Skip to content

feat(ai): Fix cost calculation for cached and reasoning tokens #4922

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jul 9, 2025
Merged

feat(ai): Fix cost calculation for cached and reasoning tokens #4922

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
- Add support for playstation data requests. ([#4870](https://github.com/getsentry/relay/pull/4870))
- Expand the NEL attributes & others. ([#4874](https://github.com/getsentry/relay/pull/4874))
- Normalize legacy AI agents attributes to OTel compatible names. ([#4916](https://github.com/getsentry/relay/pull/4916))
- Fix cost calculation for cached and reasoning tokens. ([#4922](https://github.com/getsentry/relay/pull/4922))

## 25.6.2

Expand Down
20 changes: 10 additions & 10 deletions relay-event-normalization/src/event.rs
View file
Open in desktop
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Had to rearrange the token costs and amounts to have have a realistic scenarios (e.g. gen_ai.usage.output_tokens.reasoning is subset of gen_ai.usage.output_tokens)

Original file line number Diff line number Diff line change
Expand Up @@ -2330,8 +2330,8 @@ mod tests {
"data": {
"gen_ai.usage.input_tokens": 1000,
"gen_ai.usage.output_tokens": 2000,
"gen_ai.usage.output_tokens.reasoning": 3000,
"gen_ai.usage.input_tokens.cached": 4000,
"gen_ai.usage.output_tokens.reasoning": 1000,
"gen_ai.usage.input_tokens.cached": 500,
"gen_ai.request.model": "claude-2.1"
}
},
Expand Down Expand Up @@ -2382,15 +2382,15 @@ mod tests {
input_per_token: 0.01,
output_per_token: 0.02,
output_reasoning_per_token: 0.03,
input_cached_per_token: 0.0,
input_cached_per_token: 0.04,
},
),
(
"gpt4-21-04".to_owned(),
ModelCostV2 {
input_per_token: 0.09,
output_per_token: 0.05,
output_reasoning_per_token: 0.06,
output_reasoning_per_token: 0.0,
input_cached_per_token: 0.0,
},
),
Expand All @@ -2408,7 +2408,7 @@ mod tests {
.and_then(|span| span.data.value());
assert_eq!(
first_span_data.and_then(|data| data.gen_ai_usage_total_cost.value()),
Some(&Value::F64(140.0))
Some(&Value::F64(75.0))
);
assert_eq!(
first_span_data.and_then(|data| data.gen_ai_response_tokens_per_second.value()),
Expand Down Expand Up @@ -2525,8 +2525,8 @@ mod tests {
"data": {
"gen_ai.usage.input_tokens": 1000,
"gen_ai.usage.output_tokens": 2000,
"gen_ai.usage.output_tokens.reasoning": 3000,
"gen_ai.usage.input_tokens.cached": 4000,
"gen_ai.usage.output_tokens.reasoning": 1000,
"gen_ai.usage.input_tokens.cached": 500,
"gen_ai.request.model": "claude-2.1"
}
},
Expand Down Expand Up @@ -2562,8 +2562,8 @@ mod tests {
ModelCostV2 {
input_per_token: 0.01,
output_per_token: 0.02,
output_reasoning_per_token: 0.03,
input_cached_per_token: 0.0,
output_reasoning_per_token: 0.0,
input_cached_per_token: 0.04,
},
),
(
Expand All @@ -2589,7 +2589,7 @@ mod tests {
.and_then(|span| span.value())
.and_then(|span| span.data.value())
.and_then(|data| data.gen_ai_usage_total_cost.value()),
Some(&Value::F64(140.0))
Some(&Value::F64(65.0))
);
assert_eq!(
spans
Expand Down
21 changes: 17 additions & 4 deletions relay-event-normalization/src/normalize/span/ai.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,11 +32,24 @@ fn calculate_ai_model_cost(model_cost: Option<ModelCostV2>, data: &SpanData) ->

let mut result = 0.0;

result += cost_per_token.input_per_token * input_tokens_used.unwrap_or(0.0);
result += cost_per_token.output_per_token * output_tokens_used.unwrap_or(0.0);
result +=
cost_per_token.output_reasoning_per_token * output_reasoning_tokens_used.unwrap_or(0.0);
// Cached tokens are subset of the input tokens, so we need to subtract them
// from the input tokens
result += cost_per_token.input_per_token
* (input_tokens_used.unwrap_or(0.0) - input_cached_tokens_used.unwrap_or(0.0));
result += cost_per_token.input_cached_per_token * input_cached_tokens_used.unwrap_or(0.0);
// Reasoning tokens are subset of the output tokens, so we need to subtract
// them from the output tokens
result += cost_per_token.output_per_token
* (output_tokens_used.unwrap_or(0.0) - output_reasoning_tokens_used.unwrap_or(0.0));

if cost_per_token.output_reasoning_per_token > 0.0 {
// for now most of the models do not differentiate between reasoning and output token cost,
// it costs the same
result +=
cost_per_token.output_reasoning_per_token * output_reasoning_tokens_used.unwrap_or(0.0);
} else {
result += cost_per_token.output_per_token * output_reasoning_tokens_used.unwrap_or(0.0);
}

Some(result)
}
Expand Down