Opus 5 vs Fable, OMP Harness, K3s Stack — AI Daily Jul 25
341 messages · 63 active members
Overview
Topics
Builders spent the day pitting Opus 5 against Fable, Codex 5.6, and Grok 4.5. Opus 5 is faster and cheaper and solved front-end problems Codex couldn't, but makes stuff up more and is weaker at orchestration. Fable has drawn sharp post-relaunch criticism for refusals on security/VPN/tuning tasks — many say it's now only usable for front-end, pushing power users to Opus 5 or local models like Sol and K3. Both hit 'overloaded' states in the afternoon.
Kieran noticed OMP outputs have been consistently on-target with minimal prompting, sparking a discussion on what harnesses inject beyond model behavior. Referenced write-up concluded Codex CLI is the weakest harness, OMP/OpenCode win for GPT, and Claude Code remains best for Fable/Opus — with orchestrator patterns (Opus plan → Grok work → Fable review) becoming the new default via projects like fable-advisor.
@leewardbound broke down a lab-grade setup: one Hetzner box (256GB RAM, 10TB HDD, 4TB SSD) running k3s with ~80 Django+Vite apps, 60 databases, Crunchy PG operator, Redis, Harbor registry, and self-hosted GitHub runners. Grafana+otel dashboards are generated by Claude and CI-deployed as YAML, with high-volume apps graduating to dedicated AKS clusters.
@Coybh shared a Claude Code artifact that started as pixel art and evolved via Higgsfield MCP and Mobbin MCP integrations, sourcing design references and iterating creatively. The progression drew requests for workflow details and praise for the animations, reinforcing MCP stacking as a real design pattern.
Kieran argued unique datasets (Mobbin-style UI refs, meme corpora, industry silos) are more defensible than another 'AI SDR' wrapper — potentially by acquiring dated companies sitting on valuable data and flipping them into MCPs. K3 open weights reportedly drop July 27, Flux 3 previews look promising, and split-tests show Veo 3.1 continues to beat Seedance 2 for ad creative — though 'meh' AI ads still outperform polished ones for older demos.
Key Takeaways
- Opus 5 is cheaper and faster than Fable but weaker at orchestration — Fable is now only reliable for front-end post-relaunch, so route serious work through Opus 5 or local models
- Harness choice matters as much as model: OMP for GPT, Claude Code for Fable/Opus, and avoid Codex CLI as a harness even with GPT models
- A $120/mo Hetzner + k3s can host 80 fullstack apps if you let Claude generate Grafana dashboards as YAML and deploy via CI — dedicated clusters only for high-volume graduates
- MCP stacking (Higgsfield + Mobbin inside Claude Code) is a real design workflow — and proprietary datasets are a more defensible agent-era moat than another agent wrapper
- K3 open weights drop July 27, Flux 3 shows promise, and Veo 3.1 beats Seedance 2 for ads — but polished AI ads underperform 'meh' ones in older demos
Hot Threads
Self-hosted k3s stack running 80 apps on a $120/mo Hetzner
Why OMP outputs feel more reliable — harness engineering breakdown
Opus 5 falling short on orchestration vs Fable