Skip to content

Switchyard lets LLM applications route traffic across models and providers while preserving native OpenAI and Anthropic API compatibility - enabling flexible model selection, benchmarking, and cost/performance optimization.

Rust

fix(translation): preserve Anthropic thinking-token usage (#769)

nachiketb-nvidia7508c22f4b37
diff --git a/crates/switchyard-translation/src/codecs/anthropic/buffered.rs b/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
index 9661c7674..82722c75d 100644
--- a/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
@@ -1195,10 +1195,12 @@ fn decode_anthropic_usage(value: Option<&Value>) -> Usage {
+ cache_creation_input_tokens.unwrap_or(0)
+ output
}),
- reasoning_tokens: value
- .get("output_tokens_details")
- .and_then(|details| details.get("reasoning_tokens"))
- .and_then(Value::as_u64),
+ reasoning_tokens: value.get("output_tokens_details").and_then(|details| {
+ details
+ .get("thinking_tokens")
+ .and_then(Value::as_u64)
+ .or_else(|| details.get("reasoning_tokens").and_then(Value::as_u64))
+ }),
}
}
@@ -1214,6 +1216,9 @@ fn encode_anthropic_usage(usage: &Usage) -> Value {
if let Some(cache_creation_tokens) = usage.cache_creation_input_tokens() {
value["cache_creation_input_tokens"] = json!(cache_creation_tokens);
}
+ if let Some(thinking_tokens) = usage.reasoning_tokens {
+ value["output_tokens_details"] = json!({"thinking_tokens": thinking_tokens});
+ }
value
}
diff --git a/crates/switchyard-translation/src/codecs/anthropic/stream.rs b/crates/switchyard-translation/src/codecs/anthropic/stream.rs
index 7b5663ce4..62dc0e368 100644
--- a/crates/switchyard-translation/src/codecs/anthropic/stream.rs
+++ b/crates/switchyard-translation/src/codecs/anthropic/stream.rs
@@ -574,6 +574,14 @@ fn capture_anthropic_usage(state: &mut StreamTranslationState, usage: &Value) {
if let Some(value) = usage.get("cache_read_input_tokens").and_then(Value::as_u64) {
state.usage.set_cached_input_tokens(value);
}
+ if let Some(value) = usage.get("output_tokens_details").and_then(|details| {
+ details
+ .get("thinking_tokens")
+ .and_then(Value::as_u64)
+ .or_else(|| details.get("reasoning_tokens").and_then(Value::as_u64))
+ }) {
+ state.usage.reasoning_tokens = Some(value);
+ }
}
// Builds Anthropic usage payloads from normalized and provider-extra state.
@@ -596,6 +604,12 @@ fn anthropic_stream_usage(state: &StreamTranslationState) -> Value {
if let Some(value) = state.usage.cached_input_tokens() {
usage.insert("cache_read_input_tokens".to_string(), json!(value));
}
+ if let Some(value) = state.usage.reasoning_tokens {
+ usage.insert(
+ "output_tokens_details".to_string(),
+ json!({"thinking_tokens": value}),
+ );
+ }
Value::Object(usage)
}
diff --git a/crates/switchyard-translation/tests/response_translation.rs b/crates/switchyard-translation/tests/response_translation.rs
index 29e81eabb..c4ad3e83d 100644
--- a/crates/switchyard-translation/tests/response_translation.rs
+++ b/crates/switchyard-translation/tests/response_translation.rs
@@ -371,6 +371,85 @@ fn anthropic_thinking_response_translates_to_openai_reasoning_content() -> TestR
let message = &output["choices"][0]["message"];
assert_eq!(message["content"], "Visible answer");
assert_eq!(message["reasoning_content"], "private reasoning");
+ for (details, thinking_tokens) in [
+ (json!({"thinking_tokens": 0}), 0),
+ (json!({"thinking_tokens": 5}), 5),
+ (json!({"thinking_tokens": null, "reasoning_tokens": 5}), 5),
+ ] {
+ let mut body = body.clone();
+ body["usage"]["output_tokens_details"] = details;
+ let decoded = engine
+ .decode_response(WireFormat::AnthropicMessages, &body, &normalized_policy())?
+ .response;
+ assert_eq!(decoded.usage.reasoning_tokens, Some(thinking_tokens));
+ assert_eq!(decoded.usage.output_tokens, Some(7));
+ assert_eq!(decoded.usage.total_tokens, Some(19));
+
+ for (target, details, field) in [
+ (
+ WireFormat::OpenAiChat,
+ "completion_tokens_details",
+ "reasoning_tokens",
+ ),
+ (
+ WireFormat::OpenAiResponses,
+ "output_tokens_details",
+ "reasoning_tokens",
+ ),
+ (
+ WireFormat::AnthropicMessages,
+ "output_tokens_details",
+ "thinking_tokens",
+ ),
+ ] {
+ let output = engine
+ .translate_response(
+ WireFormat::AnthropicMessages,
+ target,
+ &body,
+ &normalized_policy(),
+ )?
+ .body;
+ assert_eq!(output["usage"][details][field], thinking_tokens);
+ let mut state = switchyard_translation::StreamTranslationState::new(
+ WireFormat::AnthropicMessages,
+ target,
+ );
+ engine.translate_event(
+ &mut state,
+ WireFormat::AnthropicMessages,
+ target,
+ &json!({"type": "message_start", "message": {
+ "id": "msg_test", "model": "claude-opus", "usage": {"input_tokens": 12}
+ }}),
+ )?;
+ engine.translate_event(
+ &mut state,
+ WireFormat::AnthropicMessages,
+ target,
+ &json!({"type": "message_delta", "delta": {}, "usage": body["usage"]}),
+ )?;
+ let mut events = engine.translate_event(
+ &mut state,
+ WireFormat::AnthropicMessages,
+ target,
+ &json!({"type": "message_delta", "delta": {"stop_reason": "end_turn"},
+ "usage": {"output_tokens": 7}}),
+ )?;
+ assert_eq!(state.usage.reasoning_tokens, Some(thinking_tokens));
+ assert_eq!(state.usage.output_tokens, Some(7));
+ events.extend(engine.finish_stream(&mut state, target)?);
+ let usage = events
+ .iter()
+ .find_map(|event| {
+ event
+ .get("usage")
+ .or_else(|| event.get("response")?.get("usage"))
+ })
+ .ok_or("missing terminal usage")?;
+ assert_eq!(usage[details][field], thinking_tokens);
+ }
+ }
Ok(())
}