fix: azure voice live endpoint pricing

core/providers/azure/realtime.go

@@ -59,6 +59,10 @@ func (provider *AzureProvider) RealtimeWebSocketURL(ctx *schemas.BifrostContext,
 		if intent != "" {
 			return "", newAzureRealtimeError(fasthttp.StatusBadRequest, "invalid_request_error", "azure voice live does not take a realtime intent", nil)
 		}
+		// Voice Live turns are priced at the model's _voice_live rates.
+		if ctx != nil {
+			ctx.SetValue(schemas.BifrostContextKeyRealtimePricingSurface, schemas.PricingSurfaceVoiceLive)
+		}
 		return fmt.Sprintf("%s/voice-live/realtime?api-version=%s&model=%s", endpoint, url.QueryEscape(resolveAPIVersion(ctx, azureVoiceLiveAPIVersion)), url.QueryEscape(model)), nil
 	}
 

core/providers/azure/realtime_test.go

@@ -283,6 +283,31 @@ func TestAzureVoiceLiveWebSocketURL(t *testing.T) {
 	}
 }
 
+// TestAzureVoiceLiveMarksPricingSurface pins that a session sent to Voice Live is marked for the
+// model's _voice_live rates, and one sent to Azure OpenAI is not.
+func TestAzureVoiceLiveMarksPricingSurface(t *testing.T) {
+	t.Parallel()
+	provider := &AzureProvider{}
+	for _, tc := range []struct {
+		name      string
+		voiceLive bool
+		want      schemas.PricingSurface
+	}{
+		{"voice live", true, schemas.PricingSurfaceVoiceLive},
+		{"azure openai", false, ""},
+	} {
+		ctx := schemas.NewBifrostContext(context.Background(), schemas.NoDeadline)
+		key := schemas.Key{AzureKeyConfig: &schemas.AzureKeyConfig{Endpoint: *schemas.NewSecretVar("https://example.openai.azure.com/"), UseVoiceLiveEndpoints: tc.voiceLive}}
+		if _, err := provider.RealtimeWebSocketURL(ctx, key, "gpt-realtime", ""); err != nil {
+			t.Fatalf("%s: RealtimeWebSocketURL: %v", tc.name, err.Error.Message)
+		}
+		got, _ := ctx.Value(schemas.BifrostContextKeyRealtimePricingSurface).(schemas.PricingSurface)
+		if got != tc.want {
+			t.Errorf("%s: pricing surface = %q, want %q", tc.name, got, tc.want)
+		}
+	}
+}
+
 // TestAzureVoiceLiveRejectsWebRTCAndClientSecrets pins the 400 on the routes Voice Live does not
 // serve the OpenAI way, before any upstream call.
 func TestAzureVoiceLiveRejectsWebRTCAndClientSecrets(t *testing.T) {

core/providers/openai/realtime.go

@@ -1094,3 +1094,16 @@ func ExtractNestedVoice(audioRaw json.RawMessage) string {
 	}
 	return ""
 }
+
+// ExtractNestedVoiceType returns the type of the session.audio.output.voice object, or empty.
+func ExtractNestedVoiceType(audioRaw json.RawMessage) string {
+	var audio struct {
+		Output struct {
+			Voice schemas.RealtimeVoice `json:"voice"`
+		} `json:"output"`
+	}
+	if err := json.Unmarshal(audioRaw, &audio); err == nil {
+		return audio.Output.Voice.GetType()
+	}
+	return ""
+}

core/schemas/bifrost.go

@@ -398,6 +398,7 @@ const (
 	BifrostContextKeyA2ARequestOrigin                    BifrostContextKey = "bifrost-a2a-request-origin"                       // string (absolute HTTP origin used to reach the Agent Gateway)
 	BifrostContextKeyA2AUpstreamTransport                BifrostContextKey = "bifrost-a2a-upstream-transport"                   // string (A2A binding selected for the upstream agent)
 	BifrostContextKeyRealtimeVoice                       BifrostContextKey = "bifrost-realtime-voice"                           // string
+	BifrostContextKeyRealtimePricingSurface              BifrostContextKey = "bifrost-realtime-pricing-surface"                 // PricingSurface (set by the provider when it opens the realtime session)
 	BifrostIsAsyncRequest                                BifrostContextKey = "bifrost-is-async-request"                         // bool (set by bifrost - DO NOT SET THIS MANUALLY)) - whether the request is an async request (only used in gateway)
 	BifrostContextKeyRequestHeaders                      BifrostContextKey = "bifrost-request-headers"                          // map[string]string (all request headers with lowercased keys)
 	BifrostContextKeyRequestQuery                        BifrostContextKey = "bifrost-request-query"                            // map[string]string (request query params with lowercased keys; consumed by governance routing CEL rules)
@@ -2033,13 +2034,22 @@ func (r *BifrostMCPResponse) PopulateExtraFields(mcpRequestType MCPRequestType,
 	}
 }
 
