Anthropic Lifts Claude API Request Cap to 128 MB

Anthropic raised the raw HTTP payload ceiling on the Messages API and the Token Counting endpoint from 32 MB to 128 MB, officially launched on October 30, 2026. The company announced the change via X rather than a blog post, framing it as a fix for a specific, recurring complaint: developers running image-heavy requests were hitting the 32 MB wall while context window capacity sat unused. The new cap unblocks document analysis workflows and screenshot-laden computer-use sessions without touching context window pricing or token limits.

The 32 MB figure was never about what Claude could reason over. It measured the raw bytes of the HTTP payload hitting the API, not tokens. A multi-turn computer-use agent session could be nowhere near its token budget and still fail with a 413 error simply because file attachments pushed the request body past 32 megabytes once base64-encoded. Anthropic's wording in the X post, plenty of context window left, is a tell. Developers were paying for a large context window and then losing access to it for an unrelated reason: the size of the wrapper around the data, not the data's conceptual weight.

Why Base64 Encoding Made the Old Limit Bite Early

Base64 encoding inflates binary files like images and PDFs by roughly a third during transport. A 24 MB PDF was already flirting with the 32 MB ceiling before a single token of instructions or conversation history got added. For a computer-use session where Claude takes repeated screenshots of a screen to decide its next click, those screenshots stack up fast inside one request even though the actual reasoning load barely registers against a 200,000-token or 1-million-token context window.

Developer complaints about this gap have surfaced in Anthropic's own public issue trackers, where the 32 MB cap was described as hard-stopping entire conversations rather than letting the model recover or truncate gracefully. The error looked like a usage problem when it was actually an infrastructure ceiling that had nothing to do with how much Claude could think about.

Release Date and API Tier Breakdown

Anthropic officially released the 128 MB limit on October 30, 2026. The jump targets precisely the tier that image-heavy, interactive workloads use: the Messages API and Token Counting endpoint. Message Batches stays at 256 MB and the Files API stays at 500 MB. Anthropic's own documentation, as of this week, still lists the old limits alongside the new 128 MB figure, so teams checking the API reference will see the updated ceiling reflected for the endpoints that needed it most.

Four times the room means a computer-use agent can run roughly four times as many screenshot-laden turns, or a workflow can attach a meaningfully larger PDF, before running into the same wall again. The fix doesn't touch pricing, and Anthropic didn't announce any change to the context window itself in the same post. The update is narrower than that, and more useful for it: it closes the gap between what the model could already process and what the transport layer would let through.

What This Means for Computer-Use Agents and Document Workflows

For teams building document analysis tools or autonomous computer-use agents, this change shows up as fewer retries and fewer workarounds. No more chunking a PDF into separate requests just to dodge a limit that had nothing to do with comprehension. No more truncating screenshot histories mid-session because the HTTP envelope filled up before the context window did. The 128 MB cap is a small, specific fix. It's also an admission that the 32 MB number was never really about what Claude could handle, just about what could get to it.

This update removes a friction point that penalized exactly the workflows Anthropic has marketed most heavily: multi-turn agentic sessions and large-document reasoning. The new ceiling aligns the transport layer with the model's actual capacity, so developers can use the context window they're already paying for without hitting an unrelated infrastructure limit halfway through a session.