Claude as a Janitor AI Proxy: What It Costs per Message, the Two Routes, and Anthropic's Own Caveat

Claude is the model Janitor AI users ask about most after DeepSeek, and it is the one with no free route at all. Anthropic sells by the token, OpenRouter lists no :free Claude, and the "free Claude proxies" that circulate in Discord are stolen keys with a stranger reading the chat. So this guide does something the free-backend guides do not: it starts with the bill. At listed prices Claude is cheaper per message than most people assume, and the arithmetic is the first thing to see.

Two routes then follow, with exact values. And a caveat from Anthropic itself about the more direct of the two, quoted rather than paraphrased, because it decides which route to pick.

What a message costs, at listed prices

Every request Janitor sends carries the whole context: the character definition, your persona, the chat memory, the advanced prompt and the recent history. You pay for all of it as input tokens, every message, plus the reply as output tokens. On 24 August 2026 OpenRouter's public model list priced the current Claude models like this:

The arithmetic for a heavy roleplay request of 20,000 input tokens and a 300-token reply: on Sonnet 5, 20,000 tokens at $2 per million is $0.04, and 300 output tokens at $10 per million is $0.003, so about four cents a message. On Opus 5 the same message is about ten cents. A hundred messages an evening on Sonnet 5 is roughly $4; the same evening on a :free OpenRouter model is $0, and on DeepSeek's own API it is pennies. Claude is the expensive seat in the room, and whether its writing is worth it is a taste question this page does not settle.

Context length is the lever. Trim the chat memory and the advanced prompt and a message that cost four cents costs two.

Route one: OpenRouter

The route built for this. One OpenRouter account, a balance loaded, and the Claude models sit behind the same URL as everything else.

OpenRouter deducts the listed price from your balance per request and returns a 402 when the balance runs out, or goes negative, which its rate-limit documentation says blocks even free models until topped up. There is no daily cap on paid models the way there is on :free ones. If you already use OpenRouter for free models, this is a model-id change and a balance, nothing else; the general setup is in our OpenRouter guide.

Route two: Anthropic's own compatibility layer

Anthropic's API is not natively OpenAI-shaped, which is why Janitor could not talk to it for a long time. It now can, through a compatibility layer that Anthropic documents: point an OpenAI client at the base URL https://api.anthropic.com/v1/, use a Claude API key, and use Claude model names. In Janitor's terms:

Read Anthropic's own note before choosing this route. The documentation says the compatibility layer is primarily intended to test and compare model capabilities, and is not considered a long-term or production-ready solution for most use cases, while adding that it is intended to remain fully functional and not have breaking changes. The documented differences matter for roleplay: all system and developer messages are concatenated into a single system message at the start, temperature above 1 is capped at 1, several OpenAI fields are silently ignored, and rate limits follow Anthropic's standard limits for its messages endpoint, which are per organisation. For one person's roleplay it works; for anything you want to keep working without surprises, route one is the supported path.

What a proxy of any kind does not change

The usage policy. Anthropic's usage policy governs what Claude will write, and it applies to the model wherever the request came from. Janitor is known for uncensored roleplay; Claude is not an uncensored model, and a refused scene is Claude declining, not a broken setting. Changing the route, the URL or your IP address does not change the policy.

The rate limit. Through OpenRouter, the limit is your OpenRouter account's; through Anthropic, it is your organisation's tier. Both follow the key. Our 429 page covers the general rule.

The bill. Prices are per token and per model; no route makes Claude free, and a "free Claude proxy" is somebody else's stolen key with somebody else reading the chat. Chub's documentation puts it directly: such proxies are created by unauthorized third parties who gain access to your data and IP address and use stolen API keys. Do not.

Where the network kind of proxy fits

One place. Anthropic publishes the countries where Claude and its API are available, roughly 175 at the time of writing, and excludes places such as mainland China, Hong Kong, Macau, Russia and Iran; some school and office networks filter api.anthropic.com or openrouter.ai. If the endpoint will not answer from where you are, no key or model fixes it, and a static residential or ISP address in a supported country, kept for the session, is the tool, at the cost of reading against Anthropic's terms first. Our proxies for Claude AI post covers that side. For everything else on this page, cost, limits, refusals and keys, our product does nothing, and the honest recommendation is to load $5 on OpenRouter, pick anthropic/claude-sonnet-5, and watch the balance.

Sources

• OpenRouter, public models list: Claude model ids and per-token prices, read 24 August 2026; API rate limits: 402 on a negative balance.
• Anthropic, OpenAI SDK compatibility: the base URL, the testing-and-comparison caveat, system-message hoisting, the temperature cap, ignored fields and rate limits.
• Anthropic, supported countries and usage policy.
• Chub documentation, API connections: the warning about third-party reverse proxies.
• Janitor AI Help Center, The Absolute Beginner's Guide to Using a Proxy with Janitor, Part One: the field names and menu path.