Claude API 128 MB limit: what changed and how to use it

The Claude API now accepts requests up to 128 MB on the Messages and Token Counting endpoints, four times the old 32 MB cap. You don't have to change any code to get it. Big PDFs, stacks of images and long computer-use sessions that used to fail with a 413 error will now go through, as long as they still fit in the model's context window.

The post on X

Anthropic announced it on Thursday, October 29, in a post from its developer account, not a blog post. @ClaudeDevs said it had "raised the max request size for Messages and Token Counting" from 32 MB to 128 MB. The second half of the post explained why. Image-heavy requests like long computer-use sessions and large PDFs were hitting the size cap while they still had plenty of context window left. Two days later the post was close to 450,000 views.

The docs lagged behind. Startup Fortune pointed out that the API reference still showed 32 MB when the post went out. The API overview now shows the new number.

Why 32 MB ran out so early

There are two different limits on a request. The context window counts tokens. The request size counts bytes on the wire. A PDF or a screenshot can be small in tokens and large in bytes, and base64 encoding, which is how most people send files inline, adds about a third to the file size. Startup Fortune worked through an example: a 24 MB PDF was already close to the old ceiling before you added a single word of prompt.

Computer-use agents hit the same wall from another direction. Each step adds another screenshot, and after a long session the screenshots alone could push a request past 32 MB while the token count was nowhere near its limit.

The current limits

Here is what Anthropic's docs list for the first-party Claude API:

Go over these and you get a 413 request_too_large error. The new limit only covers Anthropic's own API. The docs say Amazon Bedrock is capped at 20 MB and Claude on Google Cloud (Google managed) at 30 MB. Claude Platform on AWS, Claude on Google Cloud (Anthropic managed) and Microsoft Foundry allow Messages requests up to 32 MB. If you set up the Anthropic-managed Google Cloud option from our Claude on Google Cloud guide, you're still on the old 32 MB cap for now.

How to use the extra room

1. Leave your code alone. The limit went up on the server side, so a request that used to bounce at 33 MB will now just work.
2. Count tokens before you send. Token Counting got the same 128 MB limit, which means you can send the whole payload there first and check that it fits in the context window before you pay for a real call.
3. Watch the page limit on PDFs. The PDF docs cap a request at 600 pages, or 100 pages when the context window is under 1M tokens. Both the page and size limits apply to the whole payload, not to each file.
4. For large documents you reuse, use the Files API. Anthropic's own advice is to upload big PDFs once and reference them by file_id, which keeps every call small. Sending base64 inline is fine for one-off jobs.
5. Move bulk jobs to Batches. If you run a folder of contracts or invoices overnight, the Batches API takes 256 MB per request, and the docs list it at half the price.

The PDF docs also warn that dense files, meaning small fonts, heavy tables or lots of graphics, can fill the context window before they reach the page limit. Each page is processed as an image, so if a large file keeps failing, try splitting it into sections or downsampling the images inside it.

Our read

On its own this is a small change, but it fixes a real annoyance, and it helps the jobs that pay off most for an ordinary business: reading long scanned documents and running agents that click through software built for people. A 40 MB inspection report or a stack of photographed receipts used to need a preprocessing step to get it under the cap. For most of those files, you can now drop that step.

We've noticed this with Claude a few times this fall: the quiet changes are often the useful ones. The Fable 5.1 cache reads change was similar. The list price stayed put and the bill went down anyway.

What it means for businesses using AI

A lot of the document automation projects we see get stuck on plumbing like this, such as file size limits or which cloud the model runs on. Those details decide whether a workflow runs on its own or needs someone checking on it every morning. If someone on your team reads a pile of PDFs or screenshots by hand, New Face Design's free process audit can tell you whether that work is ready to automate and what it would take.