Fable 5 Sunset, GLM 5.2 on 4 GPUs, GPT-5.6 Sol — AI Daily Jul 06

666 messages · 88 active members

666
messages
88
active members
@jcartu, @jasonakatiff, @samb69
top contributors

Overview

The day was dominated by the Fable 5 Max-plan sunset (11:59 PM PST July 8), triggering a credit-burn frenzy with builders rotating multiple $200 accounts, running overnight loops, and enforcing checkpoint discipline to survive mid-session limit hits. Consensus: use Fable 5 for architecture, design, and orchestration rather than raw code, and expect it back only if competitive pressure forces a subscription return. All eyes now turn to GPT-5.6 Sol, rumored for Thursday with Terminal Bench scores matching Fable 5 at roughly half the cost. On the local frontier, @jcartu reported running unpruned 768B GLM 5.2 FP8 on four RTX 6000 Pro cards at 60-80 tps with a 400k context window — a ~50k EUR rig with <1% degradation from full precision, approaching Opus 4.7 quality on max-think. He also shipped a vLLM overthinking-penalty plugin, while a Z.ai developer poll confirmed GLM 5.3 will finally include vision. Context engineering ran parallel: @thewildzeno and @samtome dug into why response-hook context compaction breaks Claude Code's remote prompt caching, often costing more than it saves without pre-API interception. Voice AI infrastructure heated up with xAI's Grok Voice launching at $0.05/min as a unified STT/TTS/LLM model, undercutting Telnyx-stitched pipelines — though builders shared open-source Convoso/Ringba clones on Telnyx SIP as the current default. @arielletolome demoed a Seedance 2 + Wan 2.2 + Remotion ad pipeline producing finished creatives in ~43 minutes, with Seedance 2.5 (3-minute output) incoming. Other threads: Codex is quietly issuing repeated free usage resets, ViralBench surfaced as a live-intel TikTok benchmark, and Resend/SES/Smartlead emerged as the email stack of choice.

Topics

Fable 5 leaves Max plans at 11:59 PM PST July 8, triggering multi-account rotation and overnight loops. Builders are prioritizing design/architecture work over raw code, keeping Ultracode on by default. GPT-5.6 Sol is rumored to drop Thursday matching Fable's Terminal Bench at ~50% the cost, potentially timed to Fable's cutoff.

@jcartu ran unpruned 768B GLM 5.2 FP8 on four RTX 6000 Pro cards at 60-80 tps with 400k context (90 tps with dcp2 at 262k), <1% degradation from full FP8, ~50k EUR total. On max-think it approaches Opus 4.7 quality. GLM 5.3 will include vision per Z.ai's developer poll, and @jcartu also shipped a vLLM overthinking-penalty plugin plus an abliterated Kimi K2.6 GGUF surfaced with no refusals.

xAI's Grok Voice launched at $0.05/min with a free phone number, running STT/TTS/LLM as a single unified model versus Telnyx-style stacked pipelines. Meanwhile builders shared open-sourced Convoso/Ringba dialer clones on Telnyx SIP trunks, and @jaypozo flagged the ElevenLabs grant program for free prototyping credits.

@thewildzeno and @samtome traded notes on why response-hook context compaction breaks Claude Code's remote prompt caching — real fix requires pre-API interception. Codex users reported repeated free usage resets (30-day expiry) as OpenAI apparently smooths reset-window contention. A shared repo of 31 natural-language prompts for cmux agent fleets is gaining traction for Telegram/Slack-linked multi-agent loops.

@arielletolome shipped a Seedance 2 + Wan 2.2 + Remotion ad pipeline producing finished creatives in ~43 minutes; Seedance 2.5 lands soon with 3-minute output. Email stack consensus: Resend for transactional, SES+Sendy for owned lists, Smartlead/Instantly for cold. @bartekadamczyk released ViralBench, three models iterating TikTok posts on live stats, sparking authority-funnel debate.

Key Takeaways

  • Fable 5 leaves Max plans 11:59 PM PST July 8 — rotate accounts, checkpoint to files, and use it for design over raw code.
  • GPT-5.6 Sol reportedly matches Fable 5 on Terminal Bench at ~50% cost with a likely Thursday drop.
  • Full 768B GLM 5.2 FP8 runs on 4 RTX 6000 Pros at 60-80 tps with 400k context — ~50k EUR local frontier is real, GLM 5.3 adds vision.
  • Grok Voice at $0.05/min unifies STT/TTS/LLM in one model; Telnyx SIP + ElevenLabs remains the open dialer default.
  • Response-hook context compaction breaks Claude Code's remote prompt caching — savings often net negative without pre-API interception.

Hot Threads

@jcartustarted

Full 768B GLM 5.2 running on 4 GPUs at 60-80 tps

20 replies6 participants
@jrizzolostarted

GPT-5.6 Sol release timing and Fable 5 endgame

14 replies8 participants
@scalingfrogstarted

Resend vs. alternatives for marketing email infrastructure

12 replies7 participants

Linked Items