Claude API now takes 128 MB requests, four times the old 32 MB cap

Anthropic has raised the maximum request size for the Messages and Token Counting endpoints on the Claude API from 32 MB to 128 MB, a fourfold increase. The company said image-heavy jobs, such as long computer-use sessions and large PDFs, had been hitting the old ceiling.

In its post announcing the change, Anthropic wrote: “We’ve raised the max request size for Messages and Token Counting on the Claude API from 32 MB to 128 MB.” The mirrored post is cut off in the text we could read, so the rest of the company’s explanation is not quoted here. The new figure is also printed in Anthropic’s own API documentation, which is where the details below come from.

What changes in the request limits

The API overview page now lists 128 MB as the maximum request size for POST /v1/messages and POST /v1/messages/count_tokens. A request over the limit returns a 413 request_too_large error, according to the same page. Nothing in the documentation suggests a new header, flag or plan requirement; the limit is listed as a property of the first-party Claude API.

The other endpoints keep their own caps. The Message Batches API accepts up to 256 MB, the Files API up to 500 MB, and the Sessions, Agents and Environments endpoints for Claude Managed Agents stay at 32 MB. Messages and Token Counting are the only ones that moved in this announcement.

Anthropic’s PDF documentation has been updated to match. It lists 128 MB as the maximum request size for PDFs on the first-party API, and a maximum of 600 pages per request, falling to 100 pages when the request’s context window is under 1M tokens. The page also warns that dense PDFs can fill the context window before the page limit is reached, and that each page is processed as an image, so downsampling embedded pictures can still help.

For developers, the practical effect is less plumbing. Computer-use loops that send a screenshot on every step build up large request bodies as the conversation grows, and a long session could previously hit the 32 MB wall well before the model’s context became the constraint. Large scanned documents encoded as base64 had the same problem, since base64 inflates a file by roughly a third. Teams that worked around the cap by splitting documents, compressing screenshots or uploading through the Files API now have more room to send everything in one call.

The new ceiling is not the same everywhere

The most useful caveat sits in a note on the overview page. The 128 MB figure applies to the first-party Claude API only. According to Anthropic’s documentation, Amazon Bedrock limits requests to 20 MB, and Claude Models on Google Cloud (Google managed) limits them to 30 MB. Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry support Messages and Token Counting requests of up to 32 MB.

That leaves a sixfold gap between the most and least generous routes to the same models. An application built around 100 MB requests will work against Anthropic’s own endpoint but fail with a size error on Bedrock, so teams that run across clouds need to design for the smallest limit or branch by platform. Anthropic recently added a way to use the full Claude API from a Google Cloud account, and its documentation lists that Anthropic-managed route at 32 MB, so the larger limit does not follow customers there.

Open questions for heavy users

A bigger request body does not make the model’s limits any wider. Context window size, page counts and per-image processing still apply, and the documentation says large PDFs can fail before reaching the page limit even through the Files API. A 128 MB request is therefore a ceiling on bytes, not a promise that any 128 MB payload will be accepted or processed well.

Several points are not established by the sources we could read. Anthropic has not said, in the material available, whether latency or timeouts change for very large payloads, whether rate limits or token billing are affected, or when the other platforms might follow. The documentation also does not say how many screenshots or pages the new limit translates to, since that depends on image format and resolution. Claims about which workloads were blocked come from Anthropic’s own description and have not been independently checked.

For now, the change is a quiet but concrete one for people building agents and document pipelines on the direct API: the bottleneck moves from the request size to the model’s context, and the cloud marketplaces are the places where the old, smaller limits remain.