GLM 5.2, LongCat Avatars, Anthropic Crackdown — AI Daily Jun 25
580 messages · 79 active members
Overview
Topics
Builders confirmed Claude 4.6 still beats 4.7/4.8 for creative/copy work, with consensus that Anthropic could 3-5x pricing and still be worth it. @jcartu pushed GLM 5.2 hard (81% on the 19-hour terminal-bench) and is piping GLM-4.6V output into it for a working vision stack. Many builders have also moved permanently from Fable back to Codex or Claude 5.5 for product work.
@arielletolome demoed LongCat-Video-Avatar-1.5 lipsync on an RTX 6000 Blackwell with ElevenLabs voice, sparking a broader move toward local rigs. RTX 6000 prices reportedly doubled to $12-14k as builders react to Anthropic instability; a viral post about Uber's $500M Anthropic bill fueled the open-source argument.
Speculation about US restrictions on Anthropic dominated mid-day, with builders blaming Anthropic's own 'Mythos' safety narrative for inviting regulation. The latest accusations against a Chinese lab triggered debate, while @MiamiMega surfaced aiapipk.com — a live marketplace of arbitraged Claude/OpenAI keys at ~5% retail where user data farming is the real margin. @robinroy gave an insider-flavored take on China's industrial dominance and Shenzhen.
@Kieran and @samb69 ran parallel threads on shipping autonomous copywriting. Consensus from @mb29266: one strong main writer + stacked editor/QA agents beats splitting hook/body/design into sub-agents, which lose context. For client delivery, wrap headless Claude Code, tmux sessions, or Hermes pipelines in a polished chat UI/dashboard — non-technical buyers reject raw terminals.
@mb29266 lost half a day of revenue when Funnelish pushed a JS update that broke his Stripe gateway; @samb69 recommends migrating to headless Checkout Champ via API. Shopify's new dispute fee structure (pay twice if you lose) is pushing more builders off-platform. On video gen, Seedance via Byteplus direct API leads on price/speed; Higgsfield's 'unlimited' plans flagged for hidden terms.
Key Takeaways
- Claude 4.6 still beats 4.7/4.8 for copy/creative work — downgrade to 4.6 1M if your prompts were built on it; Claude 5.5 outperforms Opus on product work.
- GLM 5.2 hit 81% on terminal-bench — combine with GLM-4.6V for vision and you have a serious Sonnet alternative for coding pipelines.
- Anthropic regulation fears are real enough that builders are pre-buying GPUs; RTX 6000 prices nearly doubled to ~$12-14k, and LongCat-Video-Avatar-1.5 now runs lipsync on a single one.
- For client-facing AI copy tools, use one strong writer + layered editor/QA agents (not split sub-agents), and wrap CLI pipelines in a polished chat UI — clients reject raw terminals.
- Migrate off Funnelish to headless Checkout Champ via API before the next breaking JS push; for video gen, Seedance via Byteplus direct API is the current price/speed leader.
Hot Threads
Claude pricing 3-5x scenario, 4.6 vs 4.8 regression, and GLM 5.2 as the alternative
LongCat avatar demo on RTX 6000 Blackwell and the local inference pivot
Architecting autonomous ad copywriting stacks for high-ticket clients