fix(gateway): drop pinned temperature when thinking stays on and unsupported effort under drop_params

litellm-rust/crates/llms/src/anthropic/messages/thinking.rs

@@ -306,20 +306,53 @@ fn translate_adaptive_effort_for_non_adaptive_model(
     })
 }
 
+pub(crate) fn drop_unsupported_effort(
+    request: MessagesRequest,
+    capabilities: &MessagesModelCapabilities,
+) -> MessagesRequest {
+    let rejected = |level: EffortLevel| !accepts_effort(capabilities, level);
+    let drops_reasoning_effort = capabilities.supports_adaptive_thinking
+        && matches!(
+            &request.params.reasoning_effort,
+            Some(Recognized::Known(effort)) if output_effort(*effort).is_some_and(rejected)
+        );
+    let reshaped_to_budget = !capabilities.supports_adaptive_thinking
+        && matches!(known_thinking(&request), Some(ThinkingConfig::Adaptive(_)));
+    let drops_output_effort = supports_effort_param(capabilities)
+        && !reshaped_to_budget
+        && matches!(known_effort(&request), Some(Recognized::Known(level)) if rejected(*level));
+    MessagesRequest {
+        params: MessagesOptionalParams {
+            reasoning_effort: request
+                .params
+                .reasoning_effort
+                .filter(|_| !drops_reasoning_effort),
+            output_config: if drops_output_effort {
+                without_effort(request.params.output_config)
+            } else {
+                request.params.output_config
+            },
+            ..request.params
+        },
+        ..request
+    }
+}
+
 fn drop_incompatible_temperature_for_thinking(
     request: MessagesRequest,
     context: &ThinkingContext,
 ) -> MessagesRequest {
-    if context.capabilities.supports_adaptive_thinking {
-        return request;
-    }
     let pinned = request
         .params
         .temperature
         .is_some_and(|temperature| temperature != 1.0);
-    let thinking_enabled = matches!(known_thinking(&request), Some(ThinkingConfig::Enabled(_)));
-    let effort_enabled = known_effort(&request).is_some();
-    if !pinned || !(thinking_enabled || effort_enabled) {
+    let thinking_on = matches!(
+        known_thinking(&request),
+        Some(ThinkingConfig::Enabled(_) | ThinkingConfig::Adaptive(_))
+    );
+    let effort_on_budget_model =
+        !context.capabilities.supports_adaptive_thinking && known_effort(&request).is_some();
+    if !pinned || !(thinking_on || effort_on_budget_model) {
         return request;
     }
     MessagesRequest {
@@ -1092,6 +1125,47 @@ mod tests {
         );
     }
 
+    #[rstest]
+    #[case::claude_code_payload(
+        opus_4_7(),
+        claude_code_payload("medium", 8192),
+        0.0,
+        claude_code_payload("medium", 8192)
+    )]
+    #[case::legacy_thinking_kept_on_legacy_capable_model(
+        sonnet_4_6(),
+        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "enabled", "budget_tokens": 4096}}),
+        0.5,
+        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "enabled", "budget_tokens": 4096}})
+    )]
+    #[case::bare_adaptive_thinking(
+        sonnet_4_6(),
+        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "adaptive"}}),
+        0.5,
+        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "adaptive"}})
+    )]
+    #[case::reasoning_effort_synthesized_adaptive_thinking(
+        sonnet_4_6(),
+        serde_json::json!({"max_tokens": 8192, "reasoning_effort": "high"}),
+        0.0,
+        serde_json::json!({
+            "max_tokens": 8192,
+            "thinking": {"type": "adaptive", "display": "summarized"},
+            "output_config": {"effort": "high"}
+        })
+    )]
+    fn pinned_temperature_is_dropped_when_thinking_survives_on_adaptive_model(
+        #[case] capabilities: MessagesModelCapabilities,
+        #[case] input: Value,
+        #[case] temperature: f64,
+        #[case] expected: Value,
+    ) {
+        assert_eq!(
+            translate(capabilities, with_temperature(input, temperature)),
+            Ok(request(expected))
+        );
+    }
+
     #[rstest]
     #[case::temperature_one_with_enabled_thinking(
         haiku_4_5(),
@@ -1124,17 +1198,17 @@ mod tests {
         0.0,
         serde_json::json!({"max_tokens": 8192, "output_config": {"format": {"type": "json_schema"}}})
     )]
-    #[case::adaptive_model(
-        opus_4_7(),
-        claude_code_payload("medium", 8192),
-        0.0,
-        claude_code_payload("medium", 8192)
-    )]
-    #[case::legacy_thinking_on_adaptive_model(
+    #[case::effort_without_thinking_on_adaptive_model(
         sonnet_4_6(),
-        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "enabled", "budget_tokens": 4096}}),
-        0.0,
-        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "enabled", "budget_tokens": 4096}})
+        serde_json::json!({"max_tokens": 8192, "output_config": {"effort": "high"}}),
+        0.5,
+        serde_json::json!({"max_tokens": 8192, "output_config": {"effort": "high"}})
+    )]
+    #[case::disabled_thinking_on_adaptive_model(
+        sonnet_4_6(),
+        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "disabled"}}),
+        0.5,
+        serde_json::json!({"max_tokens": 8192, "thinking": {"type": "disabled"}})
     )]
     fn temperature_is_kept(
         #[case] capabilities: MessagesModelCapabilities,

litellm-rust/crates/llms/src/anthropic/messages/transformation.rs

@@ -9,7 +9,10 @@ use litellm_llms_types::{
 use litellm_router_types::LitellmParams;
 use serde_json::{Map, Value, json};
 
