feat: implements CEL expression API for costs

Makefile

@@ -19,8 +19,10 @@ ENABLE_MULTI_PLATFORMS ?= false
 HELM_CHART_VERSION ?= v0.0.0-latest
 
 # Arguments for go test. This can be used, for example, to run specific tests via
-# `GO_TEST_EXTRA_ARGS="-run TestName/foo/etc"`.
-GO_TEST_EXTRA_ARGS ?=
+# `GO_TEST_ARGS="-run TestName/foo/etc" -v -race`.
+GO_TEST_ARGS ?= -v -race
+# Arguments for go test in e2e tests in addition to GO_TEST_ARGS, applicable to test-e2e, test-extproc, and test-controller.
+GO_TEST_E2E_ARGS ?= -count=1
 
 # This will print out the help message for contributing to the project.
 .PHONY: help
@@ -103,7 +105,7 @@ editorconfig: editorconfig-checker
 .PHONY: test
 test:
 	@echo "test => ./..."
-	@go test -v ./...
+	@go test $(GO_TEST_ARGS) ./...
 
 ENVTEST_K8S_VERSIONS ?= 1.29.0 1.30.0 1.31.0
 
@@ -115,7 +117,7 @@ test-cel: envtest apigen
 	@for k8sVersion in $(ENVTEST_K8S_VERSIONS); do \
   		echo "Run CEL Validation on k8s $$k8sVersion"; \
         KUBEBUILDER_ASSETS="$$($(ENVTEST) use $$k8sVersion -p path)" \
-                 go test ./tests/cel-validation $(GO_TEST_EXTRA_ARGS) --tags test_cel_validation -v -count=1; \
+                 go test ./tests/cel-validation $(GO_TEST_ARGS) $(GO_TEST_E2E_ARGS) --tags test_cel_validation; \
     done
 
 # This runs the end-to-end tests for extproc without controller or k8s at all.
@@ -127,15 +129,15 @@ test-extproc: build.extproc
 	@$(MAKE) build.extproc_custom_router CMD_PATH_PREFIX=examples
 	@$(MAKE) build.testupstream CMD_PATH_PREFIX=tests
 	@echo "Run ExtProc test"
-	@go test ./tests/extproc/... $(GO_TEST_EXTRA_ARGS) -tags test_extproc -v -count=1
+	@go test ./tests/extproc/... $(GO_TEST_ARGS) $(GO_TEST_E2E_ARGS) -tags test_extproc -v
 
 # This runs the end-to-end tests for the controller with EnvTest.
 .PHONY: test-controller
 test-controller: envtest apigen
 	@for k8sVersion in $(ENVTEST_K8S_VERSIONS); do \
   		echo "Run Controller tests on k8s $$k8sVersion"; \
         KUBEBUILDER_ASSETS="$$($(ENVTEST) use $$k8sVersion -p path)" \
-                 go test ./tests/controller $(GO_TEST_EXTRA_ARGS) --tags test_controller -v -count=1; \
+                 go test ./tests/controller $(GO_TEST_ARGS) $(GO_TEST_E2E_ARGS) -tags test_controller; \
     done
 
 # This runs the end-to-end tests for the controller and extproc with a local kind cluster.
@@ -146,7 +148,7 @@ test-e2e: kind
 	@$(MAKE) docker-build DOCKER_BUILD_ARGS="--load"
 	@$(MAKE) docker-build.testupstream CMD_PATH_PREFIX=tests DOCKER_BUILD_ARGS="--load"
 	@echo "Run E2E tests"
-	@go test ./tests/e2e/... $(GO_TEST_EXTRA_ARGS) -tags test_e2e -v -count=1
+	@go test ./tests/e2e/... $(GO_TEST_ARGS) $(GO_TEST_E2E_ARGS) -tags test_e2e
 
