — see Methodology.

Confidence level

How much to trust a figure: high (confirmed against public docs on the review date), medium (public but not re-confirmed), or seed (illustrative, pending verification).

Frontier lab

A company training its own frontier models and serving them through a first-party API (e.g. OpenAI, Anthropic, Mistral).

Hyperscaler

A major cloud platform offering inference alongside the rest of its cloud stack, with broad compliance and global regions.

Aggregator

A platform hosting many open models behind one API, competing on catalog breadth and per-token price.

Router

A service that forwards each request to an upstream provider, giving one API and one bill across many providers and models.

Specialist hardware

Providers running custom silicon (LPUs, wafer-scale engines, RDUs) to push throughput and latency well beyond commodity GPUs.

GPU cloud

Infrastructure-first platforms offering dedicated GPUs and serverless runtimes; you often deploy or manage the serving layer yourself.

Specialist app

An API tuned for a specific job (e.g. search-grounded answers) rather than general-purpose token serving.

Categories in the current dataset: Aggregator · Frontier lab · GPU cloud · Hyperscaler · Router · Specialist app · Specialist hardware · Subscription (flat-rate).