diff --git a/providers/openai/language_model.go b/providers/openai/language_model.go
index f9b03ff0c..b46651c8b 100644
--- a/providers/openai/language_model.go
+++ b/providers/openai/language_model.go
@@ -341,7 +341,12 @@ func (o languageModel) Stream(ctx context.Context, call fantasy.Call) (fantasy.S
 		for stream.Next() {
 			chunk := stream.Current()
 			acc.AddChunk(chunk)
-			usage, providerMetadata = o.streamUsageFunc(chunk, extraContext, providerMetadata)
+			// Some OpenAI-compatible backends emit cumulative usage on
+			// delta chunks and end with a usage-less finish chunk; keep
+			// the last usage-bearing result instead of zeroing it.
+			if chunkUsage, chunkMetadata := o.streamUsageFunc(chunk, extraContext, providerMetadata); chunkUsage != (fantasy.Usage{}) {
+				usage, providerMetadata = chunkUsage, chunkMetadata
+			}
 			if len(chunk.Choices) == 0 {
 				continue
 			}
@@ -870,8 +875,11 @@ func (o languageModel) streamObjectWithJSONMode(ctx context.Context, call fantas
 		for stream.Next() {
 			chunk := stream.Current()
 
-			// Update usage
-			usage, providerMetadata = o.streamUsageFunc(chunk, make(map[string]any), providerMetadata)
+			// Update usage, ignoring usage-less chunks so a trailing
+			// finish chunk cannot zero previously reported usage.
+			if chunkUsage, chunkMetadata := o.streamUsageFunc(chunk, make(map[string]any), providerMetadata); chunkUsage != (fantasy.Usage{}) {
+				usage, providerMetadata = chunkUsage, chunkMetadata
+			}
 
 			if len(chunk.Choices) == 0 {
 				continue
diff --git a/providers/openai/stream_usage_test.go b/providers/openai/stream_usage_test.go
new file mode 100644
index 000000000..635d797f3
--- /dev/null
+++ b/providers/openai/stream_usage_test.go
@@ -0,0 +1,96 @@
+package openai
+
+import (
+	"context"
+	"testing"
+
+	"charm.land/fantasy"
+	"github.com/stretchr/testify/require"
+)
+
+// Some OpenAI-compatible backends report cumulative usage on delta
+// chunks and end with a usage-less finish chunk; the last
+// usage-bearing chunk must win.
+func TestStreamUsageSurvivesTrailingUsagelessChunk(t *testing.T) {
+	t.Parallel()
+
+	server := newStreamingMockServer()
+	defer server.close()
+
+	server.chunks = []string{
+		`data: {"id":"c1","object":"chat.completion.chunk","created":1,"model":"laguna-xs","choices":[{"index":0,"delta":{"role":"assistant","content":"Hel"},"finish_reason":null}],"usage":{"prompt_tokens":100,"completion_tokens":1,"total_tokens":101}}` + "\n\n",
+		`data: {"id":"c1","object":"chat.completion.chunk","created":1,"model":"laguna-xs","choices":[{"index":0,"delta":{"content":"lo"},"finish_reason":null}],"usage":{"prompt_tokens":100,"completion_tokens":5,"total_tokens":105,"prompt_tokens_details":{"cached_tokens":80}}}` + "\n\n",
+		`data: {"id":"c1","object":"chat.completion.chunk","created":1,"model":"laguna-xs","choices":[{"index":0,"delta":{},"finish_reason":"stop"}],"usage":null}` + "\n\n",
+		"data: [DONE]\n\n",
+	}
+
+	provider, err := New(
+		WithAPIKey("test-api-key"),
+		WithBaseURL(server.server.URL),
+	)
+	require.NoError(t, err)
+	model, err := provider.LanguageModel(t.Context(), "laguna-xs")
+	require.NoError(t, err)
+
+	stream, err := model.Stream(context.Background(), fantasy.Call{Prompt: testPrompt})
+	require.NoError(t, err)
+
+	parts, err := collectStreamParts(stream)
+	require.NoError(t, err)
+
+	var finish *fantasy.StreamPart
+	for i, part := range parts {
+		if part.Type == fantasy.StreamPartTypeFinish {
+			finish = &parts[i]
+		}
+	}
+	require.NotNil(t, finish)
+	require.Equal(t, fantasy.FinishReasonStop, finish.FinishReason)
+	// prompt_tokens includes cached tokens; input is reported net of cache.
+	require.Equal(t, int64(20), finish.Usage.InputTokens)
+	require.Equal(t, int64(80), finish.Usage.CacheReadTokens)
+	require.Equal(t, int64(5), finish.Usage.OutputTokens)
+	require.Equal(t, int64(105), finish.Usage.TotalTokens)
+}
+
+func TestStreamObjectUsageSurvivesTrailingUsagelessChunk(t *testing.T) {
+	t.Parallel()
+
+	server := newStreamingMockServer()
+	defer server.close()
+
+	server.chunks = []string{
+		`data: {"id":"c1","object":"chat.completion.chunk","created":1,"model":"laguna-xs","choices":[{"index":0,"delta":{"role":"assistant","content":"{\"answer\":\"hello\"}"},"finish_reason":null}],"usage":{"prompt_tokens":100,"completion_tokens":5,"total_tokens":105,"prompt_tokens_details":{"cached_tokens":80}}}` + "\n\n",
+		`data: {"id":"c1","object":"chat.completion.chunk","created":1,"model":"laguna-xs","choices":[{"index":0,"delta":{},"finish_reason":"stop"}],"usage":null}` + "\n\n",
+		"data: [DONE]\n\n",
+	}
+
+	provider, err := New(
+		WithAPIKey("test-api-key"),
+		WithBaseURL(server.server.URL),
+	)
+	require.NoError(t, err)
+	model, err := provider.LanguageModel(t.Context(), "laguna-xs")
+	require.NoError(t, err)
+
+	stream, err := model.StreamObject(context.Background(), fantasy.ObjectCall{
+		Prompt: testPrompt,
+		Schema: fantasy.Schema{
+			Type: "object",
+			Properties: map[string]*fantasy.Schema{
+				"answer": {Type: "string"},
+			},
+			Required: []string{"answer"},
+		},
+	})
+	require.NoError(t, err)
+
+	parts := collectObjectStreamParts(stream)
+	require.NotEmpty(t, parts)
+	finish := parts[len(parts)-1]
+	require.Equal(t, fantasy.ObjectStreamPartTypeFinish, finish.Type)
+	require.Equal(t, int64(20), finish.Usage.InputTokens)
+	require.Equal(t, int64(80), finish.Usage.CacheReadTokens)
+	require.Equal(t, int64(5), finish.Usage.OutputTokens)
+	require.Equal(t, int64(105), finish.Usage.TotalTokens)
+}