Google: Gemini Pro Latest

Analyze this quarterly sales spreadsheet and identify the three product categories with declining revenue. Explain what might be causing each decline based on the seasonal patterns in the data.

The model would process the uploaded spreadsheet and return a structured analysis identifying Electronics (-12% QoQ), Home & Garden (-8% QoQ), and Sporting Goods (-15% QoQ) as the declining categories. For Electronics, it would note the post-holiday slump typical in Q2. For Home & Garden, it would correlate the decline with delayed spring weather patterns visible in regional breakdowns. For Sporting Goods, it would flag increased competition from a new market entrant mentioned in the notes column. The analysis would include specific row references and suggest filtering the data by region to confirm hypotheses.

Gemini Pro Latest's 1M+ token context window handles large spreadsheets without summarization loss, letting it cross-reference notes columns with numerical trends. The multimodal file input means no manual data extraction. However, at $12/Mtok output, verbose analyses of complex datasets get expensive quickly compared to text-only alternatives.

I'm attaching a 40-minute recorded team meeting (video). Summarize the three action items assigned, who owns each one, and flag any points where team members disagreed on priority.

The model would process the video file and produce a meeting summary: Action items include (1) Sarah finalizing the Q3 budget by Friday, (2) Marcus scheduling vendor demos for the new CRM by month-end, and (3) the design team delivering mockups by Tuesday. It would note that at the 18-minute mark, Sarah and Marcus disagreed on whether the CRM evaluation should precede or follow the budget lock—Sarah argued budget constraints should guide vendor selection, while Marcus wanted feature requirements defined first. The model would timestamp each action item's mention.

Native video processing eliminates transcription steps, and the extended context means the entire 40-minute meeting stays in working memory for cross-referencing. This showcases the model's multimodal strength. The trade-off: video processing burns tokens faster than text transcripts, making this approach costly for routine meeting notes compared to audio-only pipelines.

Review this Python codebase (12 files, ~3,000 lines) and explain how the authentication flow works from login request to session creation. Include which files handle each step and where potential security issues might exist.

The model would trace the authentication flow across the uploaded files: Login requests hit `api/auth.py` (lines 45-89), which validates credentials against `models/user.py`. Successful validation triggers `services/session.py` to generate a JWT token using the secret in `config/settings.py`. The token gets stored in Redis via `cache/session_store.py` with a 24-hour TTL. It would flag two concerns: the JWT secret is hardcoded rather than environment-variable-based (security risk), and `session.py` lacks rate limiting on token generation (potential DoS vector). The explanation would reference specific line numbers and function names.

The 1M token context ingests the entire codebase at once, enabling whole-system reasoning that catches cross-file security patterns. This beats chunked analysis for architectural questions. The limitation: at $2/Mtok input, repeatedly uploading large codebases for iterative questions costs more than maintaining a dedicated code-analysis tool with persistent context.

Gemini Pro Latest for low-frequency video analysis with image and audio

A 5-person brand team audits 40 hours of customer testimonial video each quarter to tag themes, sentiment, and visual brand consistency. Gemini Pro Latest processes video natively, so you skip transcription services and frame-extraction scripts—upload the MP4, prompt for themes, and get structured output. At $2/Mtok input, a 2-hour video (roughly 200k tokens of multimodal data) costs $0.40 to ingest; a 5k-token summary costs $0.06 to generate. The 1M-token context means you can batch 10 videos in one prompt for cross-video pattern detection. The trade-off: this only pencils if you run the audit quarterly or less; weekly video analysis at scale will push output costs past $50/month, at which point a cheaper text-only model with separate transcription is faster and cheaper. For infrequent, high-context video work, this is the right tool.