 # This builds a binary for the given command under the internal/cmd directory.
 #

api/v1alpha1/api.go

@@ -475,17 +475,32 @@ type LLMRequestCost struct {
 	// +kubebuilder:validation:Required
 	MetadataKey string `json:"metadataKey"`
 	// Type specifies the type of the request cost. The default is "OutputToken",
-	// and it uses "output token" as the cost. The other types are "InputToken" and "TotalToken".
+	// and it uses "output token" as the cost. The other types are "InputToken", "TotalToken",
+	// and "CEL".
 	//
-	// +kubebuilder:validation:Enum=OutputToken;InputToken;TotalToken
+	// +kubebuilder:validation:Enum=OutputToken;InputToken;TotalToken;CEL
 	Type LLMRequestCostType `json:"type"`
 	// CELExpression is the CEL expression to calculate the cost of the request.
-	// The CEL expression must return an integer value. The CEL expression should be
-	// able to access the request headers, model name, backend name, input/output tokens etc.
+	// The CEL expression must return a signed or unsigned integer. If the
+	// return value is negative, it will be error.
+	//
+	// The expression can use the following variables:
+	//
+	//	* model: the model name extracted from the request content. Type: string.
+	//	* backend: the backend name in the form of "name.namespace". Type: string.
+	//	* input_tokens: the number of input tokens. Type: unsigned integer.
+	//	* output_tokens: the number of output tokens. Type: unsigned integer.
+	//	* total_tokens: the total number of tokens. Type: unsigned integer.
+	//
+	// For example, the following expressions are valid:
+	//
+	// 	* "model == 'llama' ?  input_tokens + output_token * 0.5 : total_tokens"
+	//	* "backend == 'foo.default' ?  input_tokens + output_tokens : total_tokens"
+	//	* "input_tokens + output_tokens + total_tokens"
+	//	* "input_tokens * output_tokens"
 	//
 	// +optional
-	// +notImplementedHide https://github.com/envoyproxy/ai-gateway/issues/97
-	CELExpression *string `json:"celExpression"`
+	CELExpression *string `json:"celExpression,omitempty"`
 }
 
 // LLMRequestCostType specifies the type of the LLMRequestCost.

go.mod

@@ -11,6 +11,7 @@ require (
 	github.com/envoyproxy/gateway v0.5.0-rc.1.0.20250115172926-8b89dadfbd2c
 	github.com/envoyproxy/go-control-plane/envoy v1.32.3
 	github.com/go-logr/logr v1.4.2
+	github.com/google/cel-go v0.22.1
 	github.com/google/go-cmp v0.6.0
 	github.com/openai/openai-go v0.1.0-alpha.46
 	github.com/stretchr/testify v1.10.0
@@ -30,6 +31,7 @@ require (
 
 require (
 	cel.dev/expr v0.18.0 // indirect
+	github.com/antlr4-go/antlr/v4 v4.13.0 // indirect
 	github.com/aws/aws-sdk-go-v2/credentials v1.17.54 // indirect
 	github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.16.24 // indirect
 	github.com/aws/aws-sdk-go-v2/internal/configsources v1.3.28 // indirect
@@ -75,6 +77,7 @@ require (
 	github.com/prometheus/procfs v0.15.1 // indirect
 	github.com/rogpeppe/go-internal v1.13.1 // indirect
 	github.com/spf13/pflag v1.0.5 // indirect
+	github.com/stoewer/go-strcase v1.3.0 // indirect
 	github.com/tidwall/gjson v1.14.4 // indirect
 	github.com/tidwall/match v1.1.1 // indirect
 	github.com/tidwall/pretty v1.2.1 // indirect

