Google: Gemini 3.8 Flash

Analyze this 45-minute earnings call transcript and extract the three most significant forward-looking statements, with exact quotes and timestamps.

In a production scenario, Gemini 3.8 Flash would parse the full transcript and return structured findings like: (1) At 12:34, CFO states "We're targeting 40% gross margin expansion in Q3" — signals pricing power. (2) At 28:17, CEO mentions "Two enterprise pilots converting to full contracts in July" — near-term revenue visibility. (3) At 39:02, the phrase "regulatory approval timeline remains our biggest variable" — flags external dependency risk. Each quote would include surrounding context for verification.

The 1M+ token context window makes this model viable for analyzing hour-long transcripts or dense financial documents in a single pass, without chunking. At $0.75/Mtok input, processing a 200k-token earnings call costs ~$0.15. Trade-off: without published benchmarks, accuracy on nuanced financial language versus specialized models remains unverified.

I'm attaching a 12-page product spec PDF and a 90-second demo video. List every feature mentioned in either artifact that isn't in the other, formatted as a comparison table.

Gemini 3.8 Flash would process both the PDF text and video frames/audio, then output a table with columns: Feature | In PDF | In Video | Notes. Example rows: "Dark mode toggle" | Yes | No | Spec mentions implementation details. "Gesture-based navigation" | No | Yes | Demonstrated at 0:34 but undocumented. "API rate limits" | Yes | No | Technical detail in appendix. The model cross-references modalities to surface gaps between documentation and implementation.

Multimodal input (text, image, video, audio, file) in one request eliminates the need for separate processing pipelines. This is useful for QA workflows where specs and demos diverge. Trade-off: video understanding quality isn't benchmarked here, so frame-level accuracy for fast motion or small UI elements is uncertain.

Convert this 8,000-line Python monorepo into a dependency graph showing which modules import which, then suggest three refactoring moves to reduce circular dependencies.

In practice, the model would generate a visual representation (described textually): "Core modules: auth.py → database.py → models.py → auth.py (circular). API layer: routes.py imports 14 modules. Utils: helpers.py imported by 22 files." Then: Refactor 1: Extract shared types from models.py into types.py to break the auth→database→models→auth loop. Refactor 2: Move database connection logic from database.py to a new db_client.py so models.py doesn't need the full database module. Refactor 3: Split helpers.py into domain-specific utilities to reduce fan-in.

The large context window handles entire codebases without truncation, so the model sees all import statements and can reason about system-wide dependencies. At $0.75/Mtok, analyzing 8k lines (~150k tokens) costs ~$0.11. Trade-off: no code-specific benchmark scores published, so performance versus dedicated code models like Claude Sonnet 4 on complex refactoring tasks is unquantified.

When Gemini 3.8 Flash's native video input replaces frame-extraction workflows

A 9-person community platform moderates 400 user-uploaded videos daily for policy violations—hate speech in audio, prohibited imagery in frames, text overlays with spam links. Gemini 3.8 Flash ingests video files directly and analyzes audio, visual, and text layers in one model call, eliminating the frame-extraction and transcription preprocessing that Claude Opus workflows require. At $0.75/Mtok input, a 90-second video averaging 50K tokens costs $0.0375 to process; output flagging at $3.75/Mtok adds $0.01 for a 3K token violation report. Per-video cost: $0.05. The 1M token context means you can include the last 30 days of a user's upload history in the same call for pattern detection. If your videos are under 30 seconds and text-only moderation suffices, GPT-4o at $2.50/Mtok input handles it cheaper. Above 200 videos/day with mixed content types, Gemini 3.8 Flash's native multimodal pipeline is the right call.