Haiku 4.5 Orchestration, Multi-Agent Swarms, Claude Credits — AI Daily Oct 08

October 8 was dominated by orchestration experiments. Builders stress-tested Claude Haiku 4.5 as a cheap/fast grunt model — @WhiskeyATX swapped his Hermes COO over from Opus 5.5 and found it dramatically faster, @samb69 worked for 10 minutes before realizing he was on Haiku, while @jcartu warned it may still be too dumb to orchestrate reliably. @thewildzeno detailed his Orca + OMP + Jev stack with sub-orchestrators per project and 10–20 parallel agents, and another builder wired Grok, Claude Code, ChatGPT/Dots, and Muse together via Slack and a Notion kanban to watch them edit each other's work in real time. Hermes-in-Orca with Opus 5.5 as orchestrator and GLM 5.3 as coder is emerging as a repeatable pattern. Model economics stayed front and center. @mj888g burned ~$50 on Grok in minutes with no token visibility, pushing debate toward Hermes or OpenClaw. Anthropic shipped monthly API credits to Claude anonymous subs (claimable via Console), prompting rewiring of low-load apps to Opus 5.5 and Haiku 4.5. When Opus and Grok refused persuasive landing-page copy, members turned to GLM, Kimi, and DeepSeek — with GLM the crowd favorite for coding and copy, often paired with Opus as critic. @thewildzeno also dropped a 7-step SaaS cloning playbook, leading with 'look for reasons NOT to clone it.' On the creative side, a detailed teardown of a long-form AI video ad workflow covered still plates from fixed cameras, code-driven pushes and pinch zooms, realistic UI chrome (timestamps, scrub bars), word-level TTS captions in 4:5 safe zones, and an Opus critic pass against storyboard. @jcartu showed a near-photoreal full-body 3D avatar demo and floated it for live casino dealers. On ads, @rockdm and @samb69 found Meta geo-exclusions leaking >50% of traffic, with adset-per-geo via API and naming target states in copy as the fix. Byteplus surfaced as the upstream behind Fal and Kie for Seedance, with Higgsfield reportedly ~half Kie's price.