I've seen claims on OpenAI's website that both GPT-4.1 and GPT-4.1-mini have a context length of 1 million tokens. Has anyone actually verified this? I'm curious if this applies to both the API and the ChatGPT subscription service, as I hear there are limitations with the context window on the web app.
5 Answers
From my findings, the API boasts a 1 million token context, while the web UI only provides 32,000 tokens for Plus subscribers and 128,000 for Pro users. So, unfortunately, if you're relying solely on the web app, you're missing out!
Yes! I've tested the API, and it works up to 600k tokens, although the output quality drops as you approach that limit. I usually keep the context summarized to avoid that decline in performance.
What kind of context length are you seeing a performance drop at?
Do you think this is also true for the Plus subscription on their web app? I feel like there might be a limit there.
In my experience, the longer context can help with complex queries, but there are concerns about performance drop-off after around 64k tokens, similar to what happens with other models.
Yeah, big contexts might just hinder quality when it gets too lengthy.
Is that what happens with all models, despite the claims of higher token limits?
Absolutely, it works seamlessly with the API. In fact, some tools like Cline show how much of the 1 million tokens you're using, but it strictly uses the API.
Got it! But Cline operates via the API, right?
I've found that it can be tricky to utilize that much context effectively. If you're looking for details dispersed across a large dataset, breaking it up into sections usually yields better results than relying on a single massive context.
Makes sense! I guess maximizing the token count doesn't always translate to better accuracy.
Thanks for clarifying! That's a bummer for web users.