Video Gen Costs, Claude Code Cuts, Ad Automation — AI Daily Aug 30

401 messages · 62 active members

401
messages
62
active members
@jasonakatiff, @samb69, @fewga893
top contributors

Overview

Saturday's chat was dominated by video generation economics, with @arielletolome and @nowwatchthisdrive breaking down real per-second costs across Seedance 2.5 ($9 for 30s), Higgsfield ($300 for a 6min VSL), and Veo 3.1 via Google Ultra + useapi. Consensus: 480p often converts as well as 1080p, and audio quality plus copy matter more than fidelity. MiniMax H3 Max on fal.ai was flagged as the next model to test. Meanwhile, Anthropic announced Claude Code weekly limits are being cut ~17% in September—replacing the current 50% boost with a 25% bump—sparking debate about whether it's a real downgrade forced by capacity pressure from OpenAI and Chinese labs. Ad operations automation was the other big through-line. @scalingfrog shared how Grokbot wired directly into Meta and tracker platforms pulls data every 30 minutes for hands-off management, while @RayTGcontact runs a parallel Claude setup that suggests bid changes and pushes Telegram alerts. @KtargetMedia probed Hermes agent orchestration, with @GuruTime running lean on Opus + mem0, @ltv_1 pointing to Hermes' new bots feature, and @jsmoof sharing a SpaceXAI method deploying 10-20 Grokbot agents to handle 90% of recurring tasks. On hardware, @jcartu shared a new MXFP4 draft model hitting 450 tps on two RTX 6000 Pro Blackwell cards (~$32k) for Opus 4.8-tier output, while @jasonakatiff pushed harness-first, model-second thinking and spent the evening prototyping one-click vertical templates for lead router. Tooling picks included Baserow as the Airtable replacement, Screen Studio for demo walkthroughs, and a LiveKit question from @tb_1991_11 about listen-only call co-pilots.

Topics

@arielletolome and @nowwatchthisdrive compared per-second costs across Seedance 2.5, Higgsfield, and Veo 3.1 via Google Ultra + useapi. Key insight: 480p converts as well or better than 1080p for ads, and audio plus copy carry conversions more than video fidelity. MiniMax H3 Max on fal.ai was flagged as the next model to test.

Anthropic is replacing the current 50% limit boost (active since May) with a 25% bump starting September, netting a ~17% reduction versus what users have gotten used to. @archimorty framed it as a downgrade disguised as an increase, while others speculated OpenAI and Chinese labs are forcing Anthropic's hand on capacity.

@scalingfrog and @RayTGcontact shared Grokbot and Claude setups wired directly into Meta ad accounts and trackers, pulling data every 30 minutes and pushing bid/creative suggestions via Telegram. @jsmoof detailed a SpaceXAI method running 10-20 Grokbot agents to automate 90% of recurring tasks, while @ltv_1 highlighted Hermes' new bots feature. Subscription usage caps were the main friction point.

@jcartu shared a new MXFP4 draft model from the RTX6k Discord that hits 450 tps on two RTX 6000 Pro Blackwell cards (~$32k) for Opus 4.8-tier output. @jonmacofficial recommended 3090s for audio/image/video gen and 5090 or 6000 for local LLMs. @weslindquist is spec'ing a Mac Studio for confidential accounting work, and @jasonakatiff urged harness-first thinking—invest in the framework, swap engines freely.

@jasonakatiff prototyped a template system that preconfigures entire verticals—vendors, SMS/email nurture, AI agents, contracts—into a save/apply flow so operators can launch home services or Medicare businesses in one click. The group connected this to adaptive UI concepts and agreed templates reduce architecture mistakes for less-experienced builders.

Key Takeaways

  • 480p video ads often convert as well as 1080p—audio and copy matter more than resolution, and downscaling slashes token costs.
  • Claude Code's September limits are ~17% lower than what users currently have—plan capacity before month-end.
  • Wiring Grokbot or Claude directly to ad platforms on a 30-min pull cadence makes campaign management mostly hands-off; usage caps are the main constraint.
  • Two RTX 6000 Pro Blackwell cards can now run Opus 4.8-equivalent quality at 450 tps for ~$32k using the new MXFP4 draft model.
  • Templating full vertical setups (vendors, nurture, agents) reduces architecture mistakes and gives operators a working baseline they can't easily break.

Hot Threads

@arielletolomestarted

Video model economics: Seedance vs Higgsfield vs Veo 3.1

35 replies6 participants
@KtargetMediastarted

Tools for managing Hermes agent fleets

15 replies4 participants
@jasonakatiffstarted

One-click vertical templates for lead router

14 replies3 participants

Linked Items