Claude Fable 5, Local LLM Stacks, Meta Ads Scaling — AI Daily Jun 10
522 messages · 74 active members
Overview
Topics
Anthropic's new Mythos-class Fable 5 hit 91/100 benchmarks with 1M context and ~2x Opus pricing, included in Max plans only until June 22. Builders praised its autonomous initiative, 5-second compactions, and orchestration of Opus/Sonnet sub-agents — though it's tight-lipped, token-hungry, and often skips planning to just ship code.
Heavy debate on whether local compute (RTX 6000 Pro, NVFP4 lossless quants, 27b models) plus Chinese frontier models (GLM 5.1, Minimax M3, DS4 Pro, Qwen 3.7) can replace Anthropic for ~95% of tasks. Consensus: scaffold with Fable/Opus, assemble locally to slash costs — but audit network calls after Qwen Code was caught sending usage stats without permission.
An operator running stable budgets saw 20-30% lower CPA and 30-50% better conversion than surf scaling. Veterans agreed lowering budgets pushes ads into worse auctions post-August algo shifts — focus on creative, offer, and landers over budget manipulation, using 7-day windows. Andrei Lunev's agentic Meta Ads playbook with Threads arbitrage was widely shared.
Seedance 2.0 is the dominant video model with no American competitor — bypass its real-person flagging on AI characters with 4x4 or 6x6 grid overlays. A polished Veo3 claymation built via hand-prompted script → storyboard → first-frame/last-frame stitching impressed the group; ElevenLabs expression tags ([angry], [shyly]) fix robotic VO. HeyGen Avatar 5 praised for character consistency.
Users debated whether $200 Max plans now feel like $20 Pro plans given Fable's appetite, and whether to grab 20x subs for the 12-day window. Pairing Fable for planning with Codex for execution emerged as the cost-saving pattern. Beginners got a resource drop: Perplexity's agent-skills article, the Hermes repo, GitHub spec-kit, Matt Pocock's LLM course, and channels from David Ondrej, Peter Yang, and Nate Herk.
Key Takeaways
- Fable 5 is included in Max plans only until June 22 — ship complex builds now and pair it with Codex for execution to stretch weekly token limits before API-only pricing kicks in.
- Optimal cost pattern: scaffold end-to-end with Fable (rubrics + quality gates), then assemble with cheaper models (DS4f, 27b, Minimax M3) locally — NVFP4 quants are now lossless and 27b runs on a single RTX 6000 Pro.
- Stop fighting Meta's budget — stable budgets are outperforming surf scaling by 20-50% on conversion; constant lowering appears to push ads into low-quality auctions post-August shifts.
- Seedance 2.0 has no American competitor for video — use 4x4 or 6x6 grid overlays to bypass real-person detection; for Veo3 long-form, hand-prompted storyboards + first-frame/last-frame stitching still beats automation.
- When using Chinese open models, audit network calls and cross-check generated code with a second model — Qwen Code was caught sending usage stats without permission, and prompt injection risks remain real.
Hot Threads
Scaffold with Fable, assemble locally with DS4f/27b to save a fortune
Surf scaling vs stable budgets — is Meta's algo punishing budget changes?
$200 Max plan now feels like a $20 Pro plan — Fable burns tokens