+// PricingSurface names the product surface whose price list applies to a request.
+type PricingSurface string
+
+// PricingSurfaceVoiceLive prices an Azure Voice Live turn at the model's _voice_live rates.
+const PricingSurfaceVoiceLive PricingSurface = "voice_live"
+
 // BifrostResponseExtraFields contains additional fields in a response.
 type BifrostResponseExtraFields struct {
 	RequestType RequestType `json:"request_type"`
 	// PricingRequestType selects a catalog mode without changing the request type
 	// exposed to plugins and logs.
 	PricingRequestType RequestType `json:"pricing_request_type,omitempty"`
 	RoutingInfo        RoutingInfo `json:"routing_info"`
+	// PricingSurface selects a model's rates for the product that served the request, without
+	// changing the request type or service tier exposed to plugins and logs.
+	PricingSurface PricingSurface `json:"pricing_surface,omitempty"`
 	// Deprecated: use RoutingInfo.Provider. Still populated for backward
 	// compatibility; new consumers should read from RoutingInfo.
 	Provider ModelProvider `json:"provider,omitempty"`

core/schemas/chatcompletions.go

@@ -2022,6 +2022,10 @@ type ChatPromptTokensDetails struct {
 	CachedReadTokens        int                          `json:"cached_read_tokens,omitempty"`
 	CachedWriteTokens       int                          `json:"cached_write_tokens,omitempty"`
 	CachedWriteTokenDetails *ChatCachedWriteTokenDetails `json:"cached_write_token_details,omitempty"`
+
+	// Parts of AudioTokens that Azure Voice Live handled with Azure Speech, which it bills at their own rates.
+	StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens,omitempty"`
+	CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens,omitempty"`
 }
 
 type ChatCachedWriteTokenDetails struct {
@@ -2040,6 +2044,9 @@ func (d *ChatPromptTokensDetails) UnmarshalJSON(data []byte) error {
 		CachedWriteTokenDetails *ChatCachedWriteTokenDetails `json:"cached_write_token_details"`
 		CachedTokens            *int                         `json:"cached_tokens"`
 		CacheWriteTokens        *int                         `json:"cache_write_tokens"`
+
+		StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens"`
+		CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens"`
 	}
 	if err := Unmarshal(data, &raw); err != nil {
 		return err
@@ -2050,6 +2057,8 @@ func (d *ChatPromptTokensDetails) UnmarshalJSON(data []byte) error {
 	d.CachedReadTokens = raw.CachedReadTokens
 	d.CachedWriteTokens = raw.CachedWriteTokens
 	d.CachedWriteTokenDetails = raw.CachedWriteTokenDetails
+	d.StandardSpeechAudioTokens = raw.StandardSpeechAudioTokens
+	d.CustomSpeechAudioTokens = raw.CustomSpeechAudioTokens
 	// OpenAI spec providers send just cached_tokens, not separate read and write tokens and we handle them as read tokens in pricing calculations.
 	if raw.CachedTokens != nil && raw.CachedReadTokens == 0 && raw.CachedWriteTokens == 0 {
 		d.CachedReadTokens = *raw.CachedTokens
@@ -2076,6 +2085,9 @@ func (d ChatPromptTokensDetails) MarshalJSON() ([]byte, error) {
 		// OpenAI's field name for cache writes (mirrors cached_tokens for reads) so the
 		// OpenAI SDK — which reads cache_write_tokens, not cached_write_tokens — finds it.
 		CacheWriteTokens int `json:"cache_write_tokens"`
+
+		StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens,omitempty"`
+		CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens,omitempty"`
 	}
 	return MarshalSorted(raw{
 		TextTokens:              d.TextTokens,
@@ -2086,6 +2098,9 @@ func (d ChatPromptTokensDetails) MarshalJSON() ([]byte, error) {
 		CachedWriteTokenDetails: d.CachedWriteTokenDetails,
 		CachedTokens:            d.CachedReadTokens,
 		CacheWriteTokens:        d.CachedWriteTokens,
+
+		StandardSpeechAudioTokens: d.StandardSpeechAudioTokens,
+		CustomSpeechAudioTokens:   d.CustomSpeechAudioTokens,
 	})
 }
 
