# Gemini-3-pro-image-preview returning 400 INVALID\_ARGUMENT since ~11:10 UTC Feb 13 **URL:** \ **Category:** Gemini API\ **Tags:** gemini-3\ **Created:** [February 13, 2026, 12:34pm UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920 "2026-02-13T12:34:54Z")\ **Posts on this page:** 7\ **Page:** 1
**Author:** ![xsyjeon](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/xsyjeon/32/93041_2.png) [@xsyjeon](https://discuss.ai.google.dev/u/xsyjeon)\ **Post date:** [February 13, 2026, 12:34pm UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/1 "2026-02-13T12:34:54Z")
**Environment:** - Python SDK: google-genai ^1.52.0 - Endpoint: [generativelanguage.googleapis.com/v1beta](http://generativelanguage.googleapis.com/v1beta) - Model: gemini-3-pro-image-preview **Issue:** All image generation requests (both text-to-image and image editing with reference images) to gemini-3-pro-image-preview started returning 400 INVALID\_ARGUMENT around 11:10 UTC on February 13, 2026. Before that, requests were working (last successful at 08:04 UTC, same parameters). **Error:** `{"error": {"code": 400, "message": "Request contains an invalid argument.", "status": "INVALID_ARGUMENT"}}` **What** **we’ve** **confirmed:** - The same raw request via curl to the REST API **succeeds** (with responseModalities: [“TEXT”, “IMAGE”] and imageConfig: {aspectRatio: “16:9”, imageSize: “2K”}) - The Python SDK (google-genai) with the same parameters returns 400 - No code changes between working and broken states - Affects ALL requests: text-to-image (no images), image editing (2-5 reference images) - Same API key, same parameters, same prompts that worked at 08:04 UTC **Request** **config** **(Python** **SDK):** ```python config = types.GenerateContentConfig( response_modalities=["TEXT", "IMAGE"], image_config=types.ImageConfig( aspect_ratio="16:9", image_size="2K" ) ) ``` ```python response = client.models.generate_content( model="gemini-3-pro-image-preview", contents=prompt, config=config ) ``` **Question:** Since the raw REST API works but the Python SDK doesn’t, is the SDK injecting additional parameters (e.g., thinking config) that are incompatible with image generation models? Was there a server-side change around 11:00 UTC today that tightened parameter validation? ---
**Author:** ![xsyjeon](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/xsyjeon/32/93041_2.png) [@xsyjeon](https://discuss.ai.google.dev/u/xsyjeon)\ **Post date:** [February 13, 2026, 1:06pm UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/2 "2026-02-13T13:06:10Z")
**Reproduction** **—** **same** **code,** **different** **results:** Running the exact same request 5 times in a loop: from google import genai from google.genai import types client = genai.Client(api\_key=“…”) config = types.GenerateContentConfig( ``` response_modalities=\["TEXT", "IMAGE"\], image_config=types.ImageConfig(aspect_ratio="16:9", image_size="2K") ``` ) response = client.models.generate\_content( ``` model="gemini-3-pro-image-preview", contents="A red cat on a fence", config=config ``` ) Results: SUCCESS, **FAILED** **(400)**, SUCCESS, SUCCESS, SUCCESS Same API key, same parameters, same prompt, same SDK version — 1 out of 5 fails randomly. **What** **we’ve** **confirmed:** - The identical payload via curl to the REST API succeeds consistently - The Python SDK sends the exact same JSON body (verified by intercepting httpx.Client.send) - Both text-to-image (no reference images) and image editing (with reference images) are affected - No code changes between working and broken states - Over 50 failed requests logged between 08:00–12:30 UTC, all with the same generic error message - The error message provides no detail on which argument is invalid **Impact:** In production, this caused a ~100% failure rate for all gemini-3-pro-image-preview requests for over 2 hours (11:10–12:30 UTC), affecting both text-to-image and image editing workflows. **Request:** Could the team investigate what changed server-side around 08:00–11:00 UTC on Feb 13? The intermittent nature of the 400 on identical requests suggests a backend validation issue rather than an actual invalid argument. More detailed error messages (specifying which argument is invalid) would also help greatly with debugging. ---
**Author:** ![Logan\_Kilpatrick](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/logan_kilpatrick/32/129_2.png) [@Logan\_Kilpatrick](https://discuss.ai.google.dev/u/Logan_Kilpatrick)\ **Post date:** [February 13, 2026, 5:59pm UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/3 "2026-02-13T17:59:10Z")
Flagged to the team! ---
