From f3626706fa6855274a561f8f4aebf40666051596 Mon Sep 17 00:00:00 2001 From: Bennet Bo Fenner Date: Wed, 18 Nov 2026 12:22:00 +0000 Subject: [PATCH] Add typed support for the OpenAI Ultrafast service tier (#67143) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add an Ultrafast request speed and an independent model capability flag. Encode and decode OpenAI’s `ultrafast` service tier without changing Zed’s existing Fast toggle or enabling Ultrafast in Zed’s model catalog. Release Notes: - N/A --- crates/anthropic/src/anthropic.rs | 4 +++- crates/language_model/src/language_model.rs | 2 ++ crates/language_model_core/src/request.rs | 3 ++- crates/language_models/src/provider/anthropic.rs | 1 + crates/language_models/src/provider/open_ai.rs | 1 + .../language_models_cloud/src/language_models_cloud.rs | 1 + crates/open_ai/src/completion.rs | 10 ++++------ crates/open_ai/src/open_ai.rs | 1 + crates/openai_subscribed/src/openai_subscribed.rs | 1 + 9 files changed, 16 insertions(+), 8 deletions(-) diff --git a/crates/anthropic/src/anthropic.rs b/crates/anthropic/src/anthropic.rs index b457f947252..555266e70a0 100644 --- a/crates/anthropic/src/anthropic.rs +++ b/crates/anthropic/src/anthropic.rs @@ -1378,7 +1378,9 @@ pub fn parse_prompt_too_long(message: &str) -> Option { impl From for Speed { fn from(speed: language_model_core::Speed) -> Self { match speed { - language_model_core::Speed::Standard => Speed::Standard, + language_model_core::Speed::Standard | language_model_core::Speed::Ultrafast => { + Speed::Standard + } language_model_core::Speed::Fast => Speed::Fast, } } diff --git a/crates/language_model/src/language_model.rs b/crates/language_model/src/language_model.rs index 61e7259a51e..69c07d62bd6 100644 --- a/crates/language_model/src/language_model.rs +++ b/crates/language_model/src/language_model.rs @@ -129,6 +129,7 @@ pub struct LanguageModel { /// request. Only meaningful when `supports_thinking` is `true`. pub supports_disabling_thinking: bool, pub supports_fast_mode: bool, + pub supports_ultrafast_mode: bool, /// Whether instruction updates can retain their positions in conversation history. pub supports_system_prompt_updates: bool, /// The effort levels that can be used when thinking. @@ -222,6 +223,7 @@ impl LanguageModel { supports_thinking: false, supports_disabling_thinking: true, supports_fast_mode: false, + supports_ultrafast_mode: false, supports_system_prompt_updates: false, supported_effort_levels: Arc::default(), supports_server_side_compaction: false, diff --git a/crates/language_model_core/src/request.rs b/crates/language_model_core/src/request.rs index fa7d4dbb423..1e9c41cfeaa 100644 --- a/crates/language_model_core/src/request.rs +++ b/crates/language_model_core/src/request.rs @@ -533,13 +533,14 @@ pub enum Speed { #[default] Standard, Fast, + Ultrafast, } impl Speed { pub fn toggle(self) -> Self { match self { Speed::Standard => Speed::Fast, - Speed::Fast => Speed::Standard, + Speed::Fast | Speed::Ultrafast => Speed::Standard, } } } diff --git a/crates/language_models/src/provider/anthropic.rs b/crates/language_models/src/provider/anthropic.rs index 01249518357..0aa7c6895d7 100644 --- a/crates/language_models/src/provider/anthropic.rs +++ b/crates/language_models/src/provider/anthropic.rs @@ -1019,6 +1019,7 @@ fn language_model(model: &anthropic::Model) -> LanguageModel { }, supports_thinking: model.supports_thinking, supports_fast_mode: model.supports_speed, + supports_ultrafast_mode: false, refusal_fallback_model_id: model .id .starts_with(anthropic::FABLE_MODEL_ID_PREFIX) diff --git a/crates/language_models/src/provider/open_ai.rs b/crates/language_models/src/provider/open_ai.rs index bfb83175d72..cfd358cc7f3 