go.sum

@@ -1,5 +1,7 @@
 cel.dev/expr v0.18.0 h1:CJ6drgk+Hf96lkLikr4rFf19WrU0BOWEihyZnI2TAzo=
 cel.dev/expr v0.18.0/go.mod h1:MrpN08Q+lEBs+bGYdLxxHkZoUSsCp0nSKTs0nTymJgw=
+github.com/antlr4-go/antlr/v4 v4.13.0 h1:lxCg3LAv+EUK6t1i0y1V6/SLeUi0eKEKdhQAlS8TVTI=
+github.com/antlr4-go/antlr/v4 v4.13.0/go.mod h1:pfChB/xh/Unjila75QW7+VU4TSnWnnk9UTnmpPaOR2g=
 github.com/aws/aws-sdk-go-v2 v1.33.0 h1:Evgm4DI9imD81V0WwD+TN4DCwjUMdc94TrduMLbgZJs=
 github.com/aws/aws-sdk-go-v2 v1.33.0/go.mod h1:P5WJBrYqqbWVaOxgH0X/FYYD47/nooaPOZPlQdmiN2U=
 github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.7 h1:lL7IfaFzngfx0ZwUGOZdsFFnQ5uLvR0hWqqhyE7Q9M8=
@@ -70,6 +72,8 @@ github.com/gogo/protobuf v1.3.2 h1:Ov1cvc58UF3b5XjBnZv7+opcTcQFZebYjWzi34vdm4Q=
 github.com/gogo/protobuf v1.3.2/go.mod h1:P1XiOD3dCwIKUDQYPy72D8LYyHL2YPYrpS2s69NZV8Q=
 github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=
 github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6rSs7xps=
+github.com/google/cel-go v0.22.1 h1:AfVXx3chM2qwoSbM7Da8g8hX8OVSkBFwX+rz2+PcK40=
+github.com/google/cel-go v0.22.1/go.mod h1:BuznPXXfQDpXKWQ9sPW3TzlAJN5zzFe+i9tIs0yC4s8=
 github.com/google/gnostic-models v0.6.9 h1:MU/8wDLif2qCXZmzncUQ/BOfxWfthHi63KqpoNbWqVw=
 github.com/google/gnostic-models v0.6.9/go.mod h1:CiWsm0s6BSQd1hRn8/QmxqB6BesYcbSZxsz9b0KuDBw=
 github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
@@ -130,8 +134,15 @@ github.com/rogpeppe/go-internal v1.13.1 h1:KvO1DLK/DRN07sQ1LQKScxyZJuNnedQ5/wKSR
 github.com/rogpeppe/go-internal v1.13.1/go.mod h1:uMEvuHeurkdAXX61udpOXGD/AzZDWNMNyH2VO9fmH0o=
 github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA=
 github.com/spf13/pflag v1.0.5/go.mod h1:McXfInJRrz4CZXVZOBLb0bTZqETkiAhM9Iw0y3An2Bg=
+github.com/stoewer/go-strcase v1.3.0 h1:g0eASXYtp+yvN9fK8sH94oCIk0fau9uV1/ZdJ0AVEzs=
+github.com/stoewer/go-strcase v1.3.0/go.mod h1:fAH5hQ5pehh+j3nZfvwdk2RgEgQjAoM8wodgtPmh1xo=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=
 github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
 github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
 github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
 github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
@@ -224,6 +235,7 @@ gopkg.in/evanphx/json-patch.v4 v4.12.0 h1:n6jtcsulIzXPJaxegRbvFNNrZDjbij7ny3gmSP
 gopkg.in/evanphx/json-patch.v4 v4.12.0/go.mod h1:p8EYWUEYMpynmqDbY58zCKCFZw8pRWMG4EsWvDvM72M=
 gopkg.in/inf.v0 v0.9.1 h1:73M5CoZyi3ZLMOyDlQh031Cx6N9NDJ2Vvfl76EDAgDc=
 gopkg.in/inf.v0 v0.9.1/go.mod h1:cWUDdTG/fYaXco+Dcufb5Vnc6Gp2YChqWtbxRZE0mXw=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 k8s.io/api v0.32.1 h1:f562zw9cy+GvXzXf0CKlVQ7yHJVYzLfL6JAS4kOAaOc=

internal/controller/sink.go

@@ -16,6 +16,7 @@ import (
 
 	aigv1a1 "github.com/envoyproxy/ai-gateway/api/v1alpha1"
 	"github.com/envoyproxy/ai-gateway/filterconfig"
+	"github.com/envoyproxy/ai-gateway/internal/llmcostcel"
 )
 
