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): omit deprecated Anthropic temperature (#1305)

David Gardnerd333a047e5a9
diff --git a/crates/protocol/src/llm.rs b/crates/protocol/src/llm.rs
index 1e4e6e9c3..9713e920f 100644
--- a/crates/protocol/src/llm.rs
+++ b/crates/protocol/src/llm.rs
@@ -15,6 +15,9 @@ use serde_json::{Map, Value};
use crate::format::FormatId;
+/// Maximum temperature on the shared IR's zero-based scale.
+pub const MAX_IR_TEMPERATURE: f64 = 2.0;
+
/// Actor role normalized across provider APIs.
#[derive(Clone, Copy, Debug, Eq, PartialEq, Serialize, Deserialize, strum_macros::Display)]
#[serde(rename_all = "snake_case")]
@@ -457,7 +460,9 @@ pub enum ToolChoice {
/// Provider sampling parameters with common cross-provider names.
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
pub struct SamplingParams {
- /// Sampling temperature.
+ /// Sampling temperature on the `0.0..=MAX_IR_TEMPERATURE` scale.
+ /// Directly constructed requests use this scale; omission retains provider defaults.
+ /// See [`MAX_IR_TEMPERATURE`].
pub temperature: Option<f64>,
/// Nucleus-sampling probability mass.
pub top_p: Option<f64>,
diff --git a/crates/switchyard-translation/src/codecs/anthropic/buffered.rs b/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
index 07863e5fd..ce8b50066 100644
--- a/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
@@ -4,10 +4,12 @@
//! Buffered codec for Anthropic Messages request and response JSON.
use serde_json::{Map, Value, json};
+use switchyard_protocol::MAX_IR_TEMPERATURE;
+use super::{DEFAULT_TEMPERATURE, MAX_TEMPERATURE};
use crate::codecs::common::{
ANTHROPIC_REQUEST_KEY, is_anthropic_request, is_known_role_name, provider_extensions,
- text_from_blocks,
+ scale_temperature, text_from_blocks,
};
use crate::codecs::openai_chat::{decode_file_source, decode_image_source};
use crate::codecs::structured_output::decode_openai_schema_enforcement;
@@ -79,7 +81,10 @@ impl FormatCodec for AnthropicMessagesCodec {
response_format,
},
sampling: SamplingParams {
- temperature: body.get("temperature").and_then(Value::as_f64),
+ temperature: body
+ .get("temperature")
+ .and_then(Value::as_f64)
+ .map(|value| scale_temperature(value, MAX_TEMPERATURE, MAX_IR_TEMPERATURE)),
top_p: body.get("top_p").and_then(Value::as_f64),
top_k: body.get("top_k").and_then(Value::as_i64),
},
@@ -327,8 +332,19 @@ impl FormatCodec for AnthropicMessagesCodec {
} else {
body.insert("max_tokens".to_string(), json!(64_000));
}
- if let Some(value) = request.sampling.temperature {
- body.insert("temperature".to_string(), json!(value));
+ if let Some(value) = request
+ .sampling
+ .temperature
+ .map(|value| scale_temperature(value, MAX_IR_TEMPERATURE, MAX_TEMPERATURE))
+ && value != DEFAULT_TEMPERATURE
+ {
+ push_lossy(
+ &mut diagnostics,
+ policy,
+ format!(
+ "deprecated temperature parameter with non-default value {value} is omitted from Anthropic Messages requests"
+ ),
+ )?;
}
if let Some(value) = request.sampling.top_p {
body.insert("top_p".to_string(), json!(value));
diff --git a/crates/switchyard-translation/src/codecs/anthropic/mod.rs b/crates/switchyard-translation/src/codecs/anthropic/mod.rs
index d990cd988..5c0439485 100644
--- a/crates/switchyard-translation/src/codecs/anthropic/mod.rs
+++ b/crates/switchyard-translation/src/codecs/anthropic/mod.rs
@@ -8,3 +8,10 @@ mod stream;
pub use buffered::AnthropicMessagesCodec;
pub use stream::AnthropicMessagesStreamCodec;
+
+// Temperature range: 0.0..=1.0 for legacy models supporting this parameter.
+// https://platform.claude.com/docs/en/api/messages/create?version=2023-06-01
+// API version 2026-10-21 removes temperature; sending it returns HTTP 400.