100644 --- a/crates/language_models/src/provider/open_ai.rs +++ b/crates/language_models/src/provider/open_ai.rs @@ -713,6 +713,7 @@ fn language_model(model: &open_ai::Model) -> LanguageModel { supports_streaming_tools: true, supports_thinking: supports_selectable_thinking_effort(model), supports_fast_mode: model.supports_priority(), + supports_ultrafast_mode: false, supports_server_side_compaction: model.supports_compaction(), supports_explicit_compaction: model.supports_compaction(), supported_effort_levels: supported_thinking_effort_levels(model).into(), diff --git a/crates/language_models_cloud/src/language_models_cloud.rs b/crates/language_models_cloud/src/language_models_cloud.rs index e053909f50b..a5cd0ddff0a 100644 --- a/crates/language_models_cloud/src/language_models_cloud.rs +++ b/crates/language_models_cloud/src/language_models_cloud.rs @@ -529,6 +529,7 @@ pub fn language_model(model: &cloud_llm_client::LanguageModel) -> LanguageModel supports_thinking: model.supports_thinking, supports_disabling_thinking: model.supports_disabling_thinking, supports_fast_mode: model.supports_fast_mode, + supports_ultrafast_mode: false, supports_system_prompt_updates: match model.provider { Anthropic => anthropic::supports_mid_conversation_system_messages(&model.id.0), OpenAi => true, diff --git a/crates/open_ai/src/completion.rs b/crates/open_ai/src/completion.rs index 367ffe53701..579f392e776 100644 --- a/crates/open_ai/src/completion.rs +++ b/crates/open_ai/src/completion.rs @@ -32,11 +32,11 @@ const RESPONSE_MESSAGE_PHASE_COMMENTARY: &str = "commentary"; const RESPONSE_MESSAGE_PHASE_FINAL_ANSWER: &str = "final_answer"; /// Translates the request's `Speed` into the corresponding OpenAI service tier. -/// Only `Fast` produces a value; `Standard` leaves the field unset so that the -/// project's default tier applies. +/// `Standard` leaves the field unset so that the project's default tier applies. fn service_tier_for(speed: Option) -> Option { match speed? { language_model_core::Speed::Fast => Some(ServiceTier::Priority), + language_model_core::Speed::Ultrafast => Some(ServiceTier::Ultrafast), language_model_core::Speed::Standard => None, } } @@ -2455,15 +2455,13 @@ mod tests { Ok(()) } - /// `Speed::Fast` should translate to `service_tier: "priority"` on the - /// outgoing Responses request, while `Standard` / `None` should leave the - /// field unset so the project's default tier wins. #[test] - fn into_open_ai_response_sets_service_tier_for_fast_speed() -> Result<()> { + fn into_open_ai_response_sets_service_tier_for_speed() -> Result<()> { for (speed, expected) in [ (None, None), (Some(Speed::Standard), None), (Some(Speed::Fast), Some("priority")), + (Some(Speed::Ultrafast), Some("ultrafast")), ] { let request = LanguageModelRequest { thread_id: None, diff --git a/crates/open_ai/src/open_ai.rs b/crates/open_ai/src/open_ai.rs index da7249a11e9..c8300c5f1ce 100644 --- a/crates/open_ai/src/open_ai.rs +++ b/crates/open_ai/src/open_ai.rs @@ -790,6 +790,7 @@ pub enum ServiceTier { Scale, #[serde(alias = "fast")] Priority, + Ultrafast, } #[derive(Debug, Serialize, Deserialize)] diff --git a/crates/openai_subscribed/src/openai_subscribed.rs b/crates/openai_subscribed/src/openai_subscribed.rs index 3b84c7f6661..5fd4a86817b 100644 --- a/crates/openai_subscribed/src/openai_subscribed.rs +++ b/crates/openai_subscribed/src/openai_subscribed.rs @@ -758,6 +758,7 @@ pub fn language_model(model: &ChatGptModel) -> LanguageModel { supports_streaming_tools: true, supports_thinking: true, supports_fast_mode: model.supports_priority(), + supports_ultrafast_mode: false, supports_server_side_compaction: true, supports_explicit_compaction: true, supported_effort_levels,