Provider Outage, Fable 5.1, Astra Teaser — AI Daily Sep 3
581 messages · 87 active members
Overview
Topics
Claude, Grok, Codex and briefly ChatGPT all went down in overlapping windows, forcing builders onto GLM, Kimi and Gemini fallbacks. Speculation ranged from cyberattack threats to shared infra dependencies, though most settled on incompetence over malice. Sparked renewed interest in local model setups including 512GB Mac Studios and GLM 5.3 Flash NVFP4 on dual Blackwells.
Fable 5.1 launched at $10/$50 per million tokens (2.5x Sol pricing) claiming 70% better token efficiency and detailed thinking. OpenAI teased Astra (rumored GPT-6) mid-outage with heavy computer-use emphasis and reportedly no token cap on the weekly plan. Grok 4.7 expected in 8-9 days, driving $60/month Heavy plans via reseller bots. Meta's Muse Code also exited beta.
@sibunting shared a managed compact system for Claude Code: external orchestrator (Kit) monitors context, triggers compacts at work seams, agents write session notes/journal/handover, and tmux byte-level drills verify the compact command. @miyamotaz demoed self-hosted Hindsight via Docker/OrbStack with scoped per-repo memory and GLM 5.3 Flash coding plans, dropping infra costs to near zero. Debate on /shake in OMP surfaced cache TTL tradeoffs (5min-1hr) that could 10-100x costs if automated per turn.
@arielletolome is building a media buying agent for Meta/Snapchat/X reporting scaling recommendations at the ad level, though client app-permission constraints forced a pivot. @rstmaur detailed COMPSCRAPER: n8n pulling 250+ ads/day across five niches, combining Foreplay creative signals with Apify scale data, scoring on longevity/variants/reach, then Seedance 2.5 + Whisper + FFmpeg for production. Hourly campaign management shows diminishing returns vs solid rules + weekend oversight.
@jarvisballer reports 90% of GitHub issues completed autonomously via Orca + OMP dark factory flows. @jasonakatiff switched Vercel to Hetzner for preview builds, calling it night-and-day faster and cheaper (50+ builds/day without cost anxiety). @jasperckck built a one-prompt CC workflow scraping Apollo (capped 5/company), filling gaps from LinkedIn, enriching via MillionVerifier and bounceban, then posting to Instantly and Lemlist.
Key Takeaways
- Diversify LLM providers — Claude, Grok and Codex all went down together while GLM, Kimi and Gemini kept working
- Managed compact + three artefacts (session notes, journal, handover) lets Claude Code sessions run for weeks on one transcript
- Self-hosted Hindsight on GLM 5.3 Flash coding plans drops memory infra from thousands/month to effectively zero
- Fable 5.1 costs 2.5x per token but claims 70% better efficiency; token cache TTLs (5min-1hr) mean per-turn /shake can 10-100x costs
- Hetzner beats Vercel materially for high-volume preview builds; Orca + OMP dark factory can close 90% of GitHub issues autonomously
Hot Threads
Confirmed outages across Grok, Claude and Codex — fallback options and cache TTL implications
Fable 5.1 launch, pricing, and benchmark reactions
Managed compact system for Claude Code with tmux byte-level drills and Hindsight memory