@@ -2099,6 +2114,10 @@ type ChatCompletionTokensDetails struct {
 	ReasoningTokens          int  `json:"reasoning_tokens,omitempty"`
 	ImageTokens              *int `json:"image_tokens,omitempty"`
 	RejectedPredictionTokens int  `json:"rejected_prediction_tokens,omitempty"`
+
+	// Parts of AudioTokens that Azure Voice Live handled with Azure Speech, which it bills at their own rates.
+	StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens,omitempty"`
+	CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens,omitempty"`
 }
 
 // BifrostCost splits a request's cost into an input side, an output side, and an
@@ -2151,6 +2170,9 @@ func MergeBifrostLLMUsage(base, add *BifrostLLMUsage) *BifrostLLMUsage {
 			ImageTokens:       baseDetails.ImageTokens + addDetails.ImageTokens,
 			CachedReadTokens:  baseDetails.CachedReadTokens + addDetails.CachedReadTokens,
 			CachedWriteTokens: baseDetails.CachedWriteTokens + addDetails.CachedWriteTokens,
+
+			StandardSpeechAudioTokens: baseDetails.StandardSpeechAudioTokens + addDetails.StandardSpeechAudioTokens,
+			CustomSpeechAudioTokens:   baseDetails.CustomSpeechAudioTokens + addDetails.CustomSpeechAudioTokens,
 		}
 		if baseDetails.CachedWriteTokenDetails != nil || addDetails.CachedWriteTokenDetails != nil {
 			merged.PromptTokensDetails.CachedWriteTokenDetails = &ChatCachedWriteTokenDetails{
@@ -2175,6 +2197,9 @@ func MergeBifrostLLMUsage(base, add *BifrostLLMUsage) *BifrostLLMUsage {
 			AudioTokens:              baseDetails.AudioTokens + addDetails.AudioTokens,
 			ReasoningTokens:          baseDetails.ReasoningTokens + addDetails.ReasoningTokens,
 			RejectedPredictionTokens: baseDetails.RejectedPredictionTokens + addDetails.RejectedPredictionTokens,
+
+			StandardSpeechAudioTokens: baseDetails.StandardSpeechAudioTokens + addDetails.StandardSpeechAudioTokens,
+			CustomSpeechAudioTokens:   baseDetails.CustomSpeechAudioTokens + addDetails.CustomSpeechAudioTokens,
 		}
 		merged.CompletionTokensDetails.CitationTokens = sumOptionalInts(baseDetails.CitationTokens, addDetails.CitationTokens)
 		merged.CompletionTokensDetails.NumSearchQueries = sumOptionalInts(baseDetails.NumSearchQueries, addDetails.NumSearchQueries)

core/schemas/mux.go

@@ -1135,6 +1135,9 @@ func (cu *BifrostLLMUsage) ToResponsesResponseUsage() *ResponsesResponseUsage {
 			CachedReadTokens:        cu.PromptTokensDetails.CachedReadTokens,
 			CachedWriteTokens:       cu.PromptTokensDetails.CachedWriteTokens,
 			CachedWriteTokenDetails: cu.PromptTokensDetails.CachedWriteTokenDetails,
+
+			StandardSpeechAudioTokens: cu.PromptTokensDetails.StandardSpeechAudioTokens,
+			CustomSpeechAudioTokens:   cu.PromptTokensDetails.CustomSpeechAudioTokens,
 		}
 	}
 	if cu.CompletionTokensDetails != nil {
@@ -1146,6 +1149,9 @@ func (cu *BifrostLLMUsage) ToResponsesResponseUsage() *ResponsesResponseUsage {
 			RejectedPredictionTokens: cu.CompletionTokensDetails.RejectedPredictionTokens,
 			CitationTokens:           cu.CompletionTokensDetails.CitationTokens,
 			NumSearchQueries:         cu.CompletionTokensDetails.NumSearchQueries,
+
+			StandardSpeechAudioTokens: cu.CompletionTokensDetails.StandardSpeechAudioTokens,
+			CustomSpeechAudioTokens:   cu.CompletionTokensDetails.CustomSpeechAudioTokens,
 		}
 	}
 	usage.ToolUsage = cu.ToolUsage.DeepCopy()
@@ -1175,6 +1181,9 @@ func (ru *ResponsesResponseUsage) ToBifrostLLMUsage() *BifrostLLMUsage {
 			CachedReadTokens:        ru.InputTokensDetails.CachedReadTokens,
 			CachedWriteTokens:       ru.InputTokensDetails.CachedWriteTokens,
 			CachedWriteTokenDetails: ru.InputTokensDetails.CachedWriteTokenDetails,
+
+			StandardSpeechAudioTokens: ru.InputTokensDetails.StandardSpeechAudioTokens,
+			CustomSpeechAudioTokens:   ru.InputTokensDetails.CustomSpeechAudioTokens,
 		}
 	}
 	if ru.OutputTokensDetails != nil {
@@ -1187,6 +1196,9 @@ func (ru *ResponsesResponseUsage) ToBifrostLLMUsage() *BifrostLLMUsage {
 			RejectedPredictionTokens: ru.OutputTokensDetails.RejectedPredictionTokens,
 			CitationTokens:           ru.OutputTokensDetails.CitationTokens,
 			NumSearchQueries:         ru.OutputTokensDetails.NumSearchQueries,
+
+			StandardSpeechAudioTokens: ru.OutputTokensDetails.StandardSpeechAudioTokens,
+			CustomSpeechAudioTokens:   ru.OutputTokensDetails.CustomSpeechAudioTokens,
 		}
 	}
 	usage.ToolUsage = ru.ToolUsage.DeepCopy()

