feat(litellm): add cost tracking to agent results#1911
Open
stefanoamorelli wants to merge 4 commits intostrands-agents:mainfrom
Open
feat(litellm): add cost tracking to agent results#1911stefanoamorelli wants to merge 4 commits intostrands-agents:mainfrom
stefanoamorelli wants to merge 4 commits intostrands-agents:mainfrom
Conversation
I'm adding a 5th element to the ModelStopReason tuple to carry per-invocation cost data (in USD [1]) through the streaming pipeline. The field defaults to None so all existing model providers continue to work without changes. The MetadataEvent TypedDict also gets an optional cost field, which is where model providers will inject their cost before it reaches the stop event. Existing consumers of the stop tuple (anthropic, bedrock, and the summarizing conversation manager) now use *_ unpacking so they're forward-compatible with the new element. [1]: https://docs.litellm.ai/docs/completion/token_usage#critical-cost_per_token
EventLoopMetrics now has an accumulated_cost field (defaults to 0.0) and an update_cost() method that the event loop will call after each model invocation. The cost in USD [1] is included in get_summary() and displayed in the metrics summary output when it's greater than zero. This is the accumulation layer that sits between the streaming pipeline (which provides per-invocation cost) and the user-facing AgentResult. [1]: https://docs.litellm.ai/docs/completion/token_usage#critical-cost_per_token
process_stream() now extracts the cost field from MetadataEvent (if present) and passes it as the 5th element of the ModelStopReason tuple. On the event loop side, I unpack the cost and call EventLoopMetrics.update_cost() when a value is available. This connects the model layer (which calculates cost) to the metrics layer (which accumulates it), completing the data flow for any model provider that populates MetadataEvent.cost.
This is the actual cost calculation that makes use of the pipeline built in the previous commits. In format_chunk(), after extracting usage data, I call litellm.cost_per_token() to get prompt and completion costs and attach the total to MetadataEvent. The values returned by cost_per_token() are in USD [1][2], which is what we store in accumulated_cost. The calculation is wrapped in try/except because litellm's pricing database doesn't cover every model. When a model isn't mapped, the cost field is simply omitted and the rest of the pipeline continues as if cost tracking isn't available. I chose cost_per_token() over completion_cost() because it doesn't require constructing a fake ModelResponse object. Cache tokens (both read and creation) are forwarded to the cost function so pricing accounts for cached token discounts on providers like Anthropic. Closes strands-agents#1216 [1]: strands-agents#1216 [2]: https://docs.litellm.ai/docs/completion/token_usage#critical-cost_per_token "Returns: A tuple containing the cost in USD dollars for prompt tokens and completion tokens, respectively."
stefanoamorelli
requested a deployment
to
manual-approval
GitHub Actions
Waiting
stefanoamorelli
requested a deployment
to
manual-approval
GitHub Actions
Waiting
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Tip
Follows conventional commits. Better reviewed commit-by-commit.
Exposes LiteLLM's
cost_per_token()data (in USD) through the Strands streaming pipeline so that accumulated cost is available inEventLoopMetricswhen usingLiteLLMModel.The cost flows through:
LiteLLMModel.format_chunk()→MetadataEvent→process_stream()→ModelStopReason→ event loop →EventLoopMetrics.accumulated_cost.Cost calculation is wrapped in
try/exceptso models not yet mapped in LiteLLM's pricing database degrade gracefully (cost is simply omitted). TheModelStopReasontuple is extended from 4 to 5 elements; existing consumers use*_unpacking for forward-compatibility.Related Issues
Closes #1216
Documentation PR
N/A
Type of Change
Testing
hatch run prepareChecklist
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.