[alexcastgonz (pcbeingused333)]

Motivation and Context

The model diagnostics record token usage only when metadata["usage"] is a CompletionUsage (utils/telemetry/model_diagnostics/decorators.py). The other chat connectors build one, but AnthropicChatCompletion stores the raw anthropic.types.Usage (non-streaming) or a plain dict like {"input_tokens": ..., "output_tokens": ...} (streaming). As a result, spans for Anthropic calls never get gen_ai.usage.input_tokens / gen_ai.usage.output_tokens.

Prompt caching also isn't counted: Anthropic reports cache_read_input_tokens and cache_creation_input_tokens separately from input_tokens, and the streaming path drops them.

Description

• A new _get_completion_usage helper turns Anthropic usage into a CompletionUsage:
  • prompt_tokens = input_tokens + cache_read_input_tokens + cache_creation_input_tokens;
  • prompt_tokens_details.cached_tokens = cache_read_input_tokens (only when present);
  • completion_tokens = output_tokens.
• Non-streaming: metadata["usage"] is now that CompletionUsage.
• Streaming: message_start sets the prompt side. The final message_delta chunk carries the full usage: its own counts when it reports input_tokens, otherwise the message_start prompt counts plus the delta's output_tokens. Since the chunk metadata is merged with |, this is the usage of the combined message.
• Fields are read with getattr because older anthropic releases (the requirement is ~= 0.32) lack some of them.

Behavior change: metadata["usage"] for Anthropic responses is now a CompletionUsage, the same type as the other connectors. Code that read .input_tokens / .output_tokens from it should use .prompt_tokens / .completion_tokens. The raw Anthropic usage is still available on inner_content.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄 (see the behavior change above)

Tests, in tests/unit/connectors/ai/anthropic/services/test_anthropic_chat_completion.py, with model diagnostics enabled, 10 input + 100 cache-read + 20 cache-write + 7 output tokens:

• test_usage_is_reported_as_completion_usage (non-streaming): metadata["usage"] is a CompletionUsage (130 prompt, 100 cached, 7 completion), and the span gets gen_ai.usage.input_tokens=130, gen_ai.usage.output_tokens=7;
• test_streaming_usage_is_reported_as_completion_usage: the same checks on the combined streamed message.

Both fail on main. tests/unit/connectors/ai/anthropic + tests/unit/utils/model_diagnostics: 65 passed (test_connector_decorated.py was not run here because it imports the Mistral connector). ruff 0.15.5 check/format is clean. mypy on the changed module reports the same 3 pre-existing errors as main, all in other modules.

This change was drafted with Claude Code; the tests above were run before opening the PR.

🤖 Generated with Claude Code

[Copilot]

🟢 Approval recommended

The focused conversion and streaming changes match the diagnostics contract, with regression coverage and no identified blocking issues.

No open issues 🎉

What changed in this PR

Normalizes Anthropic token usage to CompletionUsage so model diagnostics can report input and output token counts.

Changes:

• Includes cache-read and cache-creation tokens in prompt totals.
• Preserves prompt usage when merging streaming output counts.
• Adds regression tests for response metadata and telemetry.

File	Description
python/​tests/​unit/​connectors/​ai/​anthropic/​services/​test_anthropic_chat_completion.py	Tests streaming and non-streaming usage reporting.
python/​semantic_kernel/​connectors/​ai/​anthropic/​services/​anthropic_chat_completion.py	Converts Anthropic usage and preserves streaming prompt counts.

_{Review effort: Balanced}

[PRABHU KIRAN VANDRANKI (VANDRANKI)]

I ran this end to end and the telemetry fix works. Setup: Python 3.13.3, anthropic SDK 1.15.0, opentelemetry-sdk with an in-memory exporter, SEMANTICKERNEL_EXPERIMENTAL_GENAI_ENABLE_OTEL_DIAGNOSTICS=true. The PR diff applied cleanly onto current main (fetched 2026-10-16) and I compared it with a clean main checkout.

Probe: a local HTTP server that speaks the Anthropic Messages API. The response (and the SSE message_start) carries input_tokens=10, cache_read_input_tokens=100, cache_creation_input_tokens=5. The non-streaming output_tokens is 2, and the streaming message_delta carries output_tokens=7. A real AsyncAnthropic(base_url=local) is passed to AnthropicChatCompletion.

	main	PR
non-stream metadata["usage"]	raw Anthropic Usage(...)	CompletionUsage(prompt_tokens=115, cached_tokens=100, completion_tokens=2)
stream metadata["usage"]	dicts {'input_tokens': 10} then {'input_tokens': 10, 'output_tokens': 7}	CompletionUsage with prompt 115, and completion 7 on the last chunk
SK chat claude-sonnet-4-5 span attributes (non-stream)	none	gen_ai.usage.input_tokens=115, gen_ai.usage.output_tokens=2
SK chat claude-sonnet-4-5 span attributes (stream)	none	input 115, output 7

So the missing token attributes are fixed, and the cached tokens are counted in the prompt side.

Edge cases on the streaming path:

• message_delta with output_tokens only (what I used above): prompt 115 is kept from message_start. Good.
• message_delta with input_tokens=10 and the two cache fields: prompt 115. Good.
• message_delta with input_tokens=10 but no cache fields (I made this up to test the code path): the final usage becomes prompt_tokens=10, prompt_tokens_details=None, so the 105 cache tokens from message_start are dropped. The code says "use the delta counts when it reports input_tokens". I do not know whether the real API ever sends input_tokens without the cache fields in a delta, so this may never happen. If you want to be safe, fall back to the message_start cache values when the delta leaves them out.

Tests: tests/unit/connectors/ai/anthropic gives 37 passed on the PR. When I copy the PR's test file onto main, the two new tests fail (test_usage_is_reported_as_completion_usage, test_streaming_usage_is_reported_as_completion_usage) and 35 pass.

Two notes. The PR text already says this is a behavior change for code that read .input_tokens from the old usage object. I did not search outside this repo for such code. Also, PromptTokensDetails has a cache_write_tokens field and the PR leaves it None even though cache_creation_input_tokens is known. That is optional.

I ran only against a local server, not the real Anthropic API. I did not check how this combines with #14616, which edits nearby lines in the same file.

[alexcastgonz (pcbeingused333)]

@microsoft-github-policy-service agree