core/schemas/realtime.go

@@ -203,6 +203,14 @@ func (v *RealtimeVoice) GetName() string {
 	return ""
 }
 
+// GetType returns the object voice's type, such as Azure Voice Live's "azure-standard", or empty.
+func (v *RealtimeVoice) GetType() string {
+	if v == nil || v.VoiceObject == nil {
+		return ""
+	}
+	return v.VoiceObject.Type
+}
+
 // MarshalJSON writes id, name and type next to the preserved fields.
 func (o RealtimeVoiceObject) MarshalJSON() ([]byte, error) {
 	out := make(map[string]json.RawMessage, len(o.ExtraParams)+3)

core/schemas/responses.go

@@ -1389,6 +1389,10 @@ type ResponsesResponseInputTokens struct {
 	CachedReadTokens        int                          `json:"cached_read_tokens"`
 	CachedWriteTokens       int                          `json:"cached_write_tokens"`
 	CachedWriteTokenDetails *ChatCachedWriteTokenDetails `json:"cached_write_token_details,omitempty"`
+
+	// Parts of AudioTokens that Azure Voice Live handled with Azure Speech, which it bills at their own rates.
+	StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens,omitempty"`
+	CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens,omitempty"`
 }
 
 // UnmarshalJSON maps OpenAI's cached_tokens into CachedReadTokens for compatibility.
@@ -1402,6 +1406,9 @@ func (d *ResponsesResponseInputTokens) UnmarshalJSON(data []byte) error {
 		CachedWriteTokenDetails *ChatCachedWriteTokenDetails `json:"cached_write_token_details"`
 		CachedTokens            *int                         `json:"cached_tokens"`
 		CacheWriteTokens        *int                         `json:"cache_write_tokens"`
+
+		StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens"`
+		CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens"`
 	}
 	if err := Unmarshal(data, &raw); err != nil {
 		return err
@@ -1412,6 +1419,8 @@ func (d *ResponsesResponseInputTokens) UnmarshalJSON(data []byte) error {
 	d.CachedReadTokens = raw.CachedReadTokens
 	d.CachedWriteTokens = raw.CachedWriteTokens
 	d.CachedWriteTokenDetails = raw.CachedWriteTokenDetails
+	d.StandardSpeechAudioTokens = raw.StandardSpeechAudioTokens
+	d.CustomSpeechAudioTokens = raw.CustomSpeechAudioTokens
 	// OpenAI spec providers send just cached_tokens, not separate read and write tokens and we handle them as read tokens in pricing calculations.
 	if raw.CachedTokens != nil && raw.CachedReadTokens == 0 && raw.CachedWriteTokens == 0 {
 		d.CachedReadTokens = *raw.CachedTokens
@@ -1438,6 +1447,9 @@ func (d ResponsesResponseInputTokens) MarshalJSON() ([]byte, error) {
 		// OpenAI's field name for cache writes (mirrors cached_tokens for reads) so the
 		// OpenAI SDK — which reads cache_write_tokens, not cached_write_tokens — finds it.
 		CacheWriteTokens int `json:"cache_write_tokens"`
+
+		StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens,omitempty"`
+		CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens,omitempty"`
 	}
 	return MarshalSorted(raw{
 		TextTokens:              d.TextTokens,
@@ -1448,6 +1460,9 @@ func (d ResponsesResponseInputTokens) MarshalJSON() ([]byte, error) {
 		CachedWriteTokenDetails: d.CachedWriteTokenDetails,
 		CachedTokens:            d.CachedReadTokens,
 		CacheWriteTokens:        d.CachedWriteTokens,
+
+		StandardSpeechAudioTokens: d.StandardSpeechAudioTokens,
+		CustomSpeechAudioTokens:   d.CustomSpeechAudioTokens,
 	})
 }
 
