Ni need to get these values apparently
N
Nbut i don't get where
Nno envs here
Ni do have google ai studio there
IAPI Key needs to be generated from https://dash.cloudflare.com/?to=/:account/api-tokens and needs to have the AI Gateway permission. Gateway Endpoint looks correct already, assuming that's your proper Account ID in there.
SDo CF worker AI models have rate limits for tokens? On the limts page, I can see RPM limits but not TPMs (beyond the context window of the model)
NScribe v2 best model and not availiable?
NIt is my understanding that if we enable billing we should be able to setup an authenticated gateway and then call any LLM provider supported by AI Gateway without providing the provider’s API key (since CF is handing the billing). However, if I issue a curl and do not provide the underlying provider’s API key, I get an error stating that I must provide it (in addition to the AI Gateway’s token which I generated in CF with the run permission). This makes no sense to me. Isn’t the whole point to be able to provide ONE key, access any provider and pay one bill to CF for all inferencing. This is the way the Vercel AI Gateway works but I want to host fully on CF.
CAnyone managed to connect dynamic models with
ai-sdk? this fails with a vague internal server error
CThis works if anyone searches this:
Dhey there, what's the trick to get token count and costs on the UI? I've tried sending my own
cf-aig-custom-cost to no avail
DP.S if someone is "stuck" with Boto3 for Amazon Bedrock, it's doable https://gist.github.com/DDuarte/6bd7e3851e6519c988037f883f80eeb1
AI'm trying to use Claude 4-5 Sonnet on Azure Foundry through the AI Gateway -- is this platform currently supported? I'm only seeing "Azure OpenAI" as a provider option on the page (which, as I understand, is limited only to a select set of OpenAI models).
LIs it known that AI Gateway with Unified Billing (Beta) doesn’t seem to work with OpenAI when using the Vercel AI SDK together with Cloudflare’s ai-gateway-provider package?
In my testing:
• OpenAI → fails entirely through Unified Billing
• Google Generative AI → the non-streaming functions (generateText, generateObject) work, but the streaming variants do not
• Anthropic → both streaming and non-streaming work as expected
This seems odd, since the documentation states that Unified Billing supports all of these providers. Is this a known limitation or an issue on Cloudflare’s side?
I know it’s still beta but just want to know.
In my testing:
• OpenAI → fails entirely through Unified Billing
• Google Generative AI → the non-streaming functions (generateText, generateObject) work, but the streaming variants do not
• Anthropic → both streaming and non-streaming work as expected
This seems odd, since the documentation states that Unified Billing supports all of these providers. Is this a known limitation or an issue on Cloudflare’s side?
I know it’s still beta but just want to know.
NI hope you get an answer to this because I’m struggling with the same thing.
LHoping on an answer from an employee, they might know more
NShould be able to specify ‘openai/gpt-5’ or ‘groq/gpt-oss-120b’ as the model and it should work, but it doesn’t.
LI use the vercel ai sdk with a provider registry and do it like that instead. You can specify your provider/model that way.
LStill hoping for clarification on this
VIs there an progress on supporting
v1 and v1-preview data plane API version (in azure) for ai gateway? Currently it breaks - unable to find resource (I assume url changes)Vv1 is the new GA version for a while now and the azure adapter in cercel ai sdk now defaults to it as well, breaking any usage relying on defaultsNI’m pretty close to giving up on Cloudflare and just paying $20 to Vercel.
Mis cloudflare ai gateway down?
Why can I call some models but not others?
do I need to top up more ?
Why can I call some models but not others?
do I need to top up more ?
M{
"code": "Some resource has been exhausted",
"error": "Your team 23361c02-xxxx has either used all available credits or reached its monthly spending limit. To continue making API requests, please purchase more credits or raise your spending limit."
}
what is this team concept ? I don't have team.
"code": "Some resource has been exhausted",
"error": "Your team 23361c02-xxxx has either used all available credits or reached its monthly spending limit. To continue making API requests, please purchase more credits or raise your spending limit."
}
what is this team concept ? I don't have team.
W*
Subject:
Bug Report: Google Gemini File Upload Size Mismatch (via SDK Monkey Patch)
Description:
I am reporting a critical issue with Google Gemini (AI Studio) file uploads when routed through Cloudflare AI Gateway.
I am using the official
The Issue:
While the initial handshake succeeds, the actual file data transfer fails when routed through the Gateway. Google returns a
It seems the Gateway is mishandling the
Error Log:
(Note: The
Reproduction Details:
Reference Trace (CF-Ray):
Could the team please investigate how the Gateway proxies chunked upload streams for the Google provider?
*
Subject:
Bug Report: Google Gemini File Upload Size Mismatch (via SDK Monkey Patch)Description:
I am reporting a critical issue with Google Gemini (AI Studio) file uploads when routed through Cloudflare AI Gateway.
