Intelligence,
on tap.
No token cap. Pay for time, not tokens.
Most APIs sell intelligence by the sip. We sell the tap. Open it for a day or three, and let your systems think for as long as they need to.
OPENAI_BASE_URL=https://api.endlesstokens.ai/v1 OPENAI_API_KEY=<your-key>ANTHROPIC_BASE_URL=https://api.endlesstokens.ai ANTHROPIC_AUTH_TOKEN=<your-key>curl https://api.endlesstokens.ai/v1/chat/completions \ -H "Authorization: Bearer <your-key>" \ -H "Content-Type: application/json" \ -d '{"model": "<model-name>",\ "messages": [{"role": "user", "content": "hello"}]}'
You don’t pay per movie. Why pay per token?
Streaming changed how people watch. Nobody pays per episode anymore: you subscribe, and the whole catalogue is yours for the period. Intelligence is still sold the old way, one unit at a time.
Today every thought an AI has comes with a meter. The people building the future are rationing it: trimming prompts, capping agents, and watching a quota drain while their system is mid-thought. A machine that runs all night should not wake up to a surprise bill, or to a wall that says “limit reached.”
The next wave will not be a person typing at a keyboard. It will be agents that never clock off, assistants that never sleep, and machines that work while nobody is watching, in software and, in time, well beyond it. They need intelligence that is built to run unattended and predictable to the day.
So we changed the unit. Not tokens. Time. A pass works like a subscription: a fixed price for a fixed period, and you use it as much as you like within fair limits. There is no token meter and no request cap, so run it flat out, run it all week, and the bill does not move.
We are building the subscription for the age of autonomous software. Welcome to unmetered intelligence.
Fair limits still apply: two parallel sessions per key and a 128K context window. See the limits.
Be first in
Join the waitlist at the portal. You sign in with email, GitHub or Google, and we tell you the moment passes open. Founding members get the launch price and a full refund if the launch target is missed.
Join the waitlistThe problem
Metered APIs
Every agent loop costs tokens. Bills surprise you.
Cheap flat plans
Request caps and throttling in the middle of a task.
endlesstokens
A fixed price and no ceiling. Not unlimited concurrency, not a speed promise.
How it works
Sign in and buy a pass
1 day or 3 days. Starts at payment. No subscription, no auto-renew.
Create your key
One key per account, two parallel sessions.
Set a base URL and key
In any client that lets you set both. No plugin, no lock-in.
No token or request cap
Pay for the time window.
2 parallel sessions
Per key.
Works with any client
Chat Completions, Responses, Anthropic Messages.
Images supported
Up to 5 per request.
Pricing
1-day pass
Starts at payment.
3-day pass default
20% cheaper per day.
Top-ups (3 h, 6 h) only with an active pass. Prices exclude tax; the India total with GST is always shown. No subscription, no auto-renew.
Launch prices and founding-member terms apply until launch. Passes are bought in the portal.
Works with the tools you already use
endlesstokens is an API, not a plugin. Any coding agent, editor extension, harness, SDK or script that speaks Chat Completions, Responses or Anthropic Messages works. These are a few, in alphabetical order, each with its own setup guide.
Aider
Terminal pair programmer · setup guide ↗
Claude Code
Terminal coding agent · setup guide ↗
Cline
Open-source coding agent · setup guide ↗
Codex
Coding agent CLI · setup guide ↗
Kilo
Open-source coding agent · setup guide ↗
OpenCode
Terminal coding agent · setup guide ↗
Roo
Coding agent for your editor · setup guide ↗
Your own client
Any tool or script that can set a base URL and key · setup guide ↗
Models
Qwen3.8-27B
Open weights, hosted and scaled by us. Served at 128K context, with vision, reasoning and tool calling. More popular open-source models follow as demand grows. Model details
Hosted by us. Not resold.
Every open-weight model on endlesstokens runs on GPUs we host and scale ourselves. We do not proxy your requests to another inference provider, and we do not resell anyone else’s capacity.
We host and scale it
We deploy each model, run it, and add capacity as demand grows. What you call is what we operate.
We measure it first
Every model is benchmarked and its unit measures are derived before it goes on offer. Then we keep launching the next one.
We do it in the open
Our method is built with Token By Token, an independent inference lab, and we welcome more teams who want to benchmark and engineer with us.
We collaborate openly with Token By Token on how open models are served and measured. Want to benchmark or build with us? Get in touch.
Researcher: Durgasuresh Kagitha
Frontier models stay with their labs
Closed, proprietary models belong to the labs that build them. We will not proxy or resell them. Once we agree terms with a lab, we will offer its models through that lab, with its approval.
Trust
We never store prompts
Only metadata.
GST invoice
With your GSTIN.
Limits stated up front
Sessions, context and capacity, in plain words.
Hosted, not resold
Open-weight models we run ourselves.
FAQ
Will it cut me off mid-task?
There is no token or request cap. If every GPU is full, a new session can be briefly refused and the rejected time is credited.
How many sessions do I get?
Two concurrent sessions per key. A third evicts the least recently used idle one.
Do you store my prompts?
No. We keep request metadata only.
What are the limits?
No token cap and no request cap. Two parallel sessions per key, a 128K context window, up to 5 images per request, and reasoning effort capped at medium. If every GPU is full, a new session can be briefly refused and the rejected time is credited.
Which countries?
India and the US at launch. Others can register and are told when their country opens.
Need more than two sessions?
Each account gets two parallel sessions. For a team, use one account per person, or write to [email protected] and tell us what you need.