videoMetadata` clipping on YouTube URLs no longer clips **audio** (frames still clipped) — ~20× token inflation on long videos

Gemini API
1

**Title:** Regression: `videoMetadata` clipping on YouTube URLs no longer clips **audio** (frames still clipped) — ~20× token inflation on long videos

**Environment**

  • API: `generativelanguage.googleapis.com` `v1beta` `generateContent`（and Batch API — both affected identically）
  • SDK: `google-genai` 2.16.0 (Python)
  • Models tested: `gemini-3.1-flash-lite`, `gemini-3.7-flash` — identical behavior
  • Two API keys from two different GCP projects — identical behavior
  • Regression window: between **2026-08-18 06:00 UTC** (last known-good production run) and **2026-08-19 12:00 UTC** (first bad run). No client-side changes in between (same SDK version, same request builder).

**Summary**

When passing a YouTube URL via `FileData(file_uri=…)` together with
`VideoMetadata(start_offset=…, end_offset=…)`, the clipping interval is now
applied to **video frames only**. The **audio track of the entire video** is
included in every request regardless of the offsets. Until ~Aug 18 the same
requests were correctly clipped to the window (both frames and audio).

Uploading the same media through the **File API is not affected** — offsets
clip both frames and audio correctly there.

**Repro (public 19s video, `media_resolution=LOW`)**

```python
from google import genai
from google.genai import types

client = genai.Client(api_key=“…”)
part = types.Part(
file_data=types.FileData(file_uri=“xxx”), # 19s
video_metadata=types.VideoMetadata(start_offset=“5s”, end_offset=“10s”), # 5s window
)
resp = client.models.generate_content(
model=“gemini-3.1-flash-lite”,
contents=[types.Content(parts=[part, types.Part(text=“Describe what you see and hear.”)])],
config=types.GenerateContentConfig(
media_resolution=types.MediaResolution.MEDIA_RESOLUTION_LOW),
)
print(resp.usage_metadata)
```

  • **Expected** (5s window @ LOW): ≈ 5 s × (66 frame-tok + audio-tok) ≈ **~500 tokens**
  • **Actual**: `prompt_token_count=816`, `VIDEO=794` ≈ 5 s × 66 **＋ 19 s × 25**
    → window frames ＋ **full-video audio**

**The arithmetic holds exactly across every video we tested:**

Video (public) Length Window VIDEO tokens = window×66 + full×25
jNQXAC9IVRw 19 s 5 s 794 330 + 464 ✓
dQw4w9WgXcQ 213 s 10 s 5,989 660 + 5,325 ✓
(livestream VOD) 7,650 s 60 s 195,214 3,960 + ~191,250 ✓
(livestream VOD) 7,650 s 10 s 191,914 660 + ~191,250 ✓

Note the last two rows: shrinking the window from 60 s to 10 s removes exactly
50 × 66 frame tokens — offsets *are* reaching the server and *are* applied to
frames. Only the audio ignores them.

**Control — File API is clean:** the same 10 s window on a 60 s file uploaded
via the File API yields `VIDEO=910` ≈ 10 s × 91 — frames *and* audio clipped.

**Ruled out:** model (lite = flash, token-identical), video source (2005 upload,
music video, livestream archives — all affected), explicit `mime_type` on
`FileData` (with/without — identical), API key / GCP project (two projects —
identical).

**Impact**

  • Windowed pipelines over long videos pay ~20× input tokens per call
    (e.g. 198k instead of ~9k for a 90 s window on a 2 h video), silently.
  • Worse than cost: judgment corruption. Prompts scoped to a window
    (“evaluate this moment”) now receive the entire stream’s audio as context,
    which changes model verdicts. Our detection pipeline’s verification stage
    went from a ~19 % rejection rate to rejecting **101/101** candidates,
    because “is this second a top-10 moment?” was suddenly evaluated against
    the whole 2-hour stream it could hear.

**Questions**

  1. Is this an intentional behavior change or a regression? We could not find
    any changelog entry.
  2. We notice clipping intervals are documented under the Legacy
    generateContent docs but absent from the Interactions API docs — is
    clipping being deprecated? If so, what is the recommended way to analyze a
    time window of a YouTube video without paying full-video audio tokens?

repro.py（附件用，完整可跑版）

```python
“”“Repro: YouTube URL + videoMetadata offsets → full-video audio tokens.
Usage: GEMINI_API_KEY=… python repro.py”“”
import os
from google import genai
from google.genai import types

client = genai.Client(api_key=os.environ[“GEMINI_API_KEY”])
CASES = [
(“xxxx_url”, 19, (5, 10)),
(“xxxx_url”, 213, (10, 20)),
]
for url, dur, (a, b) in CASES:
part = types.Part(
file_data=types.FileData(file_uri=url),
video_metadata=types.VideoMetadata(start_offset=f"{a}s", end_offset=f"{b}s"))
resp = client.models.generate_content(
model=“gemini-3.1-flash-lite”,
contents=[types.Content(parts=[part, types.Part(text=“Describe what you see and hear.”)])],
config=types.GenerateContentConfig(
media_resolution=types.MediaResolution.MEDIA_RESOLUTION_LOW))
um = resp.usage_metadata
win = b - a
print(f"{url} len={dur}s window={win}s"
f" prompt_tokens={um.prompt_token_count}"
f" expected≈{win*98} got_formula≈{win*66 + dur*25}")