Google: Gemini 3.6 Flash (batch)

Analyze this 45-minute product demo video and extract a structured summary: key features shown, timestamps for each demo segment, and any technical limitations mentioned by the presenter.

The model would produce a hierarchical breakdown: an executive summary paragraph, followed by a timestamped table of demo segments (e.g., '03:42 - Dashboard navigation, 12:18 - API integration walkthrough'), then a bulleted list of 8-12 features with brief descriptions, and finally a short section on limitations discussed around the 38-minute mark. The structure would be clean and scannable, with precise timestamp references that match the video's actual content.

Showcases the 1M+ token context window handling long-form video input. The batch processing mode makes this economical for bulk video analysis tasks. However, real-time use cases requiring sub-second latency would need the standard Gemini Flash variant instead.

Review this 200-page technical specification PDF for our aerospace client. Flag any sections that conflict with DO-178C software certification requirements and suggest specific remediation steps.

The model would return a compliance report identifying 6-8 conflict areas, each with: the conflicting spec section number, the relevant DO-178C clause, a plain-English explanation of the mismatch, and 2-3 concrete remediation options ranked by implementation effort. For instance, it might flag insufficient traceability documentation in Section 4.7 against DO-178C Table A-3, then suggest either enhancing the existing matrix or adopting a requirements management tool.

Demonstrates multimodal file processing with domain-specific reasoning. The $0.38/Mtok input pricing makes large document analysis viable at scale. Trade-off: batch mode introduces processing delays (typically minutes to hours), so this suits overnight report generation rather than interactive document Q&A.

Generate a Python data pipeline that ingests these 12 CSV files, normalizes timestamps across UTC and PST entries, joins on customer_id, then outputs a single Parquet file partitioned by month. Include error handling for malformed rows.

The model would produce a ~150-line script using pandas and pyarrow, with: a configurable file-path list, timezone conversion logic using pytz, an outer join strategy with explicit handling of null customer_ids, try-except blocks that log malformed rows to a separate error CSV, and partitioned write logic that creates monthly subdirectories. The code would include docstrings and inline comments explaining the UTC/PST normalization approach.

Highlights code generation with multi-file context awareness. The massive context window lets you paste all 12 CSV schemas directly into the prompt. The batch pricing model ($1.88/Mtok output) keeps costs low even for verbose, well-commented code. However, you won't get iterative refinement in real-time — submit the job, then review the complete output later.

When batch pricing makes multimodal extraction worth it at scale

A 12-person compliance team processes 800 mixed-format client submissions every night—PDFs with tables, scanned forms, video depositions with timestamps. Gemini 3.6 Flash batch mode cuts input cost to $0.38/Mtok, which matters when you're feeding 200k+ token documents through OCR-free extraction. The 1M token context window means you can drop an entire case file (contracts, emails, video transcripts) into one prompt and ask cross-reference questions without chunking. Output at $1.88/Mtok stays reasonable because you're generating structured JSON, not long-form prose. If your documents average under 50k tokens or you need sub-5-second turnaround, standard Flash is the better call. For overnight jobs where latency doesn't matter and volume is high, batch mode pays for itself after 200 documents per month.

Native video input for high-throughput safety review

A 4-person trust and safety team at a UGC platform reviews 1,200 flagged videos per day—each 30-90 seconds, mix of user reports and automated flags. Gemini 3.6 Flash batch accepts video natively, so you skip the transcription step and the model can catch visual context (gestures, on-screen text, scene changes) that audio-only models miss. At $0.38/Mtok input, processing a 60-second video with scene analysis costs roughly $0.02-0.04 depending on resolution. The batch queue adds 10-15 minutes of latency, which works fine for overnight review cycles but breaks real-time moderation. If you need decisions in under 60 seconds, you're stuck with standard Flash at 2.5x the input cost. For daily batch review where visual context matters and you're processing hundreds of videos, this is the cheapest multimodal option that doesn't require preprocessing.

Long-context synthesis when turnaround time isn't critical

A 3-person finance team at a Series B startup compiles board decks every Friday—pulling from 40+ data sources (Stripe exports, bank statements, HubSpot CSVs, Notion meeting notes). Gemini 3.6 Flash batch handles the 1M token context window, so you can load the entire week's raw data and ask for variance analysis, trend summaries, and anomaly flags in one prompt. At $0.38 input and $1.88 output per Mtok, a 300k token input with 8k token output runs about $0.13 per report. The batch queue adds 20-30 minutes, which is fine for a Friday morning job that doesn't ship until afternoon. If you need same-hour turnaround or you're generating reports more than twice a week, standard Flash justifies the 2.5x cost bump. For weekly synthesis jobs where you're trading latency for cost and context depth, batch mode is the obvious pick.