EndlessTokens is previewing an LLM API that charges for elapsed time instead of tokens. Its advertised launch prices are $15 for one day or $36 for three days, with no token or request caps during the pass. Both options include two concurrent sessions per key and a 128K context window. The service is waitlist-only, not generally available. Its public landing page directs prospective customers to register for notification when passes open.
The proposal targets a familiar agent-workflow problem: repeated model calls can make a task’s final API bill difficult to predict. EndlessTokens would fix the access price upfront, while leaving the amount of useful work dependent on model performance, concurrency and available capacity.
This analysis draws on the company’s public preview pages, which do not state an explicit announcement date. The supplied evidence contains no independent performance measurements, and the advertised terms may change before launch.
What the Time Pass Buys
EndlessTokens describes its offer as “Unlimited tokens for the length of your pass.” The important qualification is the purchased time window: the service removes token and request quotas, but retains limits on simultaneous work and individual requests. According to the pricing page, a standard pass starts at payment and has no subscription or automatic renewal.
| Advertised launch option | Price in US dollars | Effective daily price | Included access |
|---|---|---|---|
| One-day pass | $15 | $15 | Two parallel sessions per key, 128K context |
| Three-day pass | $36 | $12 | Two parallel sessions per key, 128K context |
Prices exclude applicable taxes. Using the listed dollar prices, the three-day option costs 20% less per day than the one-day option. That saving only helps if the additional days are useful; buying three days for a short task still costs more upfront.
The company also lists three-hour and six-hour top-ups at $4.49 and $7.99, available only with an active pass. Unlike the standard passes, these extensions are advertised as starting on the first request, with automatic activation after seven days if unused. That difference matters when planning when to buy access. The landing page also advertises founding-member launch pricing and a full refund if the launch target is missed.
I’d prepare the client configuration and task queue before paying for a standard pass. A clock that starts at payment can spend part of the purchased window on setup or inactivity.
Unlimited Tokens Still Come With Technical Limits
The API documentation lists the operational boundaries:
- Two parallel sessions per key, with one key per account described on the landing page.
- 128K context, meaning roughly 128,000 tokens of context capacity, not unlimited conversation memory.
- Up to five images per request.
- Reasoning effort up to medium.
- No token or request cap within the active pass.
Context and total usage measure different things. An agent could make many requests over a day without encountering a daily token quota, yet still need to keep each interaction within the model’s context window. Long-running jobs would still need strategies such as summarizing earlier steps or retrieving only relevant material.
The initial model listed is Qwen3.8-27B. EndlessTokens says it hosts and scales the open-weight model itself, with vision, reasoning and tool calling at 128K context. This is provider-hosted inference: customers would call its API, not run the model on their own machines. The company says it does not resell another inference provider’s capacity. Those are vendor descriptions of its deployment, not independently verified infrastructure findings. Source: EndlessTokens.
The included et:auto identifier adds another distinction. According to the quickstart, it chooses a model for each conversation, directing simpler jobs toward a faster model and harder work toward a stronger one. The homepage currently names one initial model, so the preview should not be read as evidence that a broad routing catalog is already operating.
For workflows that require a specific model, I’d use an explicit model identifier once available. Automatic routing may be convenient, but model consistency and routing convenience serve different needs.
Existing Clients Get Familiar API Formats
EndlessTokens advertises support for OpenAI-style Chat Completions, Responses, Completions and Embeddings endpoints, plus Anthropic Messages. Its API page gives different base URLs for the two interfaces:
Image: Flat-rate LLM API (OpenAI & Anthropic) | endlesstokens.
- OpenAI-compatible:
https://api.endlesstokens.ai/v1 - Anthropic-compatible:
https://api.endlesstokens.ai
Compatibility here concerns request and response interfaces. It does not mean a pass includes OpenAI’s or Anthropic’s proprietary models, or that a hosted open-weight model will behave identically to them. The landing page separately describes bringing a customer’s own provider key, with the customer paying that provider directly.
The following example adapts the company’s documented request format. It illustrates a request after access opens and a valid key is issued; it is not a report of a successful test.
curl https://api.endlesstokens.ai/v1/chat/completions \
-H "Authorization: Bearer <your-key>" \
-H "Content-Type: application/json" \
-d '{
"model": "et:auto",
"messages": [
{"role": "user", "content": "Summarize this task plan."}
]
}'
The command submits one user message to the OpenAI-compatible Chat Completions endpoint, using automatic model routing. The documentation says GET /v1/models lists exact, case-sensitive model identifiers. It also says responses expose the selected model through the x-et-routed-model header, which could help developers record what handled a routed request. Source: quickstart.
Endpoint compatibility can reduce integration work for existing clients. It leaves model suitability as a separate question, particularly for agents that depend on reliable tool calls or precise instruction following.
Capacity Errors Are Part of the Billing Story
EndlessTokens explicitly makes no speed promise. Its documentation also describes what happens when the service has no free GPU capacity, making error handling central to evaluating the offer.
| Status | Documented meaning | Practical response |
|---|---|---|
401 | Wrong or revoked key | Check credentials |
402 | No active pass | Check the account’s access window |
429 | More than two parallel sessions | Reduce concurrent work |
Frequently Asked Questions
3 questions
1Can I buy an EndlessTokens pass now?
The public pages say the service is waitlist-only. Users can register to be notified when passes open.