 const selectedBackendHeaderKey = "x-ai-eg-selected-backend"
@@ -261,6 +262,13 @@ func (c *configSink) updateExtProcConfigMap(aiGatewayRoute *aigv1a1.AIGatewayRou
 			fc.Type = filterconfig.LLMRequestCostTypeTotalToken
 		case aigv1a1.LLMRequestCostTypeCEL:
 			fc.Type = filterconfig.LLMRequestCostTypeCELExpression
+			expr := *cost.CELExpression
+			// Sanity check the CEL expression.
+			_, err := llmcostcel.NewProgram(expr)
+			if err != nil {
+				return fmt.Errorf("invalid CEL expression: %w", err)
+			}
+			fc.CELExpression = expr
 		default:
 			return fmt.Errorf("unknown request cost type: %s", cost.Type)
 		}

internal/controller/sink_test.go

@@ -316,6 +316,15 @@ func Test_updateExtProcConfigMap(t *testing.T) {
 							Type:        aigv1a1.LLMRequestCostTypeInputToken,
 							MetadataKey: "input-token",
 						},
+						{
+							Type:        aigv1a1.LLMRequestCostTypeTotalToken,
+							MetadataKey: "total-token",
+						},
+						{
+							Type:          aigv1a1.LLMRequestCostTypeCEL,
+							MetadataKey:   "cel-token",
+							CELExpression: ptr.To("model == 'cool_model' ?  input_tokens * output_tokens : total_tokens"),
+						},
 					},
 				},
 			},
@@ -347,6 +356,8 @@ func Test_updateExtProcConfigMap(t *testing.T) {
 				LLMRequestCosts: []filterconfig.LLMRequestCost{
 					{Type: filterconfig.LLMRequestCostTypeOutputToken, MetadataKey: "output-token"},
 					{Type: filterconfig.LLMRequestCostTypeInputToken, MetadataKey: "input-token"},
+					{Type: filterconfig.LLMRequestCostTypeTotalToken, MetadataKey: "total-token"},
+					{Type: filterconfig.LLMRequestCostTypeCELExpression, MetadataKey: "cel-token", CELExpression: "model == 'cool_model' ?  input_tokens * output_tokens : total_tokens"},
 				},
 			},
 		},

internal/extproc/processor.go

@@ -11,25 +11,32 @@ import (
 
 	corev3 "github.com/envoyproxy/go-control-plane/envoy/config/core/v3"
 	extprocv3 "github.com/envoyproxy/go-control-plane/envoy/service/ext_proc/v3"
+	"github.com/google/cel-go/cel"
 	"google.golang.org/protobuf/types/known/structpb"
 
 	"github.com/envoyproxy/ai-gateway/extprocapi"
 	"github.com/envoyproxy/ai-gateway/filterconfig"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/backendauth"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/router"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/translator"
+	"github.com/envoyproxy/ai-gateway/internal/llmcostcel"
 )
 
 // processorConfig is the configuration for the processor.
 // This will be created by the server and passed to the processor when it detects a new configuration.
 type processorConfig struct {
 	bodyParser                                   router.RequestBodyParser
 	router                                       extprocapi.Router
-	ModelNameHeaderKey, selectedBackendHeaderKey string
+	modelNameHeaderKey, selectedBackendHeaderKey string
 	factories                                    map[filterconfig.VersionedAPISchema]translator.Factory
 	backendAuthHandlers                          map[string]backendauth.Handler
 	metadataNamespace                            string
-	requestCosts                                 []filterconfig.LLMRequestCost
+	requestCosts                                 []processorConfigRequestCost
+}
+
+type processorConfigRequestCost struct {
+	*filterconfig.LLMRequestCost
+	celProg cel.Program
 }
 
 // ProcessorIface is the interface for the processor.
@@ -79,7 +86,7 @@ func (p *Processor) ProcessRequestBody(_ context.Context, rawBody *extprocv3.Htt
 	}
 	p.logger.Info("Processing request", "path", path, "model", model)
 
