NanoGPT

Pay as you go has no monthly fee, and the plan card lists one free model available on the web only. Everything else draws on a prepaid balance. The getting-started page says the API can be used without an account, and web conversation history is kept in the browser by default rather than on NanoGPT servers.

Deposits start at $1 by card through Stripe or $0.10 in Nano cryptocurrency, and other cryptocurrencies are accepted. Because a balance is prepaid, a small top-up caps what a trial can cost; there is no card on file that renews automatically unless you subscribe.

Pro covers most open-source text models, such as GLM, DeepSeek, Kimi, MiniMax and Xiaomi MiMo, up to 60 million input tokens a week across all of them, plus 100 included images a day from a short list of image models. It works on the web and through the API. Models from OpenAI, Anthropic and Google are not in the allowance and are charged from the balance, with a 5% discount on eligible paid text models.

Not every token counts the same. GLM 5, 5.1 and 5.3, Kimi K2.7 Code and DeepSeek V4 Pro count double; other included models count once. TEE variants and some provider-specific routes are excluded, and the subscription page itself advises against subscribing for the image models, which may not stay included.

• Weekly cap: 60 million input tokens, cached input included, resetting Monday 00:00 UTC.
• Daily cap: 100 included images.
• Choosing a specific provider on an included model moves that request to pay-as-you-go pricing.

The pricing page says text models are billed at the provider’s list price, with GPT-5.5 at $5 per million input tokens and $30 per million output tokens, the same as at OpenAI. Each API response returns the exact cost. The only opt-in surcharges are 5% for pinning a provider and a fee of 5% of normal cost for bring-your-own-key requests.

Editorial assessment

If most of your use is a few prompts a day on GPT, Claude or Gemini, a small balance is likely cheaper, because Pro does not cover those models. If you run long open-source sessions or coding agents every day, the weekly allowance can outlast $12 of list-price tokens. Check the Usage page after a normal week before deciding.

The answer depends on the plan. The terms say subscriptions are for personal, non-commercial use only, including no reselling, building commercial products or serving multiple end users, and each subscription is for one person. Pay as you go explicitly permits commercial use, including resale.

That permission does not override the model provider. The terms list the usage policies of OpenAI, Anthropic, Google and other providers and make you responsible for checking that each model allows your integration and downstream use. You own the output under NanoGPT’s terms, subject to those model licences.

Prompts are forwarded to the provider of the model you choose. NanoGPT says it does not train on your input or output and does not sell content, and that it picks the strictest retention settings each provider offers. Final retention and logging still depend on the provider; the terms list, for example, a broker route without guaranteed zero data retention.

Uploaded files attached to a request may be copied to NanoGPT-managed storage, and the terms say no single maximum retention period applies to those copies. Memory is off by default and stays in the browser until sync is enabled. Keep confidential documents out of the chat unless you have chosen a route with a documented zero-retention commitment.

The terms require users to be at least 18 and exclude people in comprehensively sanctioned places, including Russia, Belarus, Iran, North Korea, Cuba, Syria and occupied regions of Ukraine. They also prohibit using another person’s likeness or voice without consent and sexual content involving identifiable people without their consent.

If a payment is disputed or charged back, NanoGPT may suspend the whole account, including an active subscription and any prepaid balance, while the dispute is open or afterwards. The refund policy asks you to contact support first; it describes good-faith refund requests as generally approved, with exceptions for heavy completed usage.

1. Record the starting balance and the date of the Monday reset.
2. Use the same three models each day for the tasks you normally do.
3. At the end of the week, export or copy the per-model token counts and charges from the Usage page.
4. Apply the 1x and 2x weights to the open-source input tokens and compare the total with 60 million.
5. Add the GPT or Claude charges, which Pro would not cover, less the 5% discount.

Accept the comparison when every request has a logged model and cost. Reject weeks with missing usage records or unexplained charges. No spending or allowance result is reported until the week is run.

For a monthly multi-model subscription with one credit balance, see YesChat. For a single provider’s full product, compare ChatGPT and Claude.