Google: Gemini 3.5 Flash (batch)

Analyze this 45-minute earnings call transcript and extract: (1) three key revenue drivers mentioned, (2) any forward guidance changes, (3) management sentiment shifts compared to last quarter. Flag contradictions.

This example would produce a structured analysis identifying revenue drivers like "cloud services growth at 28% YoY" and "enterprise customer expansion in APAC," alongside specific guidance revisions with exact figures cited from the transcript. The model would highlight a sentiment shift from cautious to optimistic regarding regulatory headwinds, noting the CFO's language change from "monitoring closely" to "confident in our compliance posture." Any contradictory statements between the prepared remarks and Q&A section would be flagged with timestamps and quoted verbatim for verification.

The 1M+ token context window makes this model particularly effective for long-document analysis where other models would require chunking strategies. Batch processing at $0.75/Mtok input makes high-volume document workflows economically viable. Trade-off: batch mode introduces latency (results aren't real-time), so this suits overnight processing pipelines rather than interactive analysis.

I'm attaching a 12-minute product demo video. Create a technical FAQ covering the five features shown, written for developer documentation. Include timestamps for where each feature appears.

This example would produce a developer-focused FAQ with entries like "Q: How does the webhook retry mechanism handle exponential backoff?" followed by technical explanations derived from the video content at timestamp 3:42. Each answer would reference specific UI elements or code snippets visible in the demo, maintaining technical accuracy about API endpoints, configuration options, and error handling patterns demonstrated. The output would preserve the logical flow of the demo while reorganizing content into a searchable Q&A format.

Native video understanding eliminates the transcription-then-analyze workflow, directly processing visual and audio information together. The multimodal capability handles on-screen code, UI interactions, and spoken explanations simultaneously. Trade-off: batch processing means you can't iteratively refine the FAQ in real-time—submit the job, wait for results, then revise if needed.

Compare these three architectural diagrams (images attached) and write a 2-page technical memo recommending one approach for a distributed caching layer. Consider latency, consistency guarantees, and operational complexity.

This example would produce a structured technical memo analyzing each architecture's trade-offs: evaluating the first diagram's eventual consistency model against the second's strong consistency with higher latency, while noting the third approach's operational overhead from additional coordinator nodes. The recommendation would cite specific diagram elements—"the write-through cache in Option B"—and quantify trade-offs using the visual information provided, such as hop counts between nodes or replication paths. The memo would conclude with a justified recommendation tied to specific use-case requirements.

Image understanding allows direct reasoning over technical diagrams without manual annotation or description. The model can compare visual architectures, count components, and trace data flows—tasks that pure text models require you to describe exhaustively. Trade-off: at $4.50/Mtok output, verbose responses on large batch jobs add cost; tighter prompt constraints help manage spend.

When batch pricing makes high-volume moderation affordable at scale

A 12-person social platform runs nightly moderation sweeps on 200k user posts—text, images, and short video clips. Gemini 3.5 Flash batch mode costs $0.75/Mtok input and $4.50/Mtok output, which translates to roughly $15-25 per night for the entire queue when you're processing mixed media at that volume. The 1M token context window means you can bundle multiple posts with shared context (user history, community rules) in a single call, cutting API overhead by 40-60% compared to per-post requests. The trade-off: batch jobs run asynchronously with up to 24-hour turnaround, so this only works if you can tolerate morning delivery instead of real-time flags. If you need sub-minute moderation, pay the premium for synchronous Flash. For overnight sweeps where cost per item matters more than latency, batch mode wins.

How 1M token context handles cross-document legal review in one pass

A 4-attorney firm reviews SaaS vendor contracts—master agreement, three exhibits, a data processing addendum, and prior correspondence—totaling 180k tokens. Gemini 3.5 Flash batch mode ingests all six files in a single context window (1M tokens) and returns a structured risk summary with clause-level citations. At $0.75/Mtok input, the entire review costs under $0.20, and the multimodal support means scanned PDFs or image-based exhibits process without OCR preprocessing. The limitation: no public benchmarks exist for this model yet, so you're flying blind on accuracy compared to GPT-4 or Claude on legal reasoning tasks. Run a 20-contract pilot with manual spot-checks before committing production workflows. If the output quality holds and you're processing 500+ contracts monthly, the cost advantage over synchronous alternatives pays for the validation effort in week one.

When native video input eliminates your transcription middleware layer

A 9-person e-learning startup processes 400 hours of recorded lectures monthly—extracting key concepts, generating quiz questions, and tagging topics for search. Gemini 3.5 Flash batch mode accepts video files directly (no separate transcription API), runs overnight, and outputs structured JSON at $4.50/Mtok. A 90-minute lecture averages 60k output tokens for a full summary with timestamps, costing roughly $0.27 per video. The native video modality cuts your stack complexity: no Whisper preprocessing, no alignment logic, no S3 staging for intermediate transcripts. The catch: batch latency means same-day lecture uploads won't have summaries until the next morning, so this works for catalog content but not live-class recaps. If your content pipeline tolerates 12-24 hour lag and you're spending more than $200/month on transcription services, switch to batch mode and pocket the difference.