-	p.requestHeaders[p.config.ModelNameHeaderKey] = model
+	p.requestHeaders[p.config.modelNameHeaderKey] = model
 	b, err := p.config.router.Calculate(p.requestHeaders)
 	if err != nil {
 		return nil, fmt.Errorf("failed to calculate route: %w", err)
@@ -107,7 +114,7 @@ func (p *Processor) ProcessRequestBody(_ context.Context, rawBody *extprocv3.Htt
 	}
 	// Set the model name to the request header with the key `x-ai-gateway-llm-model-name`.
 	headerMutation.SetHeaders = append(headerMutation.SetHeaders, &corev3.HeaderValueOption{
-		Header: &corev3.HeaderValue{Key: p.config.ModelNameHeaderKey, RawValue: []byte(model)},
+		Header: &corev3.HeaderValue{Key: p.config.modelNameHeaderKey, RawValue: []byte(model)},
 	}, &corev3.HeaderValueOption{
 		Header: &corev3.HeaderValue{Key: p.config.selectedBackendHeaderKey, RawValue: []byte(b.Name)},
 	})
@@ -203,7 +210,8 @@ func (p *Processor) ProcessResponseBody(_ context.Context, body *extprocv3.HttpB
 
 func (p *Processor) maybeBuildDynamicMetadata() (*structpb.Struct, error) {
 	metadata := make(map[string]*structpb.Value, len(p.config.requestCosts))
-	for _, c := range p.config.requestCosts {
+	for i := range p.config.requestCosts {
+		c := &p.config.requestCosts[i]
 		var cost uint32
 		switch c.Type {
 		case filterconfig.LLMRequestCostTypeInputToken:
@@ -212,6 +220,19 @@ func (p *Processor) maybeBuildDynamicMetadata() (*structpb.Struct, error) {
 			cost = p.costs.OutputTokens
 		case filterconfig.LLMRequestCostTypeTotalToken:
 			cost = p.costs.TotalTokens
+		case filterconfig.LLMRequestCostTypeCELExpression:
+			costU64, err := llmcostcel.EvaluateProgram(
+				c.celProg,
+				p.requestHeaders[p.config.modelNameHeaderKey],
+				p.requestHeaders[p.config.selectedBackendHeaderKey],
+				p.costs.InputTokens,
+				p.costs.OutputTokens,
+				p.costs.TotalTokens,
+			)
+			if err != nil {
+				return nil, fmt.Errorf("failed to evaluate CEL expression: %w", err)
+			}
+			cost = uint32(costU64) //nolint:gosec
 		default:
 			return nil, fmt.Errorf("unknown request cost kind: %s", c.Type)
 		}

internal/extproc/processor_test.go

@@ -13,6 +13,7 @@ import (
 	"github.com/envoyproxy/ai-gateway/filterconfig"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/router"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/translator"
+	"github.com/envoyproxy/ai-gateway/internal/llmcostcel"
 )
 
 func TestProcessor_ProcessRequestHeaders(t *testing.T) {
@@ -65,11 +66,24 @@ func TestProcessor_ProcessResponseBody(t *testing.T) {
 			retBodyMutation: expBodyMut, retHeaderMutation: expHeadMut,
 			retUsedToken: translator.LLMTokenUsage{OutputTokens: 123, InputTokens: 1},
 		}
+
+		celProgInt, err := llmcostcel.NewProgram("54321")
+		require.NoError(t, err)
+		celProgUint, err := llmcostcel.NewProgram("uint(9999)")
+		require.NoError(t, err)
 		p := &Processor{translator: mt, config: &processorConfig{
 			metadataNamespace: "ai_gateway_llm_ns",
-			requestCosts: []filterconfig.LLMRequestCost{
-				{Type: filterconfig.LLMRequestCostTypeOutputToken, MetadataKey: "output_token_usage"},
-				{Type: filterconfig.LLMRequestCostTypeInputToken, MetadataKey: "input_token_usage"},
+			requestCosts: []processorConfigRequestCost{
+				{LLMRequestCost: &filterconfig.LLMRequestCost{Type: filterconfig.LLMRequestCostTypeOutputToken, MetadataKey: "output_token_usage"}},
+				{LLMRequestCost: &filterconfig.LLMRequestCost{Type: filterconfig.LLMRequestCostTypeInputToken, MetadataKey: "input_token_usage"}},
+				{
+					celProg:        celProgInt,
+					LLMRequestCost: &filterconfig.LLMRequestCost{Type: filterconfig.LLMRequestCostTypeCELExpression, MetadataKey: "cel_int"},
+				},
+				{
+					celProg:        celProgUint,
+					LLMRequestCost: &filterconfig.LLMRequestCost{Type: filterconfig.LLMRequestCostTypeCELExpression, MetadataKey: "cel_uint"},
+				},
 			},
 		}}
 		res, err := p.ProcessResponseBody(context.Background(), inBody)
@@ -84,6 +98,10 @@ func TestProcessor_ProcessResponseBody(t *testing.T) {
 			GetStructValue().Fields["output_token_usage"].GetNumberValue())
 		require.Equal(t, float64(1), md.Fields["ai_gateway_llm_ns"].
 			GetStructValue().Fields["input_token_usage"].GetNumberValue())
+		require.Equal(t, float64(54321), md.Fields["ai_gateway_llm_ns"].
+			GetStructValue().Fields["cel_int"].GetNumberValue())
+		require.Equal(t, float64(9999), md.Fields["ai_gateway_llm_ns"].
+			GetStructValue().Fields["cel_uint"].GetNumberValue())
 	})
 }
 
