Claude API raises the request cap from 32 MB to 128 MB
Anthropic on October 29 raised the maximum request size for the Messages and Token Counting endpoints of the Claude API from 32 MB to 128 MB. The change is for image-heavy requests, such as long computer-use sessions and big PDFs, that used to hit 32 MB with context window still left. If your team calls the first-party API, those requests get about four times the room, but through a cloud platform on AWS, Google Cloud or Microsoft Azure the limit is between 20 and 32 MB.
The new figures are in the request size limits table of Anthropic’s API documentation, and that table covers only the first-party API. Message Batches keeps its 256 MB limit and the Files API keeps 500 MB. The beta Sessions, Agents and Environments endpoints stay at 32 MB. Any request over its limit still fails with a 413 request_too_large error. Requests with more than 20 images also get a higher per-image limit: 3000 px, up from 2000 px.
The size cap counts bytes and is separate from the token context window. The API rejects an oversized body before it counts a single token. A screenshot uses few tokens but a lot of bytes, so a session could sit well inside its context window and still be refused. On the first-party API, Cloudflare returns the 413 before the request reaches Anthropic’s servers, so nothing in it gets processed.
Long Claude Code sessions have been running into this. Claude Code sends the whole conversation to the API on every turn, so images and documents read earlier in a session go out again with every request, Think Facility and TeamoRouter explain in their write-ups of the error. In one report they describe, the file being read was 49KB, while about 38 screenshots read earlier in the session were still riding along on every request.
Both write-ups also describe how Claude Code words the error under the 32 MB cap. When a request goes straight to Anthropic’s API, Claude Code measures the conversation and picks between two wordings. One reads “Request too large (max 32MB; 20.1MB of about 33.4MB is images or documents)”, after which it retries by itself with the media stripped. The other says compacting cannot make the request fit because the text alone is over the limit, and there is no retry. Through a proxy, a gateway or a cloud provider, Claude Code shows the general message, since it can’t measure what the other side counted.
The usual workaround the two outlets list is /compact, which summarizes the conversation and drops the images and attachments that have piled up. Otherwise, pressing Esc twice steps back, or /clear starts over. The docs they cite also suggest pointing Claude at a large file by its path rather than pasting it in.
Up next
Claude grades its own chats for misuse, catching 78%
Every time Claude finishes a reply, it answers one more hidden prompt about whether the chat was safe. On Fable 5 that check catches 78% of what a stronger classifier flags. Would you rather your AI flag its own chats or leave it to a separate model?
Related stories
- Claude API’s first new version since 2023 drops temperature
- Script checks whether your network breaks Claude API uploads
- Anthropic revamps Agent SDK pricing
- Claude limit meter may charge 1.4x more in US mornings
- Claude gets 40% off for every level of US government
- Opus 5.5 cut cache reads from $0.50 to $0.20 per MTok
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