+// https://platform.claude.com/docs/en/api/migration-2026-10-21#sampling-parameters-are-removed
+pub(crate) const MAX_TEMPERATURE: f64 = 1.0;
+pub(crate) const DEFAULT_TEMPERATURE: f64 = 1.0;
diff --git a/crates/switchyard-translation/src/codecs/common.rs b/crates/switchyard-translation/src/codecs/common.rs
index 303c90321..eb77bb23b 100644
--- a/crates/switchyard-translation/src/codecs/common.rs
+++ b/crates/switchyard-translation/src/codecs/common.rs
@@ -11,6 +11,14 @@ use crate::llm::{ContentBlock, LlmRequest, Tool, ToolChoice};
// Internal provenance survives mutations that invalidate exact request replay.
pub(crate) const ANTHROPIC_REQUEST_KEY: &str = "switchyard_anthropic_request";
+/// Scales temperature between ranges with a minimum of zero.
+pub(crate) fn scale_temperature(temperature: f64, source_max: f64, target_max: f64) -> f64 {
+ if source_max == target_max {
+ return temperature;
+ }
+ (temperature / source_max) * target_max
+}
+
/// Returns true when the request was decoded from an Anthropic Messages body.
pub(crate) fn is_anthropic_request(request: &LlmRequest) -> bool {
request.extensions.fields.get(ANTHROPIC_REQUEST_KEY) == Some(&Value::Bool(true))
diff --git a/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs b/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs
index 802a0dcf8..923bb2578 100644
--- a/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs
@@ -4,10 +4,12 @@
//! Buffered codec for OpenAI Chat Completions request and response JSON.
use serde_json::{Map, Value, json};
+use switchyard_protocol::MAX_IR_TEMPERATURE;
+use super::MAX_TEMPERATURE;
use crate::codecs::common::{
first_nonempty_string, is_known_role_name, provider_extensions, reasoning_text_from_blocks,
- reasoning_text_from_details, text_from_blocks,
+ reasoning_text_from_details, scale_temperature, text_from_blocks,
};
use crate::codecs::openai_media::{
ImagePayload, file_payload, file_source_text, image_payload, image_source_text,
@@ -50,7 +52,10 @@ impl FormatCodec for OpenAiChatCodec {
.map(ToOwned::to_owned),
stream: body.get("stream").and_then(Value::as_bool).unwrap_or(false),
sampling: SamplingParams {
- temperature: body.get("temperature").and_then(Value::as_f64),
+ temperature: body
+ .get("temperature")
+ .and_then(Value::as_f64)
+ .map(|value| scale_temperature(value, MAX_TEMPERATURE, MAX_IR_TEMPERATURE)),
top_p: body.get("top_p").and_then(Value::as_f64),
top_k: None,
},
@@ -245,6 +250,7 @@ impl FormatCodec for OpenAiChatCodec {
body.insert("max_completion_tokens".to_string(), json!(value));
}
if let Some(value) = request.sampling.temperature {
+ let value = scale_temperature(value, MAX_IR_TEMPERATURE, MAX_TEMPERATURE);
body.insert("temperature".to_string(), json!(value));
}
if let Some(value) = request.sampling.top_p {
diff --git a/crates/switchyard-translation/src/codecs/openai_chat/mod.rs b/crates/switchyard-translation/src/codecs/openai_chat/mod.rs
index 766f312c6..75aed3e95 100644
--- a/crates/switchyard-translation/src/codecs/openai_chat/mod.rs
+++ b/crates/switchyard-translation/src/codecs/openai_chat/mod.rs
@@ -10,3 +10,7 @@ pub use buffered::OpenAiChatCodec;
pub use stream::OpenAiChatStreamCodec;
pub(crate) use buffered::{decode_file_source, decode_image_source};
+
+// Temperature range: 0.0..=2.0.
+// https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create
+pub(crate) const MAX_TEMPERATURE: f64 = 2.0;
diff --git a/crates/switchyard-translation/src/codecs/responses/buffered.rs b/crates/switchyard-translation/src/codecs/responses/buffered.rs
index 8f7d47af9..d6f495bee 100644
--- a/crates/switchyard-translation/src/codecs/responses/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/responses/buffered.rs
@@ -6,11 +6,13 @@
use std::collections::HashMap;
use serde_json::{Map, Value, json};
+use switchyard_protocol::MAX_IR_TEMPERATURE;
+use super::MAX_TEMPERATURE;
use crate::codecs::common::{
collect_responses_reasoning_text, encrypted_reasoning_data, encrypted_reasoning_item_id,
is_anthropic_request, is_known_role_name, provider_extensions, reasoning_text_from_blocks,
- text_from_blocks,
+ scale_temperature, text_from_blocks,
};
use crate::codecs::openai_chat::{decode_file_source, decode_image_source};
use crate::codecs::openai_media::{
@@ -68,7 +70,10 @@ impl FormatCodec for OpenAiResponsesCodec {
raw: body.get("reasoning").cloned(),
},
sampling: SamplingParams {
- temperature: body.get("temperature").and_then(Value::as_f64),
+ temperature: body
+ .get("temperature")
+ .and_then(Value::as_f64)
+ .map(|value| scale_temperature(value, MAX_TEMPERATURE, MAX_IR_TEMPERATURE)),
top_p: body.get("top_p").and_then(Value::as_f64),
top_k: None,
},
@@ -288,6 +293,7 @@ impl FormatCodec for OpenAiResponsesCodec {
body.insert("reasoning".to_string(), Value::Object(reasoning));
}
if let Some(value) = request.sampling.temperature {
+ let value = scale_temperature(value, MAX_IR_TEMPERATURE, MAX_TEMPERATURE);
body.insert("temperature".to_string(), json!(value));
}
if let Some(value) = request.sampling.top_p {
diff --git a/crates/switchyard-translation/src/codecs/responses/mod.rs b/crates/switchyard-translation/src/codecs/responses/mod.rs
index 8f18e26fb..91985549e 100644
--- a/crates/switchyard-translation/src/codecs/responses/mod.rs
+++ b/crates/switchyard-translation/src/codecs/responses/mod.rs
@@ -9,6 +9,10 @@ mod stream;
pub use buffered::OpenAiResponsesCodec;
pub use stream::OpenAiResponsesStreamCodec;
+// Temperature range: 0.0..=2.0.