I am using the official
google-genai Python SDK. Since the SDK's resumable upload protocol automatically attempts to connect directly to generativelanguage.googleapis.com for the data transfer phase (Step 2), I applied a monkey patch to the SDK to intercept the upload_url and rewrite the hostname to my Cloudflare Gateway endpoint.The Issue:
While the initial handshake succeeds, the actual file data transfer fails when routed through the Gateway. Google returns a
400 Bad Request indicating that the data received is significantly larger (almost double) than the expected file size.It seems the Gateway is mishandling the
Transfer-Encoding: chunked stream sent by the Python httpx client, possibly causing data duplication or incorrect packet retries.Error Log:
(Note: The
Current size ~10MB is nearly double the Expected ~5.5MB.)Reproduction Details:
google-genai (using httpx).client._api_client.upload_file to replace googleapis.com with the Cloudflare Gateway URL.Reference Trace (CF-Ray):
9a6803f82f942a86-LAXCould the team please investigate how the Gateway proxies chunked upload streams for the Google provider?
*
Rhello, im looking to set up logpush for cloudflare ai gateway to datadog following this guide: https://developers.cloudflare.com/ai-gateway/observability/logging/logpush/.
is there a way to send these logs without encryption? it seems like datadog does not offer a way to decrypt them, and that would mean that the logpush job that i've created for these logs to be sent to datadog would be useless since they can't be decrypted at the datadog level
is there a way to send these logs without encryption? it seems like datadog does not offer a way to decrypt them, and that would mean that the logpush job that i've created for these logs to be sent to datadog would be useless since they can't be decrypted at the datadog level
Cloudflare Docs
AI Gateway allows you to securely export logs to an external storage location, where you can decrypt and process them.
You can toggle Workers Logpush on and off in the Cloudflare dashboard settings. This product is available on the Workers Paid plan. For pricing information, refer to Pricing.
You can toggle Workers Logpush on and off in the Cloudflare dashboard settings. This product is available on the Workers Paid plan. For pricing information, refer to Pricing.
LHas anyone used openrouter successfully with BYOK?
LTells me it is missing authentication, using the vercel sdk, and the openrouter provider package.
Mis the playground vide coded? so many bugs
MThese unfinished pages are definitely damaging cf's brand.
SWe're are getting the same issue. We have no server in Asia at all but all requests are getting redirected to one place which is restricted from open Ai without any control for placement from ourside.
{
"error": {
"code": "unsupported_country_region_territory",
"message": "Country, region, or territory not supported",
"param": null,
"type": "request_forbidden"
}
}
Has anyone found anyway to solve it?
{
"error": {
"code": "unsupported_country_region_territory",
"message": "Country, region, or territory not supported",
"param": null,
"type": "request_forbidden"
}
}
Has anyone found anyway to solve it?
SOur Open AI requests are also redirected to one place which is restricted. I wonder if they are related.
SThis is happening again. Our error rate has jumped to 25%. Could you have a look again? @Kathy @harshil1712
P@alex We have updated the AI SDK documentation page with examples of using the
https://developers.cloudflare.com/ai-gateway/integrations/vercel-ai-sdk/
ai-gateway-provider along with examples of using AI SDK with BYOK/Unified Billinghttps://developers.cloudflare.com/ai-gateway/integrations/vercel-ai-sdk/
PHey everyone 
@Rishit Bansal and I work on the AI Gateway team. We’re catching up on some older threads now (sorry for the slow replies there!) and will be keeping a closer eye on this space. If you have any feedback or run into a bug, drop it here or tag us, we'll jump in as quickly as we can
@Rishit Bansal and I work on the AI Gateway team. We’re catching up on some older threads now (sorry for the slow replies there!) and will be keeping a closer eye on this space. If you have any feedback or run into a bug, drop it here or tag us, we'll jump in as quickly as we can
RWe are adding support for showing usage for more and more models/providers! There should be a fix releasing out sometime this week which will also add support for price calculation on embedding models on google ai studio/workers ai / vertex and other providers. If there's something specific which you immediately see that doesn't work: do let us know, and we'll add this in on a higher priority. But we also have some internal alerting setup which lets us know when usage calculation fails, which we keep an eye on!
Regarding the playground: yes, currently it only supports text generation models which specifically use the chat completions API format. We're working on limiting the list of models which shows on the playground to a smaller list, and also simultaneously seeing if we can have support for a few image generation/embeddings models there as well.
Regarding the playground: yes, currently it only supports text generation models which specifically use the chat completions API format. We're working on limiting the list of models which shows on the playground to a smaller list, and also simultaneously seeing if we can have support for a few image generation/embeddings models there as well.