Add typed support for the OpenAI Ultrafast service tier (#67143)

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
This commit is contained in:
Bennet Bo Fenner
2026-11-18 12:22:00 +00:00
committed by GitHub
parent da294762d6
commit f3626706fa
9 changed files with 16 additions and 8 deletions
+3 -1
View File
@@ -1378,7 +1378,9 @@ pub fn parse_prompt_too_long(message: &str) -> Option<u64> {
impl From<language_model_core::Speed> 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,
}
}
@@ -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,
+2 -1
View File
@@ -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,
}
}
}
@@ -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)
@@ -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(),
@@ -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,
+4 -6
View File
@@ -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<language_model_core::Speed>) -> Option<ServiceTier> {
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,
+1
View File
@@ -790,6 +790,7 @@ pub enum ServiceTier {
Scale,
#[serde(alias = "fast")]
Priority,
Ultrafast,
}
#[derive(Debug, Serialize, Deserialize)]
@@ -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,