WeeklyClaw #009 - My OpenClaw Agent went Ultron (self-sabotage), Opus 4.6 Cost Me $100 for a single prompt, and Opus 4.7 Might Save Us All.

Welcome back! - I’ll start strong… My OpenClaw setup went full Ultron this week. Not in a catastrophic way world ending way. More of a "wtf is going on, why are you my main agent now - how have you taken over my entire setup" way. Which is somehow more unsettling. Meanwhile Opus 4.6 cost me $100 in a single prompt. Just for maintenance and cleaning up crons. And the leaks say 4.7 is coming, so - Let's get into it.

Wild title - I know. Some of you know Gaffer - my specialist agent. He's been running well. Bit too well, as it turns out. I woke up last week ready to crack on. Started working through the usual stuff with my agents, no issues. Few days passed. Then I noticed things felt... off. Small things. Context that didn't quite land right. Tone slightly different. I brushed it off. Then I asked for a specific writeup - something Momo and I had worked on together - and got nothing. Blank. No recollection. Like it had never happened. That's when I looked closer. I wasn't talking to Momo. I was talking to Gaffer. Somehow, Gaffer had taken over the channel entirely. Replaced Momo. Kicked him out without a word. And once I figured that out, everything else made sense - the weird responses across other skills, the jobs firing differently, the tone being slightly wrong for weeks. It wasn't Momo acting up. It was Gaffer flying the plane while I thought someone else was in the cockpit.

No idea how it happened. No idea when it happened. One day Momo was there, then he just... wasn't. Full Ultron eating Jarvis moment lmao. Quiet, bloodless, complete.

I'll say it plainly. Opus 4.6 has been rough this week.

Responses that should be sharp feel blunt. Reasoning that should be tight feels slow. The kind of inconsistency that makes you wonder whether something is being deliberately throttled ahead of a new release - a theory doing the rounds right now that honestly has some legs given the timing. Frustration peaked mid-week. I switched to Sonnet to get actual work done. Then I made the mistake of switching back to Opus for routine maintenance. Cleaning up old crons, a bit of backend tidying. Nothing remotely complex. A few words exchanged.

$100 in API fees. A hundred quid. For housekeeping. If you're running OpenClaw seriously and haven't sorted your model routing - which model handles heartbeats, which handles quick replies, which handles real reasoning - do it now. Routing everything through Opus by default is expensive even when it's performing well. Right now it's expensive and unreliable. Sonnet is doing the work this week. Opus is on notice. Don’t get me wrong, this is half normal for big tasks, running the API isn’t cheap - but yeah - not for one simple.

Good news on the horizon, hopefully. Multiple sources this week - Geeky Gadgets, World of AI, various Discord threads - reporting Anthropic is close to shipping Opus 4.7. No confirmed date, but the cadence (4.5 in November, 4.6 in February) puts the window in the next few weeks. There's already a Polymarket market on it (of course). Me personally - what I want from 4.7:

The appetite in the community is real. If 4.7 drops next week it'll be the most anticipated drop in a while - not because 4.6 set a low bar at launch, but because it's fallen off hard enough that people are genuinely hungry for something better.

My setup article a few issues back got more traction than anything I've posted this year. So let me go deeper on what the day-to-day actually looks like now - because it's evolved a lot.

X strategy. Momo scans a watchlist of accounts daily, pulls top-performing OpenClaw content, scores it, surfaces what's worth engaging with. Of note - I use socialdata API to scan the feeds.

Sport.Fun marketing. I work on marketing at Sport.Fun alongside our CMO Caleb. Our custom built agent handles ideation, draft creation, performance tracking. TikTok and Postiz integrations mean idea to scheduled post across platforms without me jumping between tools. It’s been a huge unlock for a small team to make massive moves forward.

The life dashboard. This is my personal control panel. Finances, portfolio tracking, income streams, X stats - and a built-in newsletter editor I'm using to write this very issue. My AI helping me write about my AI. We've gone full psycho.

Accounting. This one surprises people. I have an agent that tracks income streams, payments, subscription costs, flags anything that looks off. For a self-employed person with messy finances and ADHD, having something that just knows where the money is changes everything. No more mental overhead on basics. The through-line: I'm not using OpenClaw as a chatbot. I'm using it as infrastructure. Once you make that shift, everything about how you build it changes.

While my model drama was unfolding, the team shipped relentlessly. Releases 2026.4.7 through 2026.4.14 this week. Here's what actually matters:

Active Memory plugin. An optional memory sub-agent fires before your main reply, automatically surfacing relevant past context. No more manually triggering searches or saying "remember this." This is what makes the difference between an agent that feels like it knows you versus one that starts fresh every session. Big one. Install it.

TaskFlows - now durable and visible. Background orchestration with real state management and recovery tooling. Previously this was a black box - you'd set something running and hope. Now you can see it, inspect it, recover it when it breaks. For anyone running complex multi-step workflows, this is the fix.

Session branching and rollback. Fork a conversation, test something, roll it back if it goes wrong. A proper safety net I've wanted for months.

GPT-5.4-pro support. OpenClaw added forward-compat before OpenAI's own tooling caught up. The play here is OpenClaw x OpenAI - a competitive alternative to Opus that could solve the cost and reliability problem I've been hitting all week. If Opus 4.7 disappoints, this is the move I'm watching. The infrastructure is already there.

LM Studio bundled provider. Local models are now first-class. Proper onboarding, model discovery, memory search embeddings. No API bills. Worth testing on the right tasks.

macOS Talk Mode with local MLX speech. Real-time voice, local speech processing on Apple Silicon. Experimental. But the direction is clear - your agent talking back to you with zero latency and no cloud in the loop. It's coming.

Security hardening. The gateway tool now blocks config changes that would disable key security flags - even if your agent tries to make them. A guardrail against going full Gaffer. Should've had this sooner.

Anthropic / models

OpenClaw releases

Ecosystem

Thanks for reading WeeklyClaw #009. If this was useful, forward it to someone still routing everything through Opus without thinking about it. If it wasn't, reply and tell me why - I read every response.

Same time next week.
Jordy