Google: Gemini 2.5 Flash (batch)

Review this pull request diff and identify potential performance bottlenecks or memory leaks. Focus on the database query changes and the new caching layer.

This example would produce a structured code review highlighting three areas: First, the N+1 query pattern introduced in the user lookup loop (lines 47-52), recommending a batch fetch. Second, the unbounded cache growth in the Redis implementation—no TTL set, no eviction policy. Third, a subtle issue where the connection pool isn't released in the error path (line 89). The model would provide specific line references and suggest fixes with code snippets, maintaining the technical depth expected from a 1M+ token context window that can hold the entire codebase.

Batch processing makes this model cost-effective for large-scale code review workflows—$0.15/Mtok input means analyzing a 500k-token codebase costs $0.075. The async nature means you trade immediacy for economics; suitable for CI/CD pipelines, not live pair programming.

Analyze these 40 customer support transcripts and extract common complaint themes. Group by severity and suggest product improvements for the top three issues.

This example would produce a tiered analysis: Critical issues (authentication failures, 18 mentions), high-priority concerns (slow dashboard load times, 12 mentions), and moderate friction points (confusing onboarding flow, 9 mentions). For each theme, the model would quote representative customer language, quantify impact, and propose concrete fixes—like implementing OAuth fallback, lazy-loading dashboard widgets, or adding contextual tooltips. The multimodal capability means it could process screenshot attachments customers included.

The 1M+ token context window handles batch transcript analysis that would require chunking in smaller models. Audio input support means you could feed call recordings directly. The $1.25/Mtok output cost favors concise summaries over verbose explanations—design prompts accordingly.

Generate a technical specification document for a real-time collaborative whiteboard feature. Include WebSocket architecture, conflict resolution strategy, and data persistence approach.

This example would produce a structured spec document covering: WebSocket connection management with heartbeat protocol and reconnection logic; operational transformation for conflict-free concurrent editing using a vector clock approach; and a hybrid persistence model with Redis for active sessions plus PostgreSQL for long-term storage. The model would include sequence diagrams (described textually), API endpoint definitions, and edge case handling—like what happens when two users delete the same object simultaneously. The output would span 800-1000 words with clear section headers.

Batch mode suits document generation where you're processing multiple spec requests overnight or during off-hours. The model's file input capability means you could attach existing architecture docs as context. Output quality at this price point competes with real-time models, but latency makes it unsuitable for interactive drafting sessions.

When batch mode cuts your document pipeline cost by 80%

A 12-person legal ops team runs 400 contract summaries every night—each doc averages 18k tokens, output is 600 tokens. At $0.15 input and $1.25 output per Mtok, that's $1.08 input plus $0.30 output per batch, or $1.38 total. The synchronous Flash variant costs $7.50 for the same work. Batch mode adds 10-minute latency, but since these run at 2am and land in Slack by 7am, no one cares. The 1M-token context window means you can stuff a full NDA plus three prior amendments into one call without chunking. If you need answers before coffee, pay the real-time premium. If your pipeline runs while you sleep, this is the model.

Why multimodal batch works for trust and safety at scale

A 20-person marketplace startup reviews 8,000 user-uploaded product photos daily for policy violations. Each image gets a 120-token prompt and returns a 40-token classification. At batch pricing, that's $0.15 per million input tokens (negligible for short prompts) and $1.25 per million output tokens—roughly $0.40 per 8,000 images. The model handles image, video, and audio natively, so you can route all moderation through one endpoint instead of stitching together three vendors. Batch latency averages 8 minutes, which fits a twice-daily review cadence. If you're moderating live chat or livestreams where seconds matter, you need synchronous. For overnight or hourly sweeps, this cuts your moderation bill to one-fifth of real-time Flash.

When audio context and batch pricing beat specialist transcription APIs

A 4-person content studio publishes 6 podcasts per week, each 45 minutes (roughly 300k tokens of audio context). They need transcripts plus topic tags for their CMS. Gemini 2.5 Flash batch handles audio natively and fits the entire episode in one 1M-token context window, so no chunking or re-stitching. At $0.15 input per Mtok, each episode costs $0.045 for transcription plus $0.05 for a 40k-token output (transcript and tags). That's $0.095 per episode, or $0.57 per week. Specialist transcription APIs charge $0.25-$0.50 per episode for audio alone, then you pay again for tagging. Batch latency is 12 minutes, which works fine for a Monday-morning publish cadence. If you're transcribing live events, pay for real-time. For weekly drops, this is the cheapest multimodal path.