Fable 5.1 Ships, OMP Model Roles, Seedance 2.5 — AI Daily Sep 1
617 messages · 88 active members
Overview
Topics
Anthropic released Fable 5.1 mid-day with notably faster TPS, better cache efficiency, and a less robotic tone. Most users got fresh usage resets, though some regions were skipped and guardrails auto-demoted to Opus 4.8 in edge cases. Reasoning feels similar to Fable 5, and a system prompt leak appeared within hours.
@thewildzeno walked through OMP's model role system — SMOL, DEFAULT, SLOW, VISION, PLAN, DESIGNER, TASK, ADVISOR — and clarified the split between /model config, .omp/agent/agents files, and agents.md. Recommendations: GPT 5.6 Luna xhigh for TASK workhorse, frontier + high reasoning for PLAN, and keeping harness-level instructions out of agents.md.
@RayTGcontact demoed Seedance 2.5 one-shots for a real addiction-treatment ad, with Claude handling research, captions and branding end-to-end. Builders compared Omni output quality, shared Franky Shaw's prompt template, and debated whether ComfyUI workflows change results. anonymous recommended recursive judge loops with reference examples for consistency.
@jcartu praised Hindsight as state-of-the-art memory that seamlessly bridges OMP and Hermes sessions, but flagged Reflect as consuming 70-80% of his $300/month bill. @archimorty shared tuning tips: disable retainToolCalls in settings.json, cap reranking candidates, and use custom extraction profiles per project.
anonymous is running a CC/Fable orchestrator commanding OC/Sol ops agents to upgrade Hermes infra entirely over Telegram, with Slack backup, and is exploring product-manager specialist agents on top of dev/QA/finance/security roles. anonymous shared a Grok-bot spying agent that scans Meta Ads Library, X and Reddit on a schedule and feeds a creative-gen system.
Key Takeaways
- Fable 5.1 is materially faster with better cache efficiency, but reasoning feels flat vs. Fable 5 — expect ~3 weeks before rate-limit nerfs kick in
- Configure OMP model roles in .omp/agent files, not agents.md — GPT 5.6 Luna xhigh is a strong TASK workhorse and GLM 5.3 Flash covers vision/orchestration cheaply
- Hindsight's Reflect procedure drives 70-80% of memory costs; disable retainToolCalls and cap reranking to slash bills
- Seedance 2.5 + Claude research is a viable end-to-end SMB ad production stack — one builder shipped a 30s treatment-center ad with captions and branding in a single flow
- Specialized micro-agents (dev, QA, finance, security, PM) coordinated by an orchestrator over Telegram is outperforming monolithic single-agent setups
Hot Threads
Getting started with OMP: /model vs agents.md and multi-LLM routing
Fable 5.1 rollout, usage resets and orchestrator behavior
SaaS moats via FusionIO drives, 200k IOPS, and colocation vs AWS