-use super::{handler::shape_anthropic_messages_request, thinking::translate_thinking};
+use super::{
+    handler::shape_anthropic_messages_request,
+    thinking::{drop_unsupported_effort, translate_thinking},
+};
 use crate::base_llm::messages::context::MessagesTransformContext;
 use crate::{
     Error,
@@ -267,6 +270,11 @@ fn drop_unsupported_params(
     context: &MessagesTransformContext,
 ) -> Result<MessagesRequest, Error> {
     let capabilities = &context.thinking.capabilities;
+    let request = if context.drop_params {
+        drop_unsupported_effort(request, capabilities)
+    } else {
+        request
+    };
     let model = request.model.clone();
     let reject = |param: &str, value: String, hint: &str| -> Result<(), Error> {
         if context.drop_params {
@@ -380,6 +388,7 @@ mod tests {
 
     use super::*;
     use crate::anthropic::common_utils::ENCRYPTED_REASONING_SIGNATURE_PREFIX;
+    use crate::base_llm::messages::context::SupportedEffortTiers;
 
     type Env = &'static [(&'static str, &'static str)];
 
@@ -481,6 +490,30 @@ mod tests {
         }
     }
 
+    #[fixture]
+    fn adaptive_without_xhigh() -> MessagesModelCapabilities {
+        MessagesModelCapabilities {
+            supports_reasoning: true,
+            supports_adaptive_thinking: true,
+            supports_legacy_thinking: true,
+            supports_output_config: true,
+            effort_tiers: SupportedEffortTiers {
+                max: true,
+                ..Default::default()
+            },
+            ..Default::default()
+        }
+    }
+
+    #[fixture]
+    fn effort_without_adaptive() -> MessagesModelCapabilities {
+        MessagesModelCapabilities {
+            supports_reasoning: true,
+            supports_output_config: true,
+            ..Default::default()
+        }
+    }
+
     #[rstest]
     #[case::alone(json!({"max_tokens": null}))]
     #[case::ahead_of_the_param_gate(json!({"max_tokens": null, "speed": "fast"}))]
@@ -514,6 +547,16 @@ mod tests {
     )]
     #[case::speed_on_a_fast_mode_model(fast_mode(), false, json!({"speed": "fast"}))]
     #[case::speed_on_a_fast_mode_model_under_drop_params(fast_mode(), true, json!({"speed": "fast"}))]
+    #[case::effort_tier_the_model_lacks_is_left_to_the_provider(
+        adaptive_without_xhigh(),
+        false,
+        json!({"output_config": {"effort": "xhigh"}})
+    )]
+    #[case::effort_tier_the_model_has_under_drop_params(
+        adaptive_without_xhigh(),
+        true,
+        json!({"output_config": {"effort": "max"}})
+    )]
     #[case::native_context_management_edits(unmapped(), false, json!({"context_management": {"edits": [{
         "type": "clear_tool_uses_20250919",
         "trigger": {"type": "input_tokens", "value": 30000},
@@ -568,6 +611,21 @@ mod tests {
         json!({"speed": "fast", "temperature": 1.0}),
         json!({"temperature": 1.0})
     )]
+    #[case::reasoning_effort_tier_the_adaptive_model_lacks(
+        adaptive_without_xhigh(),
+        json!({"reasoning_effort": "xhigh"}),
+        json!({})
+    )]
+    #[case::output_config_effort_tier_the_adaptive_model_lacks(
+        adaptive_without_xhigh(),
+        json!({"output_config": {"effort": "xhigh", "format": {"type": "json_schema"}}}),
+        json!({"output_config": {"format": {"type": "json_schema"}}})
+    )]
+    #[case::output_config_effort_tier_the_effort_model_lacks(
+        effort_without_adaptive(),
+        json!({"output_config": {"effort": "max"}}),
+        json!({})
+    )]
     fn removed_params_are_dropped_under_drop_params(
         #[case] capabilities: MessagesModelCapabilities,
         #[case] fields: Value,
@@ -576,6 +634,25 @@ mod tests {
         assert_eq!(transform(fields, capabilities, true), Ok(body(expected)));
     }
 
+    #[rstest]
+    fn drop_params_keeps_the_effort_sized_budget_on_a_model_without_adaptive_thinking(
+        effort_without_adaptive: MessagesModelCapabilities,
+    ) {
+        let adaptive = json!({"max_tokens": 64000, "thinking": {"type": "adaptive"}});
+        let with_xhigh = merged(
+            adaptive.clone(),
+            json!({"output_config": {"effort": "xhigh"}}),
+        );
+
+        let dropped = transform(with_xhigh.clone(), effort_without_adaptive, true);
+
+        assert_eq!(
+            dropped,
+            transform(with_xhigh, effort_without_adaptive, false)
+        );
+        assert_ne!(dropped, transform(adaptive, effort_without_adaptive, true));
+    }
+
     #[rstest]
     #[case::temperature(
         sampling_removed(),

litellm-rust/crates/llms/src/azure_ai/messages/transformation.rs

@@ -363,7 +363,7 @@ mod tests {
             "metadata": {"user_id": "u1"},
             "stop_sequences": ["STOP"],
             "stream": false,
-            "temperature": 0.4,
+            "temperature": 1.0,
             "top_p": 0.9,
             "top_k": 40,
             "tools": [{"name": "get_weather", "input_schema": {"type": "object"}}],