Futureweb

Distributed Cloud

The auto router opened in September

Cloudflare's model router for the AI Gateway is a public beta dated September 30. An October 26 news account is not a second launch, and a savings line is the company's.

The short version

No. Cloudflare's auto router post is dated September 30, 2026, a public beta in AI Gateway under the model value cloudflare/auto. The docs page was last updated October 2. InfoQ wrote it up on October 26. A company line about up to 30 percent savings is from internal use, and the company also says its router can lose its own benchmark.

September 30 is the post

Cloudflare published the auto router as a public beta on September 30, 2026, inside AI Gateway, selected with the model value cloudflare/auto. The developer docs for the feature show an update on October 2. InfoQ's October 26 story is how the feature reached this morning's wire. A cache rule about which bytes an origin will return is a different Cloudflare feature, already filed, and it does not route a model. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.A byte range is not a model pick.

The classifier is a description, not a recipe

The company says the router rates recent turns on task categories and on four dimensions it names: complexity, ambiguity, stakes, and how much the turn depends on earlier context. That is enough to know what the product claims to look at. The arithmetic that turns those ratings into a model choice stays in the documentation. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.

Thirty percent is an internal ceiling

Up to 30 percent cheaper than using only frontier models is Cloudflare's figure for its own OpenCode use. Up to is the company's hedge. A customer with a different mix of calls can land somewhere else, including, on the company's own account, behind the baseline. Asked directly, Where does 30 percent come from? Cloudflare's account of its own internal OpenCode use, as a ceiling of up to 30 percent against using only frontier models. It is not a receipt from a customer's bill.

A loss on the house benchmark

InfoQ records the company's admission that the router can lose Cloudflare's internal comparison. A beta that sometimes picks a worse model than the simple policy is still a beta. A closed test of a different Cloudflare feature, geographic keys, is a separate product and does not grade this router. Asked directly, Does this page print the session header? No. The docs name how a conversation stays on one route. The header itself is an implementation detail and stays in the docs. The document named in the sources is the record this page will keep.A closed test is a different product.

Session stickiness stays in the docs

The docs describe a way for later turns of one conversation to keep the route the first turn chose. This page does not print the header or a request body. A name in a certificate is not a route, and a disabled credential at another company is not this gateway. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.A name is not a route.

Who is in the beta

Customers of AI Gateway who opt into the model value are the people the post addresses. A reader who never sends traffic there is not receiving the 30 percent sentence. A context window still bounds each model the router might pick. Asked directly, What does the company say it measures? A classifier looks at recent turns across task categories and scores complexity, ambiguity, stakes, and dependence on earlier context. The company describes a score that trades expected quality against cost. The formula stays in the docs.Each model still has a window.

What an October dateline cannot do

Filing the story on October 26, because that is when a newsroom explained it, does not move the launch to October 26. The post is September 30. A credential disabled at a database company is a different October incident and does not configure this router. Asked directly, Can the router lose? InfoQ notes that Cloudflare says its own router can lose the company's internal benchmark. A product that publishes a loss is still a beta, not a proof it always wins. The document named in the sources is the record this page will keep.A disabled credential is not a router.

Public beta is the status word

The post says public beta. A beta can change its categories, its prices, and its default. Treating the September behavior as a permanent tariff would outrun the document. Asked directly, What is the product called, and when did the post go up? Auto router, in AI Gateway, model value cloudflare/auto. The blog post is September 30, 2026. The docs say they were updated October 2. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.

What would make the savings a bill

An invoice, or a method a customer can rerun on their own traffic and publish. The internal ceiling is not that invoice. The short answer this page will stand on is this: No. Cloudflare's auto router post is dated September 30, 2026, a public beta in AI Gateway under the model value cloudflare/auto. The docs page was last updated October 2. InfoQ wrote it up on October 26. A company line about up to 30 percent savings is from internal use, and the company also says its router can lose its own benchmark.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Cloudflare blog
  2. Cloudflare docs
  3. InfoQ

Questions

What is the product called, and when did the post go up?

Auto router, in AI Gateway, model value cloudflare/auto. The blog post is September 30, 2026. The docs say they were updated October 2.

What does the company say it measures?

A classifier looks at recent turns across task categories and scores complexity, ambiguity, stakes, and dependence on earlier context. The company describes a score that trades expected quality against cost. The formula stays in the docs.

Where does 30 percent come from?

Cloudflare's account of its own internal OpenCode use, as a ceiling of up to 30 percent against using only frontier models. It is not a receipt from a customer's bill.

Can the router lose?

InfoQ notes that Cloudflare says its own router can lose the company's internal benchmark. A product that publishes a loss is still a beta, not a proof it always wins.

Does this page print the session header?

No. The docs name how a conversation stays on one route. The header itself is an implementation detail and stays in the docs.