diff --git a/internal/registry/model_registry.go b/internal/registry/model_registry.go
index 7b8b262e..e036a04f 100644
--- a/internal/registry/model_registry.go
+++ b/internal/registry/model_registry.go
@@ -47,6 +47,10 @@ type ModelInfo struct {
 	MaxCompletionTokens int `json:"max_completion_tokens,omitempty"`
 	// SupportedParameters lists supported parameters
 	SupportedParameters []string `json:"supported_parameters,omitempty"`
+	// SupportedInputModalities lists supported input modalities (e.g., TEXT, IMAGE, VIDEO, AUDIO)
+	SupportedInputModalities []string `json:"supportedInputModalities,omitempty"`
+	// SupportedOutputModalities lists supported output modalities (e.g., TEXT, IMAGE)
+	SupportedOutputModalities []string `json:"supportedOutputModalities,omitempty"`
 
 	// Thinking holds provider-specific reasoning/thinking budget capabilities.
 	// This is optional and currently used for Gemini thinking budget normalization.
@@ -499,6 +503,12 @@ func cloneModelInfo(model *ModelInfo) *ModelInfo {
 	if len(model.SupportedParameters) > 0 {
 		copyModel.SupportedParameters = append([]string(nil), model.SupportedParameters...)
 	}
+	if len(model.SupportedInputModalities) > 0 {
+		copyModel.SupportedInputModalities = append([]string(nil), model.SupportedInputModalities...)
+	}
+	if len(model.SupportedOutputModalities) > 0 {
+		copyModel.SupportedOutputModalities = append([]string(nil), model.SupportedOutputModalities...)
+	}
 	return &copyModel
 }
 
@@ -1067,6 +1077,12 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 		if len(model.SupportedGenerationMethods) > 0 {
 			result["supportedGenerationMethods"] = model.SupportedGenerationMethods
 		}
+		if len(model.SupportedInputModalities) > 0 {
+			result["supportedInputModalities"] = model.SupportedInputModalities
+		}
+		if len(model.SupportedOutputModalities) > 0 {
+			result["supportedOutputModalities"] = model.SupportedOutputModalities
+		}
 		return result
 
 	default:
diff --git a/internal/runtime/executor/antigravity_executor.go b/internal/runtime/executor/antigravity_executor.go
index 7e480a97..e697b64e 100644
--- a/internal/runtime/executor/antigravity_executor.go
+++ b/internal/runtime/executor/antigravity_executor.go
@@ -1176,6 +1176,29 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 				OwnedBy:     antigravityAuthType,
 				Type:        antigravityAuthType,
 			}
+
+			// Build input modalities from upstream capability flags.
+			inputModalities := []string{"TEXT"}
+			if modelData.Get("supportsImages").Bool() {
+				inputModalities = append(inputModalities, "IMAGE")
+			}
+			if modelData.Get("supportsVideo").Bool() {
+				inputModalities = append(inputModalities, "VIDEO")
+			}
+			modelInfo.SupportedInputModalities = inputModalities
+			modelInfo.SupportedOutputModalities = []string{"TEXT"}
+
+			// Token limits from upstream.
+			if maxTok := modelData.Get("maxTokens").Int(); maxTok > 0 {
+				modelInfo.InputTokenLimit = int(maxTok)
+			}
+			if maxOut := modelData.Get("maxOutputTokens").Int(); maxOut > 0 {
+				modelInfo.OutputTokenLimit = int(maxOut)
+			}
+
+			// Supported generation methods (Gemini v1beta convention).
+			modelInfo.SupportedGenerationMethods = []string{"generateContent", "countTokens"}
+
 			// Look up Thinking support from static config using upstream model name.
 			if modelCfg != nil {
 				if modelCfg.Thinking != nil {