+// https://developers.openai.com/api/reference/resources/responses/methods/create
+pub(crate) const MAX_TEMPERATURE: f64 = 2.0;
+
pub(crate) fn is_native_output(kind: &str) -> bool {
matches!(
kind,
diff --git a/crates/switchyard-translation/tests/lossless_roundtrip.rs b/crates/switchyard-translation/tests/lossless_roundtrip.rs
index db99a71fd..b507c19b4 100644
--- a/crates/switchyard-translation/tests/lossless_roundtrip.rs
+++ b/crates/switchyard-translation/tests/lossless_roundtrip.rs
@@ -73,6 +73,18 @@ fn embedded_preservation_roundtrips_requests_for_every_source_target_pair_exactl
})
.body;
+ let expected_temperature = match (source, target) {
+ (WireFormat::AnthropicMessages, WireFormat::AnthropicMessages) => Some(0.2),
+ (WireFormat::AnthropicMessages, _) => Some(0.4),
+ (_, WireFormat::AnthropicMessages) => None,
+ _ => Some(0.2),
+ };
+ assert_eq!(
+ translated.get("temperature"),
+ expected_temperature.map(Value::from).as_ref(),
+ "request {source:?} -> {target:?} temperature"
+ );
+
if source == target {
assert_eq!(translated, original, "request {source:?} self-translation");
} else {
diff --git a/crates/switchyard-translation/tests/request_translation.rs b/crates/switchyard-translation/tests/request_translation.rs
index d025b4182..d1a4654ce 100644
--- a/crates/switchyard-translation/tests/request_translation.rs
+++ b/crates/switchyard-translation/tests/request_translation.rs
@@ -3869,25 +3869,56 @@ fn openai_tool_strictness_is_preserved_for_anthropic() -> TestResult {
Ok(())
}
-// Verifies Anthropic-bound OpenAI requests get the required max_tokens fallback.
+// Verifies Anthropic defaults and the policy for discarded non-default temperatures.
#[test]
-fn openai_request_to_anthropic_adds_required_default_max_tokens() -> TestResult {
+fn openai_request_to_anthropic_parameters() -> TestResult {
let engine = TranslationEngine::default();
let body = json!({
"model": "gpt-5",
"messages": [{"role": "user", "content": "hi"}]
});
- let output = engine
- .translate_request(
- WireFormat::OpenAiChat,
- WireFormat::AnthropicMessages,
- &body,
- &TranslationPolicy::default(),
- )?
- .body;
-
- assert_eq!(output["max_tokens"], 64_000);
+ for (temperature, is_lossy) in [
+ (None, false),
+ (Some(1.0), true),
+ (Some(0.5), true),
+ (Some(2.0), false),
+ ] {
+ let mut body = body.clone();
+ if let Some(temperature) = temperature {
+ body["temperature"] = json!(temperature);
+ }
+ for lossy_conversion_policy in [
+ LossyConversionPolicy::AllowWithDiagnostics,
+ LossyConversionPolicy::Reject,
+ ] {
+ let policy = TranslationPolicy {
+ lossy_conversion_policy,
+ ..normalized_policy()
+ };
+ let result = engine.translate_request(
+ WireFormat::OpenAiChat,
+ WireFormat::AnthropicMessages,
+ &body,
+ &policy,
+ );
+ if is_lossy && lossy_conversion_policy == LossyConversionPolicy::Reject {
+ assert!(
+ matches!(result, Err(TranslationError::LossyConversion(message))
+ if message.contains("temperature"))
+ );
+ continue;
+ }
+ let output = result?;
+ assert_eq!(output.body["max_tokens"], 64_000);
+ assert!(output.body.get("temperature").is_none());
+ assert_eq!(output.diagnostics.len(), usize::from(is_lossy));
+ if is_lossy {
+ assert_eq!(output.diagnostics[0].code, "lossy_conversion");
+ assert!(output.diagnostics[0].message.contains("temperature"));
+ }
+ }
+ }
Ok(())
}
diff --git a/docs/reference/protocols/llm_requests.md b/docs/reference/protocols/llm_requests.md
index d64ecb734..f2432d5d4 100644
--- a/docs/reference/protocols/llm_requests.md
+++ b/docs/reference/protocols/llm_requests.md
@@ -69,7 +69,7 @@ providers apply their own defaults and accepted ranges.
