Google: Gemini 3.5 Flash Lite (batch)

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

The model would process the uploaded file and return a structured analysis identifying Electronics (-18% MoM), Home Goods (-12% MoM), and Apparel (-9% MoM) as the declining categories. It would note that Electronics declined sharply in March coinciding with a competitor product launch visible in the notes column, Home Goods showed weekend sales drops suggesting inventory issues, and Apparel decline correlates with unseasonable weather data in the regional breakdown tab. The response would include specific cell references and suggest filtering the data by region to confirm the weather hypothesis.

Demonstrates multimodal file processing with the 1M+ token context window handling large spreadsheets in a single pass. The batch pricing ($0.15/$1.25 per Mtok) makes this economical for high-volume document analysis, though the 'Lite' designation suggests simpler reasoning than full Gemini models—complex financial modeling or causal inference may require verification.

I'm uploading a 40-minute customer support call recording. Summarize the main complaint, list every commitment our agent made, and flag any policy violations you detect in the conversation.

The model would transcribe and analyze the audio file, producing a summary showing the customer's complaint centered on a delayed refund (mentioned at 3:47, 12:22, and 28:15). It would extract four specific commitments: expedited processing within 48 hours, waived restocking fee, direct manager follow-up, and complimentary shipping on next order. The policy violation section would flag that the agent offered a discount exceeding the standard 15% threshold without supervisor approval at timestamp 31:40, requiring compliance review.

Showcases native audio processing across extended recordings within the 1M token context—no need to chunk the file. The batch API suits overnight processing of support call archives. However, nuanced tone analysis or detecting sarcasm may be less reliable than specialized speech models, and timestamp accuracy depends on audio quality.

Review this 8-page technical specification PDF and this 3-minute product demo video. Write a gap analysis: what features are documented but not shown, and what's demoed but missing from the spec?

The model would process both the PDF and video simultaneously, producing a two-column gap analysis. The 'Documented but not demoed' section would list API rate limiting (page 4), offline mode (page 6), and webhook retry logic (page 7)—none visible in the video walkthrough. The 'Demoed but undocumented' section would identify the dark mode toggle shown at 1:23, the bulk export feature at 2:10, and the collaborative editing shown at 2:45, none of which appear in the specification. It would recommend updating section 3.2 of the spec to match the implemented UI.

Highlights cross-modal reasoning—comparing structured text against visual demonstration within a single context window. The batch processing model makes this affordable for documentation audits at scale. The 'Flash Lite' variant may miss subtle UI details or complex logical inconsistencies that require deeper reasoning; critical compliance gaps should be human-verified.

When Gemini 3.5 Flash Lite wins on overnight tagging jobs

A 4-person compliance team needs to tag 20,000 insurance claims every night—each claim is 800-1,200 tokens, mostly boilerplate with a few critical fields. Gemini 3.5 Flash Lite (batch) hits $0.15/Mtok on input, which means tagging all 20,000 claims costs about $3.60 in input tokens plus minimal output. The 1M token context window handles even the longest claims without truncation. Batch mode adds latency (results arrive in hours, not seconds), but that's fine when you're processing overnight and reading results at 9am. If you need sub-minute turnaround or you're under 5,000 documents/day, the batch discount stops mattering and you should compare real-time alternatives. For high-volume classification where speed isn't the constraint, this is the cheapest multimodal option in the Gemini lineup.

How a 10-person trust team uses Flash Lite for flagged uploads

A social platform's trust team reviews 3,000 flagged videos per day—each video is 15-90 seconds, already transcoded to frames. They need basic yes/no moderation calls (violence, spam, policy violations) before a human escalation. Gemini 3.5 Flash Lite (batch) supports video input natively, so they send the raw video file instead of pre-extracting frames or transcripts. At $0.15/Mtok input, processing 3,000 videos costs roughly $12-18/day depending on video length. The batch API queues everything and returns results in 2-4 hours, which fits their SLA (human reviewers work the queue the next morning). If you need real-time moderation (block uploads instantly), batch latency kills you and you need a synchronous model. For deferred review workflows where cost matters more than speed, this setup works.

When a 3-person research shop batches weekly literature reviews

A biotech consultancy downloads 40-60 PDFs every Friday (clinical trial reports, patent filings, preprints) and needs a structured summary by Monday morning. Each PDF is 8,000-25,000 tokens; total weekly volume is around 800k tokens. Gemini 3.5 Flash Lite (batch) accepts file uploads directly and the 1M token context means they can bundle multiple PDFs into a single prompt without chunking. Input cost is $0.15/Mtok, so 800k tokens costs $0.12 in input—output summaries add maybe $0.50/week at $1.25/Mtok. Batch mode delivers results in 6-12 hours, which is fine for a Friday-to-Monday cadence. If you're summarizing one document at a time or need instant results, batch overhead isn't worth it. For weekly research digests where you're processing dozens of files in one shot, this is the lowest-cost multimodal play.