Google: Gemini 3 Flash Preview

Analyze this 45-minute earnings call transcript and extract the three most significant forward-looking statements, then flag any contradictions with prior quarter guidance.

The model would process the full transcript within its 1M-token context window and return a structured analysis identifying key forward-looking statements about revenue guidance, product launches, and market expansion. It would cross-reference these against historical statements, highlighting a discrepancy where Q2 guidance suggested 15% growth but Q3 now projects 8-12%. The response would include exact timestamps and quote the relevant passages verbatim, maintaining the nuance of hedging language used by executives.

This showcases Gemini 3 Flash Preview's massive context window handling long documents in a single pass, plus its multimodal capability if the transcript includes audio. The 'Preview' designation means production stability isn't guaranteed — expect occasional inconsistencies in citation formatting or edge-case hallucinations on numerical comparisons.

Here's a video of our manufacturing line. Identify safety violations, estimate cycle time per unit, and suggest two efficiency improvements based on worker movement patterns.

The model would analyze the video input and produce a timestamped report noting three safety issues: missing hard hats at 0:47, unsecured cable routing at 1:23, and blocked emergency exit at 2:15. It would calculate an average cycle time of 47 seconds per unit based on tracking worker hand movements across 12 complete cycles. Efficiency suggestions would focus on repositioning the parts bin 0.8 meters closer to reduce reach time, and implementing a gravity-feed system for the fastener tray to eliminate restocking delays.

Demonstrates native video understanding without requiring frame extraction or separate vision API calls. The Flash variant prioritizes speed over the deepest reasoning, so complex spatial relationships or subtle defects might be missed compared to the full Gemini 3 model. Useful for rapid triage, less reliable for final audit reports.

Convert this Python data pipeline into idiomatic Rust with proper error handling. Preserve the async behavior and add inline comments explaining ownership decisions.

The model would rewrite the Python code into Rust, replacing dictionary lookups with Result<T, E> enums, converting async/await syntax to Tokio futures, and wrapping file I/O in proper error propagation chains. Inline comments would explain why certain values are moved versus borrowed, where Arc<Mutex<T>> is necessary for shared state across tasks, and how the borrow checker enforces thread safety that Python's GIL handled implicitly. The output would compile without warnings and maintain the original pipeline's logical flow.

Gemini 3 Flash Preview handles cross-language translation well, especially for modern languages in its training data. The $3/Mtok output pricing makes it 6× more expensive than input for code generation tasks — fine for one-off conversions, but consider caching or cheaper models for iterative refactoring workflows where you're regenerating similar code blocks.

When Gemini 3 Flash handles support tickets with screenshots and voice clips

A 12-person SaaS support team gets 200+ tickets daily mixing text descriptions, error screenshots, screen recordings, and occasional voice memos from mobile users. Gemini 3 Flash Preview processes all five modalities in a single call at $0.50/$3.00 per Mtok, letting you route tickets without pre-processing files into separate text summaries. The 1M token context means you can include the last 30 days of conversation history for repeat customers without chunking. If your tickets average under 50K tokens each and you're running classification + sentiment + urgency scoring, you'll spend roughly $0.15-0.45 per ticket at current rates. This works until you hit 500+ tickets daily or need sub-200ms response times—then the $3.00 output cost starts hurting and you should test a text-only model for the 60% of tickets that don't actually need multimodal.

Why Gemini 3 Flash wins on full-contract review without chunking

A 4-attorney firm reviews 15-20 commercial leases monthly, each running 80-150 pages with exhibits. Gemini 3 Flash Preview's 1M token window fits an entire lease plus rider documents in one prompt, so you extract key terms, flag non-standard clauses, and generate redline summaries without splitting the file or losing cross-references between sections. At $0.50 input per Mtok, a 120-page lease (roughly 180K tokens) costs $0.09 to ingest; the 8K token summary output runs $0.024. Total per-document cost sits around $0.12, and you avoid the engineering overhead of chunking strategies or vector search. This holds until you're processing 100+ documents weekly—then batch discount models or fine-tuned smaller models start penciling out better on pure economics.

When Gemini 3 Flash moderates live-stream clips frame-by-frame

A 20-person community platform moderates 300 hours of user-uploaded video weekly, flagging policy violations before content goes live. Gemini 3 Flash Preview ingests video directly, so you send 30-second clips (sampled at 1 fps = 30 frames) and get back violation scores, timestamp markers, and severity ratings in one call. The multimodal input saves you from running separate vision + audio pipelines. At $0.50/$3.00 per Mtok, a 30-second clip costs roughly $0.03-0.08 depending on resolution and output verbosity. You're spending $25-40 per 100 hours moderated, which beats hiring contract moderators until you cross 1,000+ hours weekly. Past that threshold, the output token cost ($3.00/Mtok) becomes the bottleneck and you should test a vision-specialist model with cheaper generation.