Google: Gemini 3.1 Flash Lite (batch)

Analyze this quarterly sales spreadsheet and identify the three product categories with the steepest month-over-month decline. Explain possible causes based on the data patterns you see.

In this illustrative example, the model would parse the uploaded spreadsheet and return a structured analysis: "Electronics (-18% MoM), Home Goods (-14% MoM), and Apparel (-11% MoM) show the steepest declines. Electronics dropped sharply in week 3, coinciding with a competitor launch visible in your notes column. Home Goods declined steadily across all regions, suggesting a category-wide issue rather than geographic. Apparel's decline concentrated in sizes M-L, while XS-S held steady—possible inventory or targeting mismatch." The response would reference specific cells and include a brief recommendation section.

Flash Lite's 1M-token context window handles large spreadsheets without chunking, and the batch API makes this analysis cost-effective at $0.13/Mtok input. The model connects data patterns to business context when that context exists in the file. Trade-off: without public benchmarks, reasoning depth on complex financial models remains unverified.

Transcribe this 40-minute team meeting recording and create a summary with action items, decisions made, and unresolved questions. Tag each item with the speaker's name.

In this illustrative example, the model would process the audio file and return: a full transcript with speaker labels, followed by a structured summary. "Decisions: Migration to new CRM approved (Sarah, 12:34); Q3 budget increased 15% (Marcus, 28:10). Action items: Dev team to audit API endpoints by Friday (Priya); Marketing to draft launch email (Jordan, due Monday). Unresolved: Whether to sunset legacy dashboard—Sarah and Marcus disagreed, tabled for next week." The output would preserve technical terms and acronyms accurately.

Audio processing in a single API call simplifies workflows that previously required separate transcription services. The batch pricing model makes long-form audio economical for internal tooling. Trade-off: speaker diarization accuracy degrades in crosstalk-heavy sections, and the model may miss nuance in sarcasm or tone without explicit verbal cues.

Review this 20-page product requirements document and this 8-minute demo video. List any features shown in the video that aren't documented in the PRD, and flag any PRD requirements that seem unaddressed in the demo.

In this illustrative example, the model would cross-reference both inputs and return: "Undocumented features in demo: dark mode toggle (0:45), bulk export to CSV (3:12), inline comment threading (6:30). PRD requirements not shown: role-based permissions (section 4.2), audit log (section 5.1), API rate limiting (section 7.3). Note: The demo's search functionality appears more advanced than the 'basic keyword search' specified in PRD section 3.4—possible scope creep or outdated documentation."

Multimodal context—text document plus video—lets the model catch specification drift that manual review often misses. The 1M-token window accommodates lengthy PRDs without summarization loss. Trade-off: the model interprets visual UI elements but can't verify backend implementation details like rate limiting that aren't visually demonstrated.

When Flash Lite wins on overnight content review queues

A 4-person community team running a Discord with 12k members needs to review 800-1200 flagged messages every morning. Gemini 3.1 Flash Lite in batch mode costs $0.13/Mtok input and runs overnight, so you're reviewing text + image attachments for $0.40-0.60/day instead of $3-5 with real-time models. The 1M token context means you can load the last 48 hours of channel history for repeat-offender detection without chunking. Accuracy on clear-cut violations (spam, slurs, explicit content) matches pricier models; edge cases (sarcasm, context-dependent insults) need human review either way. If you need sub-10-second response times or are processing under 200 items/day, pay for real-time. Otherwise, queue it and let batch pricing do the work.

When to batch-process 500-page RFPs with multimodal input

A 9-person procurement consultancy receives 40-80 RFP documents per week (PDFs with tables, diagrams, and 200-600 pages each). They need to extract requirements, flag compliance gaps, and route sections to specialists. Flash Lite's file + image modalities let you ingest the full PDF without pre-processing, and the 1M context window handles even the longest documents in one pass. At $0.13/Mtok input, a 500-page RFP with embedded images costs $1.20-1.80 to process versus $8-12 on real-time multimodal models. Batch mode adds 6-12 hour latency, but RFPs arrive Monday-Thursday and get reviewed Friday morning anyway. If you're processing under 10 documents/week or need same-day turnaround, real-time makes sense. Above that threshold, batch cuts your AI budget by 75% with zero workflow change.

When batch video analysis beats real-time for internal comms

A 30-person remote startup records 8-15 hours of Zoom meetings and Loom updates every week. The ops lead needs searchable summaries and action-item extraction but doesn't need it live. Flash Lite's video modality processes the recordings overnight at $0.13/Mtok input, so 10 hours of video (roughly 600k tokens of transcript + visual context) costs $0.08 versus $0.60-1.20 on real-time video models. The 1M context window means you can process a 90-minute all-hands in one call without splitting clips. Accuracy on speaker identification and slide text is solid; nuanced tone detection (sarcasm, frustration) is weaker than premium models but fine for action-item extraction. If you need live transcription during the call or are processing under 2 hours/week, real-time is worth it. Above that, batch saves you $40-60/month per team.