@@ -1461,6 +1476,10 @@ type ResponsesResponseOutputTokens struct {
 	CitationTokens           *int `json:"citation_tokens,omitempty"`
 	// Deprecated: use ResponsesResponseUsage.ToolUsage.WebSearch. Populated, will be removed in 3.0.0.
 	NumSearchQueries *int `json:"num_search_queries,omitempty"`
+
+	// Parts of AudioTokens that Azure Voice Live handled with Azure Speech, which it bills at their own rates.
+	StandardSpeechAudioTokens int `json:"standard_speech_audio_tokens,omitempty"`
+	CustomSpeechAudioTokens   int `json:"custom_speech_audio_tokens,omitempty"`
 }
 
 // =============================================================================

docs/architecture/framework/model-catalog.mdx

@@ -210,6 +210,17 @@ type PricingEntry struct {
     OutputCostPerVideoPerSecond *float64 `json:"output_cost_per_video_per_second,omitempty"`
     OutputCostPerSecond         *float64 `json:"output_cost_per_second,omitempty"`
 
+    // Azure Voice Live rates, used on Voice Live turns in place of the matching rates above when set
+    InputCostPerTokenVoiceLive                     *float64 `json:"input_cost_per_token_voice_live,omitempty"`
+    OutputCostPerTokenVoiceLive                    *float64 `json:"output_cost_per_token_voice_live,omitempty"`
+    CacheReadInputTokenCostVoiceLive               *float64 `json:"cache_read_input_token_cost_voice_live,omitempty"`
+    InputCostPerAudioTokenVoiceLive                *float64 `json:"input_cost_per_audio_token_voice_live,omitempty"`
+    OutputCostPerAudioTokenVoiceLive               *float64 `json:"output_cost_per_audio_token_voice_live,omitempty"`
+    InputCostPerStandardSpeechAudioTokenVoiceLive  *float64 `json:"input_cost_per_standard_speech_audio_token_voice_live,omitempty"`
+    OutputCostPerStandardSpeechAudioTokenVoiceLive *float64 `json:"output_cost_per_standard_speech_audio_token_voice_live,omitempty"`
+    InputCostPerCustomSpeechAudioTokenVoiceLive    *float64 `json:"input_cost_per_custom_speech_audio_token_voice_live,omitempty"`
+    OutputCostPerCustomSpeechAudioTokenVoiceLive   *float64 `json:"output_cost_per_custom_speech_audio_token_voice_live,omitempty"`
+
     OutputCostPerVideoPerSecond480p  *float64 `json:"output_cost_per_video_per_second_480p,omitempty"`
     OutputCostPerVideoPerSecond720p  *float64 `json:"output_cost_per_video_per_second_720p,omitempty"`
     OutputCostPerVideoPerSecond1024p *float64 `json:"output_cost_per_video_per_second_1024p,omitempty"`

docs/openapi/schemas/management/governance.yaml

@@ -2494,6 +2494,40 @@ PricingPatch:
     output_cost_per_second:
       type: number
       minimum: 0
+    input_cost_per_token_voice_live:
+      type: number
+      minimum: 0
+      description: >-
+        Azure Voice Live input rate per text token, used on Voice Live turns in place of
+        input_cost_per_token. Each _voice_live rate falls back to its normal rate when unset.
+    output_cost_per_token_voice_live:
+      type: number
+      minimum: 0
+    cache_read_input_token_cost_voice_live:
+      type: number
+      minimum: 0
+    input_cost_per_audio_token_voice_live:
+      type: number
+      minimum: 0
+    output_cost_per_audio_token_voice_live:
+      type: number
+      minimum: 0
+    input_cost_per_standard_speech_audio_token_voice_live:
+      type: number
+      minimum: 0
+      description: >-
+        Azure Voice Live rate per input audio token handled by Azure speech to text.
+    output_cost_per_standard_speech_audio_token_voice_live:
+      type: number
+      minimum: 0
+      description: >-
+        Azure Voice Live rate per output audio token from an Azure standard voice.
+    input_cost_per_custom_speech_audio_token_voice_live:
+      type: number
+      minimum: 0
+    output_cost_per_custom_speech_audio_token_voice_live:
+      type: number
+      minimum: 0
     output_cost_per_video_per_second_480p:
       type: number
       minimum: 0