Fable Refusals, OMP Harness, Video Routing — AI Daily Jul 14

Fable 5's broad safeguards are flagging routine coding and sysadmin tasks, burning tokens before refusing or downgrading to Opus 4.8. @jcartu reported Sol completed a media-player install that Fable rejected, and a shared kilo.ai analysis showed Sol produces better execution plans regardless of which model runs them.

@jcartu and @samb69 championed omp.sh as a single harness that wraps all OAuth/API accounts and local GPUs, auto-switching models by role. @thewildzeno confirmed the speed but noted terminal rendering glitches outside Arch Linux. Users are pairing it with GPT-5.6 + Grok 4.5 for fast default execution.

@iamgalba detailed an agent-based routing setup across Seedance, Veo, WAN 2.7, and Grok Imagine 1.5 that produced 70-segment midform videos for ~$30. @samb69 reported similar wins, dropping Seedance video costs from $8-12 to ~$2 per video, with @arielletolome praising WAN 2.7 for B-roll generation.

@jcartu shared a MXFP8/NVFP4/NF3 hybrid quantization of GLM 5.2 that stays within 1% KLD of BF16 — better than z.ai's FP8 cloud serving. He clarified GLM is mainly useful for compliance-sensitive tasks, as Fable and GPT-5.6 outperform it when they comply.

Grok 4.5's ~1000 tps speed is convincing members to restructure subs — cancelling Anthropic seats, downgrading GPT to $100, and upgrading Grok to the $300 tier (currently $99 for 3 months). @samb69 uses Sol as advisor and Grok as the fast executor for hours-long chained workflows.

Fable refusing sysadmin tasks while Sol complies

Agent-routed video pipeline cutting UGC costs in half

Running GLM 5.2 locally with near-lossless hybrid quantization