Feb 15, 2026
We're making some changes to the NanoGPT subscription. We want to be fully transparent about why, and what this means for you.
A few things have come together that make the current unlimited setup unsustainable:
Abuse. We've been dealing with a growing number of accounts that exploit the subscription — multiple accounts depositing minutes to each other, maxing out input tokens around the clock on the most expensive models. These are often the same users that then do chargebacks, which compounds the problem.
Extreme usage concentration. The top 1-5% of users account for over half our total token usage, and well over half the total cost.
Model costs have gone up. The subscription used to be mostly cheaper model usage (various Deepseek variants). The shift to GLM 4.7, Kimi K2.5, and now GLM 5 has been great for output quality but not great for costs. There was plenty of spare capacity for Deepseek, so good deals were available. There is essentially zero spare capacity for K2.5 and GLM 5 at any provider, meaning almost no volume discounts. These models are more expensive at list price, and the lack of discounts means per-token costs have multiplied several times over.
Growth outpacing capacity. The number of subscribers is growing faster than we can increase our rate limits with providers. This means worse performance for most users (slower responses, more 429 errors) and us falling back to more expensive backup providers.
Starting Tuesday, February 17th at noon CET, the following limits take effect:
A maximum of 10 concurrent requests (already in place).
A new burst bucket of 10 requests per 10 seconds, in addition to the existing 60 requests per minute limit.
This is the biggest change. Input tokens used to be unlimited, which meant a very small group of users were consuming billions of tokens per month. We're introducing a cap of 60 million input tokens per week.
Based on data from the last month, this will affect roughly 5% of users (and this 5% includes accounts that are actually violating our Terms of Service). The average and median user will very likely not notice this at all — though of course your mileage may vary.
A limit of 100 free images per day in the subscription. This will impact virtually no one, except a few accounts that appear to be using us as an image generation backend for another service — you'd be hard pressed to manually generate images 24/7 at the rates we're seeing from some accounts.
If you are a legitimate user who is impacted by these changes, we genuinely apologize. We'd love to cater to every usage level, but it's currently just not possible to do so without the subscription becoming deeply unsustainable.
To be clear — aside from those clearly breaking our Terms of Service, we absolutely don't blame anyone for getting the most out of the subscription. We'd love to keep things unlimited because we know many of you are very happy with it. But with the way things are going now, we'd be subsidizing a very small group for a fairly large sum.
If you'd like to cancel your subscription, email us at support@nano-gpt.com or open a ticket in our Discord with your support key and we will refund your subscription no questions asked.
How about a more expensive subscription tier?
We've considered it. The issue is that a more expensive tier would need to offer higher limits (obviously), and since the current $8 tier already isn't profitable when people use it to the limit, a $20 tier would just mean high-usage users self-select into the bigger plan and exacerbate the problem.
How about weighting different models differently?
This is a good idea and something we may move toward. For now, we needed a simple, easy-to-understand change that we can build on.
Can you guarantee there won't be more changes?
Honestly, no. We wish we could say yes, but the reality is that the subscription only works for us if it's not too loss-making. We're hoping these changes accomplish that, but we don't have a crystal ball.
We're also hoping we can make more targeted changes later — different model weighting, for example. But we needed to start with something straightforward that we can iterate from. The subscription started out mostly for roleplay, but the hype around K2.5, GLM 5, and agentic coding more broadly is changing our average user profile and increasing costs significantly.
Thank you for understanding, and thank you for using NanoGPT.