Cloudflare Markdown for Agents now accepts 6 MiB: which response headers must clients update?

What does the higher 6 MiB limit actually cover?

Cloudflare announced on October 9 that Markdown for Agents now uses an in-process streaming engine at the edge. It starts processing HTML as it arrives, rather than buffering the full response and sending it to a separate service. The conversion limit also rises from 2 MiB to 6 MiB, or 6,291,456 bytes.

That limit applies to HTML after decompression. A small compressed download can still exceed it once decompressed; a shorter Markdown result does not make oversized source HTML eligible. The announcement does not specify an activation time for each site, so check your own configuration.

The current documentation lists Pro, Business, Enterprise and SSL for SaaS customers as eligible, with no additional charge for conversion itself. Existing plan fees still apply, and this does not establish the same eligibility for the Free plan.

Why might an Accept header still return HTML?

Content negotiation lets a client say which response format it wants. For this feature, the request's Accept header must include text/markdown, Markdown for Agents must be enabled for the site's Cloudflare zone, and the source response must be HTML.

A site administrator can select the relevant zone in the Cloudflare dashboard and enable the feature in AI Crawl Control. To enable it only for a particular subdomain or path, set the matching conditions in Configuration Rules and set Markdown for Agents to On.

The client should check that the actual response's Content-Type is text/markdown, rather than merely confirming that it sent the request. If the feature is not enabled, the source is not HTML or access is restricted, adding Accept does not replace site configuration or access authorization.

How do clients measure usage without token-count headers?

Conversion responses no longer generate x-markdown-tokens or x-original-tokens. Clients that previously read counts from those headers must calculate them themselves. A missing header does not mean zero tokens and, by itself, does not prove that conversion failed.

Once the complete content has arrived, count it using a method appropriate to the target model. To compare HTML and Markdown, retain each actual input and the basis for its count. The claim that the new engine reduces conversion overhead does not guarantee a fixed saving for every page.

How can a client tell the response is complete without Content-Length?

A streaming response cannot provide the final Markdown length in advance, so Cloudflare removes Content-Length instead of recalculating it. Clients should consume the stream using their HTTP library and process the complete document after the response ends normally.

Review code that allocates storage from the stated length or rejects responses without that header, while retaining read-size limits and timeouts. Its absence alone does not mean the body is incomplete. The original HTML length should not be used to validate the converted Markdown either.

What changes for caching, cookies and content-use policies?

Conversion changes Content-Type to Markdown and adds Accept to Vary while retaining the origin's existing Vary dimensions. This helps caches distinguish HTML and Markdown versions. Origin headers related to security, caching, cookies and CORS are preserved, so existing access permissions still apply.

Headers describing the original body, including ETag and Last-Modified, are removed because the converted result cannot honor their original conditional-request semantics. Clients using If-None-Match or If-Modified-Since should stop assuming that the original HTML validators will produce the same caching behavior for a Markdown response.

If the origin already provides content-signal, conversion preserves its content-use policy. Otherwise, the defaults added are ai-train=yes, search=yes and ai-input=yes, covering training, search and AI input respectively. Check that these settings match the site's intentions before enabling the feature; do not enable training by default merely to pursue GEO visibility.

Which page should you check first before integrating?

Consider a hypothetical team whose agent reads a product documentation page. The compressed download is only a few hundred KiB, but the decompressed HTML is 7 MiB. It still exceeds the new 6 MiB limit. Empty token-count headers should not be interpreted as zero cost or successful conversion. This example has not been executed.

First, confirm the enablement scope in a zone you are authorized to configure. Then request a known HTML page with the appropriate Accept header. Check Content-Type, Vary and the actual body, confirm that the client finishes reading the stream, and handle the missing token headers with local counting logic.

Next, inspect cache validation behavior and the actual Content-Signal value, retaining unexpected responses alongside the original HTML for comparison. Format conversion does not prove that a model understands the page, that a search engine has indexed it or that an AI system has cited it. If the result matters to a task, also check that the converted body preserves the information the task needs.