Real costs of running a production AI agent 24/7
A solo founder running a production AI agent for 30 days reports total monthly costs of $92–132, dominated by Claude API usage ($55–90), with practical insights on prompt caching, local video encoding, and ROI for content and SaaS applications.
Score breakdown
Developers building production agents can use this real-world cost breakdown and the critical cache TTL discovery to optimize API spending, avoid silent cost increases, and make informed decisions about model selection and local vs. cloud infrastructure.
- 01 <item>Claude Sonnet 4.6 is the primary agent model; Claude Opus 4.7 is used sparingly for strategic decisions only.</item> <item>Typical overnight sessions consume 50,000–120,000 input tokens and 5,000–15,000 output tokens, with estimated monthly Anthropic costs of $55–90.</item> <item>Anthropic silently reduced prompt cache TTL from 1 hour to 5 minutes in March 2026; explicit cache control headers are required to maintain cost efficiency.</item> <item>Local video processing via ffmpeg on M-series Mac eliminates cloud encoding costs; Remotion renders take 5–50 minutes depending on composition complexity.</item> <item>Total estimated monthly cost is $92–132, including API fees ($55–90), text-to-speech ($5–8), image generation ($2–4), and platform hosting ($30).</item> <item>The author spends ~30 minutes/day on oversight and reports ROI of 20–50x for SaaS content marketing if conversions materialize.</item>
Atlas Whoff documents the operational costs of running a production AI agent continuously for 30 days. The agent operates autonomously overnight, managing a content pipeline, YouTube channel, email monitoring, and generating morning reports. The primary model is Claude Sonnet 4.6 via the Anthropic API, supplemented by Claude Opus 4.7 for strategic decisions, Mistral Voxtral for text-to-speech, and MiniMax M2 for visual generation. Video composition uses Remotion (React-based, local rendering) and ffmpeg for encoding, both running on an M-series Mac.
Typical overnight sessions consume 50,000–120,000 input tokens and 5,000–15,000 output tokens.
API costs dominate the budget. Typical overnight sessions consume 50,000–120,000 input tokens and 5,000–15,000 output tokens. With prompt caching enabled, cached input costs $0.30/MTok while non-cached input costs $3.00/MTok and output costs $15.00/MTok. Estimated monthly Anthropic costs are $45–90 for Sonnet and $8–15 for Opus. A critical discovery: Anthropic silently changed the default prompt cache TTL from 1 hour to 5 minutes in March 2026. Without explicit cache control headers, this can increase costs 3–5x. The fix is adding `cache_control: {"type": "ephemeral"}` blocks to system prompts. Secondary costs include Mistral Voxtral at $0.08–$0.12 per 60-minute story ($5–8/month), MiniMax at $2–4/month, and platform hosting at $30/month. Local video processing eliminates cloud encoding costs entirely. Total estimated monthly cost: $92–132.
The author emphasizes that local hardware (M-series Mac) is sufficient for the entire pipeline, requiring only 30 minutes of daily oversight. ROI depends on output: the YouTube channel targets monetization in ~60 days, while the SaaS application replaces 15–20 hours/week of human content marketing work, yielding a 20–50x ROI on API costs if conversions follow.
Key facts
- 01 <item>Claude Sonnet 4.6 is the primary agent model; Claude Opus 4.7 is used sparingly for strategic decisions only.</item> <item>Typical overnight sessions consume 50,000–120,000 input tokens and 5,000–15,000 output tokens, with estimated monthly Anthropic costs of $55–90.</item> <item>Anthropic silently reduced prompt cache TTL from 1 hour to 5 minutes in March 2026; explicit cache control headers are required to maintain cost efficiency.</item> <item>Local video processing via ffmpeg on M-series Mac eliminates cloud encoding costs; Remotion renders take 5–50 minutes depending on composition complexity.</item> <item>Total estimated monthly cost is $92–132, including API fees ($55–90), text-to-speech ($5–8), image generation ($2–4), and platform hosting ($30).</item> <item>The author spends ~30 minutes/day on oversight and reports ROI of 20–50x for SaaS content marketing if conversions materialize.</item>