NEW OpenClaw Update is INSANE!

OpenClaw update 223 ships a 1M-token context window for Claude Opus and Sonnet 4.6, plus a Kilo hardening pass — but the creator still warns you'll blow past the $30/mo hobbyist budget the moment you route everything through Opus.

1M-token context, beta only on Opus/Sonnet

The headline feature in v223 is a beta 1,000,000-token context window on Claude Opus and Sonnet. The creator calls it out as the upgrade that actually fixes a real pain: agents that "forget" mid-conversation because the entire chat history is re-sent each turn. His rough benchmark: 1M tokens fits "maybe like a quarter of the Harry Potter book." It's labeled beta, and he flags it as expensive — a recurring warning, not a footnote.

Model freedom: route by task, not by default

The slide is titled Model Freedom, and the takeaway is to stop sending everything to Opus. Use Opus for "deep thinking" — life planning, restaurant reservations, scheduling. The MiniMax mishap with the car-wash scheduling is cited as a concrete reason not to downgrade that workload. For dashboards, prototypes, and "build something fast" tasks, he suggests routing to Kimi or MiniMax on the fly to keep the bill down.

Sub-agents: one orchestrator, not eight

Agent orchestration is positioned as the workflow to use, not a new toy. The creator's rule: tell the bot "you are the orchestrator. You command the sub-agents. I don't want to interact with the sub-agents." He mocks a partner who was juggling 8 spawned agents manually — "don't try to set up your own fantasy company."

Cron jobs, dashboards, and VPS hosting

Cron jobs got cleanup work; the creator's own failure mode was stale jobs from older schedules firing all at once. His fix: prune the old cron list, then ask the agent to make a dashboard of all cron jobs so you can verify what's actually running. Dashboard access hinges on hosting — run it on a VPS, not a MacBook, or the local port won't be reachable from outside.

Security, Kilo, and the "95% casino" rule

The update is delivered by literally asking the bot: "Hey, update yourself to latest open claw." He says it works 95% of the time, and the other 5% you'll be debugging — "when you're using AI, you're playing casino." The companion Kilo update is described as a hardening pass; the creator admits he doesn't use Kilo himself and uses Anthropic more. Security advice stays conservative: keep the agent behind a VPS, "don't let it on every aspect of your life yet," and skip the Apple Watch integration for now.

Video understanding and self-criticism

Video input is back (moonshot dropped, vision is back in). The creator is blunt: it's expensive because of the compute needed just for a transcript, and "it's not going to watch the Avengers and write you the next chapter just yet." On the upside, he notes Opus is now proactively flagging unreliable data — "this data might not be reliable. I've never seen it. I only see it today" — which he frames as Stark learning to detect its own hallucination rather than bluffing through it.

Cost reality check

The $3–5 to $30/mo API figure is called out as a hobbyist floor. The creator admits he and his team spend "way more than that" running Opus for content production, and the next company-wide push is explicitly a cost-saving experiment — picking cheaper models per task instead of a single default.