Google: Gemini 3.7 Flash (batch)

Analyze this 45-minute earnings call transcript and extract: key revenue drivers, management's forward guidance, and any red flags investors should watch. Prioritize quantitative statements.

In this illustrative example, the model would parse the full transcript and return a structured summary highlighting three revenue drivers (e.g., 23% YoY growth in cloud services, new enterprise contracts worth $180M ARR, geographic expansion into APAC contributing 12% of Q4 revenue). Forward guidance would be quoted directly: management expects 15-18% revenue growth next fiscal year, with operating margins improving to 28-30%. Red flags flagged might include rising customer acquisition costs (up 34% QoQ) and delayed product launches in two segments. The model would cite specific timestamps and speaker names for each claim.

The 1M+ token context window handles long-form documents without chunking, making this ideal for financial analysis where full context matters. Batch processing keeps costs low at $0.38/Mtok input. However, without public benchmarks, users can't compare reasoning accuracy against models like Claude or GPT-4 on complex inference tasks.

I'm uploading a 12-minute product demo video. Generate a shot-by-shot breakdown with timestamps, identify UI elements shown, and write a technical FAQ covering features demonstrated.

In this illustrative example, the model would process the video and return a timeline: 0:00-1:23 (login flow, OAuth integration visible), 1:24-3:45 (dashboard overview showing real-time analytics widgets), 3:46-6:10 (bulk import feature with CSV validation), and so on. For each segment, it would describe UI components (button labels, modal dialogs, navigation patterns). The FAQ section might include 8-10 entries like 'Does the platform support SSO?' (answered based on the OAuth screen at 0:45) and 'What file formats does bulk import accept?' (CSV, JSON, XML per the 4:20 timestamp). Technical details would be extracted directly from on-screen text and voiceover.

Native video understanding eliminates the need for separate transcription and frame-extraction pipelines. The model handles multiple modalities in one prompt, which simplifies workflows. Batch mode suits non-urgent video processing jobs. The $1.88/Mtok output price is higher than text-only models, so verbose responses on large video sets add cost.

Here's a 200-page PDF of our legacy Java codebase documentation from 2018. Cross-reference it with these 15 recent bug reports (attached) and propose a refactoring plan that addresses recurring issues without breaking existing integrations.

In this illustrative example, the model would ingest the full PDF and all bug reports, then identify patterns: 8 of 15 bugs stem from a deprecated authentication module (documented on pages 67-72), 4 relate to thread-safety issues in the caching layer (pages 103-108), and 3 involve API versioning mismatches. The refactoring plan would propose: (1) migrate auth to a modern OAuth2 library while maintaining backward-compatible endpoints for three legacy clients; (2) replace the custom cache with a proven concurrent data structure; (3) introduce API versioning middleware. Each recommendation would cite specific page numbers from the docs and bug report IDs, with estimated effort and risk scores.

The massive context window lets the model reason across hundreds of pages plus auxiliary files simultaneously, which is critical for legacy system analysis. Batch processing makes this economical for periodic audits. Without coding benchmarks (HumanEval, MBPP), it's unclear how the model's refactoring suggestions compare to specialist code models in correctness and idiomaticity.

When batch pricing makes large-scale document analysis viable

A 12-person legal ops team needs to extract clauses from 800 vendor contracts every month. Gemini 3.7 Flash batch mode runs at $0.38/Mtok input and $1.88/Mtok output—roughly 60% cheaper than real-time equivalents. The 1M token context window handles most contracts in a single pass without chunking. You queue the job at 5pm, results land by 7am, and the team reviews extracts over coffee. The trade-off: batch jobs can take up to 24 hours to complete, so this only works if you can tolerate overnight latency. If you need answers in under an hour, pay for real-time. If your workflow runs on a daily or weekly cadence and you're processing 50+ documents per batch, the cost savings justify the wait.

Why batch multi-modal makes sense for high-volume review workflows

A 4-person trust and safety team at a creator platform reviews 2,000 flagged posts daily—mix of text, images, and short videos. Gemini 3.7 Flash batch handles all three modalities in one API call and costs a fraction of running separate vision and text models. The 1M token window means you can bundle 20-30 posts per request without hitting limits. You run the batch overnight, export a CSV of flagged content by severity, and the team triages in the morning. The constraint: batch latency means this is a next-day workflow, not real-time moderation. If you need sub-minute decisions (live chat, payment fraud), this isn't the model. If you're reviewing user-generated content on a daily cycle and volume exceeds 500 items, batch pricing cuts your bill by half.

When to batch-process support tickets for executive reporting

A 20-person SaaS company collects 1,200 support tickets and NPS comments weekly. The ops lead needs a 2-page summary for the exec team every Monday. Gemini 3.7 Flash batch ingests the full week's text (plus attached screenshots and PDFs) in one job, runs over the weekend, and delivers a structured summary by Sunday night. At $0.38 input per Mtok, processing 400k tokens of feedback costs under $0.20. The model's file and image support means you don't need a separate OCR step for attachments. The boundary: if you need daily or intraday summaries, batch latency (up to 24 hours) breaks the cadence. If your reporting cycle is weekly or monthly and you're summarizing 200+ items, batch mode is the cheapest path to structured output.