Google: Gemini 3.1 Pro Preview Custom Tools

Analyze this quarterly sales spreadsheet and create a Python script that identifies underperforming regions, then draft an email to the regional managers explaining the findings.

This example would demonstrate the model processing a multi-sheet Excel file, extracting sales data across regions, and producing both a working Python analysis script and a professionally-toned email. The script would include pandas operations for aggregation and filtering, with inline comments explaining the logic. The email would reference specific numbers from the analysis, maintain appropriate business tone, and suggest actionable next steps—all generated in a single conversational turn without requiring the user to switch contexts between coding and writing tasks.

Showcases the 1M+ token context window handling large file uploads and the multimodal capability to reason across structured data and natural language simultaneously. The custom tools feature would allow this model to execute the Python script and verify results before finalizing the email, though latency at $12/Mtok output makes iterative refinement expensive for budget-conscious teams.

Watch this 8-minute product demo video and generate a technical FAQ covering the five most complex features shown, with timestamps linking back to relevant moments.

This example would show the model ingesting video content, identifying key technical demonstrations (API integration sequences, configuration workflows, error-handling scenarios), and producing a structured FAQ document. Each answer would include a timestamp reference like "See 3:42 for the webhook setup process" and explain the feature in clearer terms than the presenter used. The model would distinguish between features that were clearly explained versus those that need additional clarification, organizing answers by complexity rather than chronological order.

Highlights native video understanding without requiring transcription preprocessing. The model's ability to cross-reference visual demonstrations with spoken explanations makes it useful for documentation teams working from recorded content. However, without public benchmarks, accuracy on domain-specific technical content (medical devices, financial systems) remains unvalidated compared to specialized alternatives.

I'm debugging a React app where the useEffect hook fires twice on mount in development. Here's my component code, console logs, and a screenshot of the Network tab showing duplicate API calls. What's happening?

This example would demonstrate the model analyzing code, log output, and a browser screenshot together to diagnose the issue. The response would explain React 18's Strict Mode behavior causing intentional double-mounting in development, reference the specific lines in the provided code where side effects occur, and point to the duplicate POST requests visible in the Network tab screenshot. It would then provide a corrected code snippet using a ref to prevent duplicate API calls, explaining why this pattern works and when the double-mount behavior disappears in production builds.

Demonstrates multimodal debugging where code, logs, and visual browser state must be synthesized. The audio input capability (not used here) would theoretically allow developers to verbally walk through reproduction steps. The $2 input pricing makes uploading screenshots and large codebases economical, though the model's code-specific performance relative to Claude 3.5 Sonnet or GPT-4 remains unclear without published coding benchmarks.

When support tickets arrive as screenshots, PDFs, and voice clips

A 12-person SaaS support team gets 200+ tickets daily—half arrive as phone recordings, annotated screenshots, or scanned contracts. Gemini 3.1 Pro Preview Custom Tools handles all four input types (text, audio, image, file) in a single API call, so you skip the pre-processing pipeline that costs you 90 seconds per mixed-media ticket. At $2/$12 per Mtok, a 500-token triage prompt with a 2MB image costs roughly $0.03, which pencils if your team's hourly rate exceeds $40. The 1M-token context window means you can dump an entire 40-page PDF contract into the prompt without chunking. If your tickets are text-only or you're processing under 50/day, Claude Sonnet 4 at $3/$15 will save you the custom-tools learning curve.

Diffing 300-page agreements without chunking or retrieval overhead

A 4-attorney firm reviews merger agreements that average 80,000 words each. Gemini 3.1 Pro Preview's 1M-token window fits two full contracts plus a 5,000-word diff prompt in one pass—no vector store, no retrieval step, no context-window juggling. You paste both PDFs (via file input), ask for clause-level changes, and get a structured response in 15-20 seconds. At $2 input per Mtok, a 200k-token comparison costs $0.40, versus $60 of associate time for the same first-pass review. The trade-off: no public benchmarks yet, so you'll want to spot-check the first 20 diffs against human review before trusting it on high-stakes deals. If your agreements are under 30 pages, GPT-4o at $2.50/$10 is cheaper and has published MMLU scores.

Scanning user-uploaded videos for policy violations at scale

A 20-person social app moderates 800 user videos daily (30-90 seconds each). Gemini 3.1 Pro Preview ingests video natively, so you skip the frame-extraction step that adds 12 seconds and $0.08 per video in your current pipeline. You send the raw MP4, a 200-token policy checklist, and get a violation report in one call. At $2/$12 per Mtok and an average 8,000-token video encoding, each moderation costs roughly $0.12 in API fees—half your current spend when you factor in the frame-extraction overhead. The 1M-token context means you can batch-review a user's last 10 uploads in a single prompt to catch repeat offenders. If you're under 200 videos/day, GPT-4o Vision at $2.50/$10 is simpler to integrate and has stronger benchmark coverage for edge cases.