| Protocol field and type | Meaning | Anthropic Messages | OpenAI Responses | Chat Completions |
|---|---|---|---|---|
-| <a href="../../../rust/switchyard_protocol/llm/struct.SamplingParams.html#structfield.temperature"><code>SamplingParams.temperature: Option&lt;f64&gt;</code></a> | Sampling temperature. | `temperature` | `temperature` | `temperature` |
+| <a href="../../../rust/switchyard_protocol/llm/struct.SamplingParams.html#structfield.temperature"><code>SamplingParams.temperature: Option&lt;f64&gt;</code></a> | Sampling temperature. | (omitted; refer to the note below) | `temperature` | `temperature` |
| <a href="../../../rust/switchyard_protocol/llm/struct.SamplingParams.html#structfield.top_p"><code>SamplingParams.top_p: Option&lt;f64&gt;</code></a> | Nucleus-sampling probability mass. | `top_p` | `top_p` | `top_p` |
| <a href="../../../rust/switchyard_protocol/llm/struct.SamplingParams.html#structfield.top_k"><code>SamplingParams.top_k: Option&lt;i64&gt;</code></a> | Maximum candidate-token count. | `top_k` | No standard field | No standard field |
| <a href="../../../rust/switchyard_protocol/llm/struct.OutputParams.html#structfield.max_output_tokens"><code>OutputParams.max_output_tokens: Option&lt;u64&gt;</code></a> | Output token budget. | `max_tokens` | `max_output_tokens` | `max_completion_tokens`, or legacy `max_tokens` |
@@ -78,6 +78,32 @@ providers apply their own defaults and accepted ranges.
| <a href="../../../rust/switchyard_protocol/llm/struct.ReasoningParams.html#structfield.effort"><code>ReasoningParams.effort: Option&lt;String&gt;</code></a> | Requested reasoning effort. | `output_config.effort` | `reasoning.effort` | `reasoning_effort` |
| <a href="../../../rust/switchyard_protocol/llm/struct.ReasoningParams.html#structfield.raw"><code>ReasoningParams.raw: Option&lt;Value&gt;</code></a> | Provider-specific reasoning configuration. | `thinking` controls | `reasoning` object | No standard equivalent for the whole object |
+**Anthropic temperature:** Switchyard omits `temperature` when rebuilding Anthropic
+requests. Older models still support this parameter under API version `2023-06-01`,
+but models released after Claude Opus 4.6 reject values other than the default
+`1.0`. For model restrictions, refer to the [Anthropic parameter documentation](https://platform.claude.com/docs/en/api/messages/create?version=2023-06-01).
+API version `2026-10-21` removes the parameter entirely and rejects it even at its
+default value. For removal details, refer to [Anthropic's migration guide](https://platform.claude.com/docs/en/api/migration-2026-10-21#sampling-parameters-are-removed).
+Switchyard scales OpenAI temperatures from the `0.0–2.0` range to Anthropic's
+`0.0–1.0` range before comparing with the default. An OpenAI temperature of `2.0`
+scales to Anthropic's default of `1.0` and is omitted without diagnostics or rejection.
+Omitting a non-default value after scaling to Anthropic's range follows the
+[lossy-conversion policy](translation.md#policies). Exact preserved Anthropic
+requests can still retain `temperature`.
+
+For older Anthropic models that support non-default temperatures under API version
+`2023-06-01`, add these settings to the target's TOML configuration:
+
+```toml
+omit_body_fields = ["temperature"]
+extra_body = { temperature = 0.5 }
+```
+
+After translation, Switchyard removes any existing `temperature` and sets it to
+`0.5` in Anthropic's units, without further scaling. This also overrides temperature
+in preserved requests. Refer to the [target configuration reference](../toml_schema.md#targetsname)
+for these settings.
+
Preserved same-format bodies take precedence over edits to these settings,
including schema enforcement. See **[encoding an edited request](translation.md#preservation-and-edits)**.