[aikins01]

A failed model call now records the provider's error code in error.type and exception.type when its class name says nothing about the cause. Before, an exhausted openai quota and a rate limit were both RateLimitError, and an Anthropic stream that failed with overloaded_error, a Gemini 429, and an AI SDK RetryError gave generic class names. telemetry.dev groups failures and alert incidents by error type, so these failures could not be told apart.

The TypeScript SDKs and the Python SDK use the same rule, in errorTypeOf (packages/otel/src/errors.ts) and _error_type (sdks/python/src/telemetry_dev/_spans.py):

• An error with an HTTP status from 400 to 599 and a code that is not only digits records the code, such as insufficient_quota or context_length_exceeded on an openai error.
• A generic provider error class records its code, else its HTTP status as a string (429), else the provider's error type (overloaded_error). The generic classes are listed in docs/sdk-conformance.md. In Python, a boto ClientError gives its response.Error.Code.
• The AI SDK's AI_RetryError records the type of the last error that it retried.
• Any other error records its class name, as before.

Numeric codes, such as Z.ai's "1302", are not used, because a number names no cause and the class (RateLimitError) does. A getter on a third-party error that throws gives the fallback type and does not break the span.

The core SDKs, the AI SDK v6 and v7 integrations (also the media and batch spans, which set no error.type before), and TanStack AI call the shared function, so integrations that end spans through the core SDK, such as LangChain, get the rule too.

Failures that a provider reports in a response body or a stream event, not as an HTTP error, recorded Error or RuntimeError. The integrations now build them with the provider's code and a type that names the failure:

Failure	error.type
openai or OpenRouter response.failed, stream error event, openai Realtime error	the code, such as rate_limit_exceeded, else response.failed or response.error
OpenRouter numeric code	the code as an HTTP status, such as 502
LiteLLM Responses failure	the code, else response.failed
failed Google video operation	the google.rpc.Code name, such as RESOURCE_EXHAUSTED
Bedrock stream exception in Python	ModelStreamErrorException, as in TypeScript, in place of EventStreamError

A TypeScript openai image stream that ends with a failed event now marks its span failed, as the Python SDK does.

After the release, an alert that groups incidents by error type sees the new values as new error types one time, for example insufficient_quota in place of RateLimitError.

Tests through the real provider clients with a fake network: openai (TypeScript and Python) for insufficient_quota, context_length_exceeded, a 500 with no code, and a numeric code; Anthropic (TypeScript and Python) for a stream that fails with overloaded_error; Google GenAI (TypeScript 400, Python 429); and the AI SDK v7 for RetryError, an APICallError 529, and a stream provider error. The existing tests of each body or stream failure now check the code, and new tests cover a failed image stream, a failed Realtime response, and a failed Google video operation in TypeScript.

I ran vp test run in the changed packages (openai also with OPENAI_SDK_VERSION=6), vp check, and pytest, ruff, and pyright in each Python package.