sub2tok

sub2tok is a third-party model gateway: it does not train models, and it is not an official channel either. It simply converts three kinds of upstream channels — Claude, Codex and Antigravity — into an API compatible with the OpenAI, Anthropic and Gemini protocols, billed per token. A lineup of 22 models is not broad in coverage, yet unit prices are converted from official reference prices by a multiplier.

Core capabilities

• Three-protocol compatibility: the same model is reachable through the Anthropic Messages, OpenAI Chat Completions and Responses endpoints, so changing clients does not mean changing the integration
• One-step setup: install.sh automatically writes the configuration for Claude Code, Codex CLI, the Claude desktop app and the magpie plugin, and any OpenAI/Anthropic client works too
• Itemized billing: input, output, cache writes and cache reads are metered separately, and the balance is kept in nano-USD so every charge can be recalculated line by line; models outside the price list are refused outright instead of being charged at a fallback price
• Pure prepaid: no monthly fee and no subscription tiers; top-ups by Alipay, USDT and USDC; a $1 signup credit
• Supply from shared subscriptions: upstream capacity comes from the operator's own channels plus subscriptions that users share, and holders can list a share by time slot in exchange for a revenue cut (approved users only)

Use cases Suited to developers running long tasks with Claude Code or Codex CLI, and to small teams that want to hold the API budget down while accepting fluctuation in supply; it is not suited to enterprises that need domestic Chinese models, invoices or data residency.

Unique advantages OpenRouter and AIHubMix resell at the official list price and cannot quote this rate; sub2tok draws supply from its own channels and from consumer subscriptions shared by users, with the whole Claude line at 0.2x and the gpt-6 series at 0.25x: claude-opus-5-5 costs $0.80 input and $4 output per 1M tokens.

Editorial assessment If the goal is to push Claude and GPT series unit prices down to 0.2x to 0.25x of official reference prices, this is the least troublesome one to configure among gateways of this subscription-to-API kind. The price is that supply depends on shared subscriptions, so a model may be "temporarily without a taker", and there is no SLA and no refund on the balance. Start with a small trial run and move production workloads back to the official API or to a platform with enterprise terms.