OMP Harness, Kimi K3 Sold Out, Fable 5 Pricing — AI Daily Jul 19

529 messages · 69 active members

529
messages
69
active members
@jasonakatiff, @c_1media, @jarvisballer
top contributors

Overview

The day was dominated by a rapid harness and model reshuffle around OMP (omp.sh), which emerged as the coding harness of the moment. Builders migrated from Claude Code and stacked OMP with CMUX, Codex, Kimi K3, Grok 4.5, GLM, and Sol Max — with 'Grokini' (Grok 4.5 orchestrator + Gemini/Codex subagents) becoming a favored daily driver. Praised features included multi-model OAuth, per-token cost visibility on subscriptions, and time-traveling stream rules that abort mid-token to inject corrections that survive compaction. Known bugs: stale AGENTS.md/CLAUDE.md pickups and agent-selection quirks worth tracking on GitHub. Kimi K3 stole the spotlight after @jarvisballer's blind eval showed it beating GPT-5.6, Fable, and Opus on code review, ad copy, research memos, and compliance-edge content — even catching an RCE vulnerability every Claude model missed. The $100 plan is punching above its weight for UI mockups (4 interactive designs in 5 minutes), and upgrading to $200 resets the weekly quota. Within hours Moonshot paused new subscriptions citing compute limits, leaving many locked out. Meanwhile Fable 5 was made permanent on Max/Team plans, but refusals on legitimate builds and enterprise bills hitting $7.5k/employee/month are pushing workloads to GLM 5.2 and K3 at 70–90% discounts. Other threads covered scaling AI SEO content and thousands of ChatGPT images per day via isolated CMUX workspaces per OAuth account, self-hosting tradeoffs (one SaaS cut $4k/mo migrating to Hetzner), and GPU/ffmpeg workflows on Runpod, Vast, and AWS Lambda. Growing skepticism toward MCPs emerged — modern models can just read API docs directly.

Topics

OMP.sh became the harness of the week, with builders pairing it with CMUX, Codex, and multiple OAuth models for agnostic orchestration. The 'Grokini' setup (Grok 4.5 driver + Codex Sol planning + Claude via skills for review) is emerging as a fast daily driver. Standout features: time-traveling stream rules that inject corrections mid-token and survive compaction, plus per-token cost visibility even on subscriptions. Watch GitHub issues for stale AGENTS.md/CLAUDE.md pickups and agent-selection bugs.

@jarvisballer's blind evaluation showed K3 beating GPT-5.6, Fable, and Opus on code review, ad copy, research, and compliance work — including catching an RCE vuln every Claude model missed. The $100 plan handles heavy usage and produces strong UI mockups with trivial prompts; upgrading to $200 resets the weekly quota. Moonshot paused new subscriptions mid-day citing compute limits, with weights reportedly dropping in ~5 days.

Fable 5 was made permanent on Max/Team plans, but refusals on legitimate builds are driving cancellations and fallback strategies (Fable as orchestrator with K3/cheaper subagents doing the work). Sentiment is that Anthropic's pricing won't hold against GLM 5.2 and Kimi K3 at 70–90% discounts, with enterprise AI bills hitting $7.5k/employee/month accelerating migration. Users report Fable feeling 'lazy' while Codex will grind for hours.

@robinroy claimed #1 Google rankings in Scandinavia with 100% AI-written content across 8-9 OAuth ChatGPT accounts, needing thousands of images per day. @c_1media outlined the architecture: one isolated CMUX workspace per account, a session manager for OAuth, a central orchestrator, and optionally a HAR-based lightweight CLI client. @arielletolome flagged Ideogram 4 as on par with Nano Banana Pro and GPT Image 2.

OpenShip sparked debate — great for personal projects but not serious SLA workloads. Real anecdotes: one SaaS cut $4k/mo migrating from Render to Hetzner, an SRE moved a blockchain firm from $150k/mo RDS to $3k baremetal. For parallel ffmpeg on 15–20 videos: Runpod serverless, Vast, or AWS Lambda (1000 concurrency) — consensus is ffmpeg rarely needs GPU. Coolify mentioned alongside OpenShip.

Key Takeaways

  • OMP + Grok 4.5 orchestrator is the harness stack of the week — its time-traveling stream rules abort mid-token and inject corrections that survive compaction.
  • Kimi K3 credibly beats Claude and GPT-5.6 on code review, ad copy, and security audits; the $200 upgrade resets your weekly quota — useful arbitrage while subscriptions reopen.
  • Fable refusals plus $7.5k/employee/month enterprise bills are pushing real workloads to GLM 5.2 and K3; use Fable as orchestrator with cheaper subagents.
  • For scaling OAuth ChatGPT image generation, isolate one CMUX workspace per account with a session manager and central orchestrator — or reverse a HAR file into a lightweight CLI.
  • Self-hosting on Hetzner/baremetal can cut infra costs 10–50x when you don't need strict SLAs; skip MCPs when models can just read API docs directly.

Hot Threads

@geiltstarted

OMP + Grokini stack running blazing fast as daily driver

50 replies14 participants
@jarvisballerstarted

Kimi K3 blind eval results vs GPT-5.6, Fable, and Opus

25 replies10 participants
@robinroystarted

Scaling AI SEO content and thousands of ChatGPT images per day

30 replies8 participants

Linked Items