@@ -168,7 +186,7 @@ func TestProcessor_ProcessRequestBody(t *testing.T) {
 				{Name: "some-schema", Version: "v10.0"}: factory.impl,
 			},
 			selectedBackendHeaderKey: "x-ai-gateway-backend-key",
-			ModelNameHeaderKey:       "x-ai-gateway-model-key",
+			modelNameHeaderKey:       "x-ai-gateway-model-key",
 		}, requestHeaders: headers, logger: slog.Default()}
 		resp, err := p.ProcessRequestBody(context.Background(), &extprocv3.HttpBody{})
 		require.NoError(t, err)

internal/extproc/server.go

@@ -8,6 +8,7 @@ import (
 	"log/slog"
 
 	extprocv3 "github.com/envoyproxy/go-control-plane/envoy/service/ext_proc/v3"
+	"github.com/google/cel-go/cel"
 	"google.golang.org/grpc/codes"
 	"google.golang.org/grpc/health/grpc_health_v1"
 	"google.golang.org/grpc/status"
@@ -17,6 +18,7 @@ import (
 	"github.com/envoyproxy/ai-gateway/internal/extproc/backendauth"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/router"
 	"github.com/envoyproxy/ai-gateway/internal/extproc/translator"
+	"github.com/envoyproxy/ai-gateway/internal/llmcostcel"
 )
 
 // Server implements the external process server.
@@ -64,14 +66,27 @@ func (s *Server[P]) LoadConfig(config *filterconfig.Config) error {
 		}
 	}
 
+	costs := make([]processorConfigRequestCost, 0, len(config.LLMRequestCosts))
+	for i := range config.LLMRequestCosts {
+		c := &config.LLMRequestCosts[i]
+		var prog cel.Program
+		if c.CELExpression != "" {
+			prog, err = llmcostcel.NewProgram(c.CELExpression)
+			if err != nil {
+				return fmt.Errorf("cannot create CEL program for cost: %w", err)
+			}
+		}
+		costs = append(costs, processorConfigRequestCost{LLMRequestCost: c, celProg: prog})
+	}
+
 	newConfig := &processorConfig{
 		bodyParser: bodyParser, router: rt,
 		selectedBackendHeaderKey: config.SelectedBackendHeaderKey,
-		ModelNameHeaderKey:       config.ModelNameHeaderKey,
+		modelNameHeaderKey:       config.ModelNameHeaderKey,
 		factories:                factories,
 		backendAuthHandlers:      backendAuthHandlers,
 		metadataNamespace:        config.MetadataNamespace,
-		requestCosts:             config.LLMRequestCosts,
+		requestCosts:             costs,
 	}
 	s.config = newConfig // This is racey, but we don't care.
 	return nil