Google: Gemini 3.6 Flash

Analyze this 45-minute earnings call transcript and extract the three most significant forward-looking statements, with timestamps and confidence levels for each claim.

In this illustrative example, the model would parse the full transcript and return: (1) at 12:34, the CFO projects 18-22% revenue growth in Q3, citing new enterprise contracts — high confidence based on signed LOIs; (2) at 28:17, the CEO mentions exploring M&A in the logistics space, no timeline given — speculative, medium confidence; (3) at 39:02, the COO commits to reducing operational costs by 15% through automation by year-end — high confidence, references existing pilot data. Each entry would include the exact quote and a two-sentence risk assessment.

With a 1M+ token context window, Gemini 3.6 Flash can ingest hour-long transcripts in a single prompt without chunking. The $0.75/Mtok input price makes this economical for batch document analysis. Trade-off: without public benchmarks, accuracy on nuanced financial language or domain-specific jargon remains unverified in production scenarios.

I'm attaching a 90-second product demo video. Write a technical FAQ covering the five questions a developer would ask after watching, with timestamps for where each answer appears in the video.

In this illustrative example, the model would process the video and generate: Q1: What API rate limits apply? (See 0:34, mentions 1000 req/min on Pro tier). Q2: Does it support webhooks or polling only? (See 1:12, shows webhook config UI). Q3: What's the cold-start latency? (Not addressed in video — assume 200-500ms based on similar services). Q4: Can I self-host? (See 0:48, SaaS-only currently). Q5: What's the retry behavior on 5xx errors? (See 1:22, exponential backoff with 3 attempts). Each answer would cite the video timestamp or flag gaps in the demo.

Native video understanding removes the need for separate transcription or frame-extraction steps. The multimodal input (video + text prompt) fits within the 1M token budget, making this viable for tutorial indexing or support automation. Trade-off: video processing quality and frame-sampling strategy are undocumented, so results on fast-cut or low-contrast footage may vary.

Here's a 300-page PDF of our internal API documentation. Generate a migration guide for developers moving from v2 to v3, focusing on breaking changes and required code updates.

In this illustrative example, the model would scan the full PDF and produce a structured guide: Section 1 lists deprecated endpoints (e.g., /users/legacy replaced by /accounts/users), with side-by-side code samples in Python and JavaScript. Section 2 covers authentication changes (OAuth2 now required, API keys sunset June 2025). Section 3 highlights new required fields in POST /orders (shipping_method now mandatory). Section 4 provides a regex-based search-and-replace checklist for common patterns. The guide would include a risk matrix ranking changes by migration effort (low/medium/high).

The 1M token context window handles large PDFs without splitting, preserving cross-references between sections. At $0.75/Mtok input, processing a 300-page doc costs roughly $0.15-0.30 depending on formatting. Trade-off: without file-parsing benchmarks, table extraction or diagram interpretation accuracy is unknown — complex visual layouts may require preprocessing.

When your team needs to digest PDFs, videos, and audio in one pass

A 4-person product team ships weekly feature briefs by pulling insights from customer calls (audio), competitor demos (video), and internal docs (PDFs). Gemini 3.6 Flash handles all three modalities in a single prompt at $0.75/Mtok input, so you skip the transcription-then-summarization pipeline that doubles your token spend on other models. The 1M-token context window means you can drop 20+ customer calls and a dozen PDFs into one request without chunking. If your synthesis workload stays under 500K tokens per run and you're comparing across formats (not just text), this is the cheapest multi-modal option that won't force you into separate API calls. Over that threshold or if you need cited sources at the paragraph level, consider a model with stronger retrieval benchmarks.

Why this model works for support teams routing 200+ screenshots daily

A 12-person customer success team gets 300 bug reports a day, half with screenshots or screen recordings. Gemini 3.6 Flash reads the image, extracts error codes, and routes tickets to the right eng pod in under 2 seconds per request. At $0.75 input per Mtok, processing 300 images with 500-token prompts costs about $0.11/day—10x cheaper than GPT-4o for this narrow task. The trade-off: no public benchmarks means you can't predict accuracy on edge cases (blurry screenshots, non-English UI text) without a pilot. Run a 100-ticket test batch first. If your error-code extraction accuracy hits 92%+, deploy it; if not, the cost savings won't offset the manual review overhead.

When your exec team needs 90-minute board meeting summaries by end-of-day

A 3-person ops team records quarterly board meetings (90 minutes, 1080p) and needs action-item summaries within 4 hours. Gemini 3.6 Flash ingests the full video file without pre-chunking, pulls speaker-attributed quotes, and flags unresolved questions in one prompt. The 1M-token context fits the entire meeting plus your 2,000-token summary template. Output cost is $3.75/Mtok, so a 50K-token summary runs $0.19—cheap enough to generate three drafts with different focus areas (finance, product, hiring) and let the CEO pick one. The risk: without video-benchmark data, you're trusting the model to catch nuance in off-camera remarks or overlapping speakers. If your meetings are structured (slide decks, clear agendas), this works. If they're freeform strategy debates, test it on two past recordings before committing.