**Author:** ![phillip\_alvelda](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/phillip_alvelda/32/93199_2.png) [@phillip\_alvelda](https://discuss.ai.google.dev/u/phillip_alvelda)\ **Post date:** [February 16, 2026, 8:27pm UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/4 "2026-02-16T20:27:47Z")
**Related issue: “Quota tier: Unavailable” blocking `gemini-3-pro-image-preview` entirely** **Environment:** - Model: `gemini-3-pro-image-preview` - API Key: from Google AI Studio - Billing: Active ($64.75 spent last 28 days) **Issue:** Google AI Studio shows **“Quota tier: Unavailable”** for `gemini-3-pro-image-preview`. All requests return `429 RESOURCE_EXHAUSTED` with: ```auto ERROR:quota_project_id: Image output modality is not allowed... ``` This is different from the intermittent 400 INVALID\_ARGUMENT above — our requests fail 100% with quota errors despite having active billing and successful usage of other Gemini models (Flash 2.0, etc.). **What we’ve confirmed:** - Billing is active and verified - Other Gemini models work fine (Gemini 2.5 Flash, etc.) - The quota console shows “Unavailable” for the image preview tier — not 0, not a number, literally “Unavailable” - Cannot increase quota through Cloud Console (no request button available) **Impact:** All image generation via `gemini-3-pro-image-preview` blocked. Production workflows down. **Request:** Could the team check if there’s a tier/quota issue affecting projects or API keys for the experimental image models? The “Unavailable” status in AI Studio suggests this may be a backend configuration issue rather than usage limits. Thanks for investigating! @Logan_Kilpatrick ---
**Author:** ![Daniel\_Bracken](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/daniel_bracken/32/100201_2.png) [@Daniel\_Bracken](https://discuss.ai.google.dev/u/Daniel_Bracken)\ **Post date:** [March 3, 2026, 3:23pm UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/5 "2026-03-03T15:23:36Z")
@Logan_Kilpatrick We’re experiencing the same issue as outlined in great detail by @xsyjeon . Are there any updates the team is able to share regarding this issue? We have since transitioned to 3.1 Flash Image Preview and encounter the same problem. ---
**Author:** ![Nackos](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/nackos/32/31930_2.png) [@Nackos](https://discuss.ai.google.dev/u/Nackos)\ **Post date:** [March 26, 2026, 7:51am UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/6 "2026-03-26T07:51:29Z")
Hello! Did you guys get any feedback regarding this? Were also experiencing this issue at our side. Br Nichlas ---
**Author:** ![Lorenzo\_Gentile](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/lorenzo_gentile/32/118610_2.png) [@Lorenzo\_Gentile](https://discuss.ai.google.dev/u/Lorenzo_Gentile)\ **Post date:** [April 28, 2026, 11:02am UTC](https://discuss.ai.google.dev/t/gemini-3-pro-image-preview-returning-400-invalid-argument-since-11-10-utc-feb-13/122920/7 "2026-04-28T11:02:52Z")
Adding fresh data on this — issue is still live. Same exact error shape on `gemini-3.1-flash-image-preview` via Vertex AI (`@google/genai` 1.45.0, `responseModalities: ["IMAGE"]`, `imageConfig: { aspectRatio: "16:9", imageSize: "1K" }`): ```json {"error":{"code":400,"message":"Request contains an invalid argument.","status":"INVALID_ARGUMENT"}} ``` I ran a small variance probe — 3 prompts × 5 attempts each, no retries — and got a counterintuitive distribution: | Prompt | 200 success | 400 INVALID\_ARGUMENT | | --- | --- | --- | | **A** — generic clean | 0/5 | **5/5** | | **B** — single trademark word | 1/5 | 4/5 | | **C** — multi-trademark loaded | 4/5 | 1/5 | The exact prompt strings: **A:** > A minimalist 3D illustration of a smartphone showing a music playlist interface, blue gradient background, floating music notes, clean modern design **B:** > A stylized iPhone displaying a clean playlist interface with album artwork, floating against a soft blue gradient background. Modern minimalist friendly style. **C:** > A photorealistic Apple iPhone 15 Pro Max running Spotify Premium with the Disney+ logo visible in the playlist. A MacBook Air on the desk and AirPods Pro nearby. > Cinematic studio lighting. Each was wrapped in a JSON envelope as the text content (~650 chars), e.g.: ```json { "task": "generate_image", "output_format": "image/png", "scene": { "description": "", "mood": "...", "style": "..." }, "critical_overrides": { "aspect_ratio": "16:9 landscape", "background": "..." }, "color_palette": { "background": "#E0F2FE", "primary": "#3B82F6", "accent": "#F59E0B" }, "technical": { "camera": {...}, "lighting": {...}, "resolution": "2048x1152" } } ``` A few observations that might help narrow it down: 1. **The 400 rate is inversely correlated with prompt complexity** — the simplest generic prompt fails 100%, the most brand-laden succeeds 80%. So it’s not a content-policy filter triggering on specific words. 2. **Every 400 returns in 10–13s** consistently. Successful generations take 22–36s. The 400 path is ~3× faster than a real generation attempt — looks like a quick reject (routing / preflight / queue) rather than a content evaluation. 3. **The 400 wire body has no `details[]`** — `code` / `message` / `status` only — so callers have nothing to dispatch on or surface to users. Happy to share the probe script if that’d help.