Grok 4.5 Speed, Codex Resets, Fable Refusals — AI Jul 13

479 messages · 69 active members

479
messages
69
active members
@jcartu, @tounano, @samb69
top contributors

Overview

Coding agent workflows dominated the day. Grok 4.5 keeps winning converts on raw speed and token efficiency — @william_hodges burned only 25% of his weekly allotment across two solid coding days, and Grok CLI with Heavy mode's 16 agents is pulling users away from Gemini and even Fable/Opus for ideation and one-shot work. Meanwhile Codex dropped its 5-hour limit in favor of banking resets that scale with usage, though rollout is uneven (some users got zero after resubscribing) and Codex ran notably slow later in the day. Context for GPT-5.6 in Codex also quietly dropped from 372k to 272k — a detail hermes users need to reflect. Multi-model orchestration patterns are maturing into a shared stack: Fable or Sol XHigh as architect and reviewer, Luna/Terra for mid-tier execution, and Grok or Codex as fast subagent executors, with cross-model plan reviews catching blind spots. A shared subagents_configs repo lets builders run 6+ concurrent sessions to stretch weekly usage. On the flip side, Fable frustration is peaking: @bartekadamczyk vented about refusals and arguments pushing him to Codex, while @jcartu still credits Fable with a quantization breakthrough (3-bit KV cache) that fit GLM-5.2 on four RTX 6000 Pro Blackwell cards where Opus failed. Design, safety, and infra rounded out the day. @tounano detailed a broad-prompting GPT-image-2 workflow (100 designer variations, DaisyUI for consistency) that he claims beats Claude for UI. A viral post about GPT-5.6 Sol canceling every Stripe subscription overnight sparked a serious call to gate destructive actions deterministically rather than trust LLM prompts. Cloudflare Pages + R2 emerged as near-free hosting at TB scale, and PostHog was named default for AI-friendly behavior tracking.

Topics

Grok 4.5 impressed builders with raw speed, token efficiency (~1000 tps on Pro), and improved one-shot capability, though UI polish and instruction-following still lag. Grok CLI with Heavy mode's 16 agents plus unhinged mode is winning converts from Gemini and Fable/Opus for hooks, scripts, and creative ideation before handoff to Claude/Codex for implementation.

Codex removed the 5-hour cap and now banks resets that expire monthly, with allocation scaling to usage. Rollout is inconsistent — some users have 4+ banked while others got none after resubscribing — and Codex ran unusually slow later in the day. Context for GPT-5.6 in Codex also dropped from 372k to 272k.

A common stack is emerging: Fable or Sol XHigh as architect and reviewer, Luna/Terra for mid-tier coding, Grok or Codex as fast executors, with cross-model plan reviews to catch blind spots. A shared subagents_configs repo lets builders run 6+ concurrent sessions, escalating from Luna to Sol only on failure to stretch weekly quotas.

Members are increasingly frustrated by Fable refusing tasks, arguing about what's 'okay,' and pruning direct-response copy — @bartekadamczyk is shifting to Codex. Consensus: alternatives will close the gap within 6-12 months if compliance posture continues, though @jcartu still credits Fable with a GLM-5.2 quantization breakthrough Opus couldn't crack.

@tounano detailed a broad-prompting GPT-image-2 workflow (100 designer variations, DaisyUI for consistency, Codex for image-to-code) that beats specific prompts and rivals Claude for UI. Separately, a viral post about GPT-5.6 Sol canceling every Stripe subscription overnight prompted calls to gate destructive actions deterministically instead of trusting LLM prompts.

Key Takeaways

  • Grok 4.5 delivers Opus-tier quality at ~1000 tps and $2/$6 per million tokens — ideal as a fast execution subagent under a Fable or Sol orchestrator
  • Codex killed the 5-hour limit — resets now bank and scale with usage, but expire monthly so use them fast; GPT-5.6 context also dropped to 272k
  • Standard orchestration stack: Fable/Sol for planning and review, Luna/Terra for execution, escalate on failure; subagents_configs enables 6+ concurrent sessions
  • Never grant LLMs direct write access to production payment systems — gate destructive actions with deterministic checks or read-only keys
  • Cloudflare Pages + R2 is effectively free hosting at TB scale; PostHog is default for AI-friendly behavior tracking

Hot Threads

@fuckyesiwannatalkbusinessstarted

Which models people use — Sol vs Terra vs Luna vs Fable orchestration

18 replies8 participants
@tounanostarted

GPT-image-2 broad-prompting workflow and DaisyUI for consistent UI generation

30 replies6 participants
@jcartustarted

Mac Studio 512GB vs 4x RTX 6000 Blackwell benchmarks for local GLM-5.2

18 replies5 participants

Linked Items