Historical video trial · 19 November 2026
This is an earlier attempt. Read the latest trial result.
The corrected paid attempt reached Google’s current model and received a completed text response. Our integration rejected that response during JSON parsing or evidence validation, then discarded the unvalidated text and detailed error. This is an InsightDesk integration defect, not a credential or billing setup problem.
Claude Just Changed: The 6 Big Updates You Need to Know
Duncan Rogoff | Learn Claude Code · Open original video
Recorded at . This is a saved, read-only operator report. Opening this page makes no provider request.
What the corrected attempt established
- Model:
gemini-3.5-flash-lite. API: Interactions, static video input at a requested one frame per second. - One token-count request succeeded in 179 ms and returned 48,567 input tokens.
- One generation request returned HTTP 200 with status “completed” in 9,335 ms. Total elapsed time: 9.514 seconds.
- The provider reported 42,942 generation input tokens and 1,276 output tokens, with 0 thinking tokens and 0 tool-use tokens.
- Reported input breakdown: text: 261 tokens; video: 42,681 tokens. Token counting and generation reported different input totals; neither is proof of faithful whole-video understanding.
- No automatic retry, additional video, source download, billing activation, top-up or public paid-generation endpoint was used.
What failed in InsightDesk
The response reached local validation, but no accepted analysis was saved. The runner did not retain the rejected text or the exact validation exception. We cannot identify the rejected field, repair that response offline, or independently check its speech quotations, timestamps and visual descriptions. The stateless provider request was not stored for later retrieval.
The runner now captures a bounded, key-redacted original response before decoding its envelope. Offline tests cover documented text extraction, fenced JSON, incomplete output and field-specific evidence errors. Rejected candidate fields are available only for private diagnosis; they are not a verified report. Rejected drafts are never treated as source evidence or automatically published. Tests exercise this correction without making paid requests. This does not recover the lost response or establish that a future analysis will pass.
Coverage and quality
Requested range: 0–469 seconds. Verified coverage: unknown. InsightDesk downloaded no local audio or frames. Provider video-token usage shows the submitted video input was processed, but does not independently establish which speech or visuals were understood.
Factual fidelity, timestamp accuracy, useful visual evidence, filler removal and substantive omissions could not be evaluated. No verified summary is available. Long-form, no-creator-subtitle and the other selected channel cases remain untested.
Cost limits versus actual charges
Reviewed standard prices for this model: $0.30 per million input tokens across text, image, video and audio; $2.50 per million output tokens including thinking. Pricing reference. The YouTube URL feature is documented as a no-charge preview; the estimate nevertheless reserves normal input charges rather than assuming a free invoice.
- Estimated generation charge from reported usage: $0.0160726 (about 1.6 cents). Invoice: not verified. Any charge for preflight: unknown.
- Hard request output cutoff: 4,096 tokens, including thinking. Thinking level: minimal, not a promise of zero thinking. No tools or agentic video processing.
- New preflight reserve: $0.3145728. New generation reserve: $0.3248128. Each reserves the model’s full 1,048,576-token input context separately.
- Total reserve for this corrected attempt: $0.6393856, below its $1 authorization.
- Cumulative reserve including the previous failed preflight: $0.9539584, below the $5 cumulative authorization. This is not an actual bill or remaining-credit balance.
These are conservative price-based reserves, not an account-level spending cap. The founder-reported $5 credit is not a guarantee that provider billing stops at $5.
Earlier attempt
The earlier gemini-2.5-flash-lite request failed during token counting with HTTP 404 (NOT_FOUND) in 122 ms. It made zero generation requests. Catalogue availability had been mistaken for account eligibility. Its actual charge remains unknown; $0.3145728 remains included in the cumulative reserve.
Current state
The single corrected attempt is complete and its permission is consumed. No further paid request will run automatically. Another attempt needs fresh authorization and the same cost review. The metadata form cannot resolve this failure and does not need to be retested by the founder. A successfully deployed status page is not successful video understanding.