Google: Gemini 3.8 Flash (batch)

Analyze this 45-minute earnings call transcript and extract: (1) three revenue growth drivers mentioned, (2) any forward guidance changes, (3) analyst sentiment shifts. Format as a structured summary.

This example would produce a clean three-section summary. Revenue drivers: the model identifies product line expansions in APAC (mentioned at 12:40), a new enterprise tier launch (18:22), and improved retention metrics (31:05). Forward guidance: Q4 revenue range revised upward from $280-290M to $295-305M, citing stronger pipeline conversion. Analyst sentiment: tone shifted cautiously optimistic after the CFO addressed margin pressure concerns; two analysts upgraded price targets mid-call. The model maintains transcript timestamps and distinguishes management statements from analyst questions throughout.

The 1M+ token context window handles full earnings transcripts without chunking, preserving cross-references between sections. Batch processing at $0.38/Mtok input makes this economical for weekly analyst workflows. Output structure is consistent but occasionally merges related points — you may need a follow-up prompt to separate overlapping themes.

Review this 8-page product spec PDF and this 22-minute design review recording. Flag any requirements the design doesn't address, then suggest three UI improvements that would close those gaps.

This example would produce a gap analysis followed by actionable recommendations. Unaddressed requirements: the spec mandates offline mode (section 3.2) but the design shows no local storage UI; accessibility targets (section 5.1) require keyboard navigation that isn't visible in the prototype; the API rate limit handling (section 4.3) has no user-facing error state. Suggested improvements: (1) add a sync status indicator in the header with offline queue visibility, (2) implement focus trapping in the modal flows shown at 14:30 in the recording, (3) design a toast notification system for API errors with retry actions. Each suggestion references the spec section it addresses.

Multimodal input (PDF + video) lets the model cross-reference static requirements against dynamic design walkthroughs. At $1.88/Mtok output, longer responses cost more than text-only models, but the batch API offsets this for non-urgent reviews. The model occasionally misses subtle design details visible only in video — verify UI element claims against your own review.

I'm attaching 140 customer support tickets from Q3. Cluster them by root cause, rank clusters by revenue impact, and draft a one-page executive brief explaining what engineering should prioritize next quarter.

This example would produce a data-driven brief with three sections. Cluster analysis: 38 tickets (27%) trace to the payment retry logic failing on expired cards—estimated $180K MRR at risk; 31 tickets (22%) report mobile app crashes on Android 12+ during file uploads—affecting 15% of mobile users; 24 tickets (17%) describe confusing onboarding flow for enterprise SSO—correlates with 40% drop-off in trial-to-paid conversion for teams over 50 seats. Remaining tickets scatter across minor bugs. Recommendation: prioritize payment retry fix (2-week eng effort, highest revenue protection), then SSO onboarding redesign (4-week effort, unlocks enterprise segment growth). The brief includes a simple table mapping each cluster to affected user count and revenue exposure.

Batch processing 140 tickets at once reveals patterns that per-ticket analysis misses, and the 1M token window accommodates the full dataset without sampling. The model infers revenue impact from ticket metadata (plan tiers, user counts) but can't access your actual revenue database—validate financial estimates against internal data before presenting to leadership.

When batch pricing makes overnight doc analysis viable for small teams

A 4-person legal tech startup needs to extract clauses from 200-500 contracts nightly, feeding a compliance dashboard by 6 AM. Gemini 3.8 Flash batch hits the sweet spot: the 1M token context window handles full contracts without chunking, multimodal support parses scanned PDFs and images inline, and $0.38/Mtok input pricing means a 500-doc run costs under $40. The batch queue adds 10-30 minute latency, but that's invisible when you're running overnight cron jobs. If you need sub-minute turnaround or real-time user responses, this isn't your model—switch to standard Gemini Flash. But for any workflow where you can queue work and wait, batch pricing turns a $200/night bill into $40 without sacrificing context or accuracy.

How multimodal batch cuts video review costs by 80% at scale

A 12-person creator platform moderates 3,000 user-uploaded videos per day, flagging policy violations before publish. Standard video analysis APIs run $0.15-0.25 per video; Gemini 3.8 Flash batch drops that to $0.03-0.05 by accepting delayed processing. The model ingests video, audio, and text prompts in one call, so you're not stitching together separate transcription and vision services. At 3k videos/day, that's $1,500/month instead of $7,500. The catch: batch latency means videos sit in a queue for 15-45 minutes. If your platform promises instant publish, you need real-time inference. But if you can hold uploads in review for an hour, batch pricing makes video moderation economically feasible for teams under 15 people.

When 1M token context eliminates chunking in multi-source reporting

A 6-person fintech aggregates transaction data, PDFs, and spreadsheets into investor reports once a month. Each report pulls from 40-60 files totaling 800k-950k tokens. Gemini 3.8 Flash batch's 1M context window means you load everything in one prompt—no RAG pipeline, no chunking logic, no retrieval errors. At $0.38/Mtok input, a 900k token job costs $0.34; output at $1.88/Mtok adds another $1-2 for a 15-page summary. Total per-report cost: under $3. The batch queue adds 20-40 minutes, but monthly reporting has zero time pressure. If you're generating reports hourly or need instant answers, pay for standard inference. For monthly or weekly cadences, batch mode turns a complex RAG problem into a single API call at 5% the cost.