$12 debugging receipt for a 400 model not found

01 / The receipt starts with $12 of developer time

A request returns HTTP 400 with “model not found.” You send it again. Same result. Before changing prompts or buying more credits, check the address and model identifier. This error commonly points to request configuration or model availability, not prompt quality. The response body is your first clue, not a complete diagnosis.

Here is a planning receipt, not a measured benchmark or a CheapRelay price quote. Assume developer time costs $60 per hour and reserve 12 minutes for diagnosis. That makes the labor budget $12. A small, successful verification request has a separate token cost; failed-request charges must be checked rather than assumed.

• ■ Read the error and capture request details: 2 minutes, $2.
• ■ Verify the API base and route: 3 minutes, $3.
• ■ Verify the model identifier and access: 3 minutes, $3.
• ■ Run a minimal request and inspect client overrides: 4 minutes, $4.
• ■ Total planned labor: 12 minutes, $12.

02 / Spend the first $2 reading the actual response

Record the status, error message, timestamp and request ID if provided. Capture the final request URL and submitted model value. Redact authorization headers and sensitive prompt content before sharing logs.

Do not treat every 400 as the same problem. Read any error code or parameter field alongside the message. A relay may reject an identifier before contacting an upstream service, while another error may describe account access or route compatibility. The status alone does not establish the cause.

Pause automatic retries during diagnosis. Repeating an unchanged request will not repair an incorrect identifier. Also, do not conclude that every failed request is free: inspect usage records and the service’s billing rules.

03 / Spend $3 checking the host and route together

CheapRelay’s API base is https://api.cheapapirelay.com. OpenAI-compatible routes sit under /v1; Gemini native routes sit under /v1beta. Choose the route family your client expects. A native request format and an OpenAI-compatible request format are not interchangeable merely because both reach an AI service.

For an OpenAI SDK, check whether its base URL should include /v1. Inspect the final outgoing URL rather than guessing: a wrapper might append the version itself, producing an accidental duplicate path. Also confirm that the request actually reaches the relay rather than a provider’s default host.

Check the endpoint as well as the prefix. A model’s availability does not guarantee support for every operation. Use the documented endpoint for the selected model and request format.

04 / Spend $3 replacing the display name with an exact ID

A catalog label such as Gemini 2.5 Flash is not necessarily the API identifier. Copy the exact identifier from current service documentation, the dashboard or a model-list response where supported. Do not invent capitalization, punctuation, aliases or version suffixes.

If your relay exposes GET /v1/models, query it with the same key used by the failing application. Treat the result as evidence of what that key can discover, then check endpoint support separately. If the identifier is missing, investigate current availability and account access before editing the prompt.

05 / The token receipt: $0.0004 under explicit assumptions

For arithmetic only, assume an input rate of $1 per million tokens and an output rate of $4 per million tokens. These are illustrative rates, not advertised CheapRelay prices. A successful test using 200 input tokens and 50 output tokens would cost:

• ■ Input: 200 divided by 1,000,000, multiplied by $1 = $0.0002.
• ■ Output: 50 divided by 1,000,000, multiplied by $4 = $0.0002.
• ■ Total: $0.0004 per successful test; ten identical tests total $0.004.

Actual charges depend on the selected model, recorded usage and applicable cache rates. CheapRelay bills per real token, including cache reads and writes, with no monthly fee or subscription.

06 / Use the final $4 to isolate the client

Send one minimal request with a short prompt, a small output limit and no tools or attachments. If it succeeds, compare its host, route and model field with the application’s outgoing request. Check environment variables, project settings and command-line overrides; the configuration file you edited may not control the running process.

If you need a pay-as-you-go relay for this workflow, sign up for CheapRelay. Keep the working request as a smoke test. The useful receipt is straightforward: a $12 troubleshooting budget, explicitly priced test tokens, and no blind retry loop hiding the real configuration error.