Medium conversations, real-time, web search + RAG common. ~1,000 conversations/day is a typical mid-market starting point.

Short conversations dominated by retrieval context — input-token-heavy. Caching pays off when the same retrieved chunks repeat across users.

Long conversations with high output token counts. Quality-tier sensitive — Sonnet/GPT-5/Opus typically beat budget tiers on code reasoning.

Single-turn, scheduled — the canonical Batch API use case. 50% off list pricing on output tokens, no real-time pressure.

High-volume single-turn. Mid- or budget-tier models often win — quality difference is small for narrow tasks. Caching the system prompt is the biggest lever.

Long conversations to plan + script + per-scene prompt + retry. Generation cost dominates — Veo or Sora at 720p, ~4 clips × 8s per finished video, ~10 videos/day. Brain LLM is the minority of the bill.

Single-turn brain calls (or short), one-image-per-call generation at 1024². ~100 images/day across DALL-E 3 or Flux. Caching the brain prompt + sharing infrastructure across calls is where the savings hide.

Brain LLM drafts a script (~5,000 chars), TTS synthesizes it. ~50 scripts/day. ElevenLabs for expressive reads, OpenAI / Google for cost-sensitive volume.

Brain LLM iterates on prompt + style cues; generation produces full songs. High reject rate — most tracks regenerated 2-3× before keep. ~20 finished tracks/day.