diff --git a/backend/README.md b/backend/README.md
index 419bec0..9f4f001 100644
--- a/backend/README.md
+++ b/backend/README.md
@@ -113,8 +113,8 @@ On model/auth/timeout failure the service returns HTTP 502 with
 | `GCP_LOCATION`    | `us-east1`         | Vertex AI region                         |
 | `GEMINI_MODEL`    | `gemini-2.5-flash` | Model id                                 |
 | `GEMINI_TIMEOUT_MS` | `30000`          | Per-request model timeout (ms)           |
-| `GEMINI_MAX_OUTPUT_TOKENS` | `700`     | Analysis output cap for the single structured call |
-| `GEMINI_CHAT_MAX_OUTPUT_TOKENS` | `512` | Result-chat output cap |
+| `GEMINI_MAX_OUTPUT_TOKENS` | `4096`    | Analysis output cap (shared by gemini-2.5-flash thinking + JSON answer) |
+| `GEMINI_CHAT_MAX_OUTPUT_TOKENS` | `2048` | Result-chat output cap (shared by thinking + answer) |
 | `ENABLE_OPENAPI_DOCS` | `false`        | Enables `/docs`, `/redoc`, and `/openapi.json` only for local/dev use |
 | `CORS_ALLOWED_ORIGINS` | empty         | Comma-separated browser origins allowed to call the API; empty disables CORS middleware |
 | `PORT`            | `8080`             | Server port (set by Cloud Run)           |
diff --git a/backend/main.py b/backend/main.py
index 0431660..c0a16e0 100644
--- a/backend/main.py
+++ b/backend/main.py
@@ -33,8 +33,12 @@
 GCP_LOCATION = os.getenv("GCP_LOCATION", "us-east1")
 GEMINI_MODEL = os.getenv("GEMINI_MODEL", "gemini-2.5-flash")
 REQUEST_TIMEOUT_MS = int(os.getenv("GEMINI_TIMEOUT_MS", "30000"))
-FULL_ANALYSIS_MAX_OUTPUT_TOKENS = int(os.getenv("GEMINI_MAX_OUTPUT_TOKENS", "700"))
-CHAT_MAX_OUTPUT_TOKENS = int(os.getenv("GEMINI_CHAT_MAX_OUTPUT_TOKENS", "512"))
+# ponytail: gemini-2.5-flash "thinking" tokens share this output budget. 700/512 let thinking
+# eat the whole cap on complex labels, truncating the JSON to garbage -> 502 model_response_unparseable.
+# Raised to comfortably fit thinking + answer for real OCR inputs. If a pathological label still
+# truncates, bound thinking with types.ThinkingConfig(thinking_budget=N) on the generate_content calls.
+FULL_ANALYSIS_MAX_OUTPUT_TOKENS = int(os.getenv("GEMINI_MAX_OUTPUT_TOKENS", "4096"))
+CHAT_MAX_OUTPUT_TOKENS = int(os.getenv("GEMINI_CHAT_MAX_OUTPUT_TOKENS", "2048"))
 MAX_INGREDIENT_CHARS = 20_000
 MAX_CHAT_QUESTION_CHARS = 1_000
 MAX_CHAT_HISTORY_MESSAGES = 12