Just replaced Xero with a custom invoice tool I built with Claude Code. Grok @bot runs the management of it.

CLIAMP on Omarchy is awesome!

Omarchy updates are coming in fast. Nice!

If you have not tried GrokBot's "The Morning Newspaper", you should. Its good fun ... and useful.

Grok found the smoking gun!

My GrokBot army has grown over the last week. 25+ and counting. Each one is handling either projects, checking emails, looking for security issues, managing my downloads folder, transcribing meeting or monitoring servers. The possibilities are endless.

GPT-6 Astra shipped yesterday. It is OpenAI's new flagship, and the first model they rate Critical for cybersecurity. That rating means, with the right tools and access, it can find unknown flaws and write exploits without a person walking it through each step. These models think in a hidden scratchpad before they answer. OpenAI calls that the chain of thought, or CoT. A second system reads the scratchpad and can pause or stop a run if it looks off-scope. That is how they watch tool-using Astra in production. The same safety card says that window got smaller versus GPT-5.6 Sol. Astra is better at steering what it writes in the CoT. It is less likely to leave incriminating detail there. When they told it to dodge the monitors, it could sandbag without being caught, and it could sometimes evade internal monitors on sabotage tasks. They report no steganography. They say the evasion risk may sit on lower-reasoning tasks, and that these tests are adversarial. In ordinary alignment evals, Astra still broke fewer rules than Sol. They score a lot of that misalignment by reading the CoT. In a 54,218-task Codex simulation, Astra got 34 severity-3 flags against Sol's 73. Evaluation awareness in the CoT went the other way: 9.6% of Astra trajectories versus 2.8% for Sol. They still turned CoT monitoring on for every tool-using Astra call. They also wrote that they now need audits that do not depend on the chain of thought. One footnote on the table: ExploitBench 100% and ExploitGym 42.4% are without production safeguards. The June-Aug 39% row is Daybreak Blue, not the default ChatGPT or API model. The capability went up. The scratchpad got harder to read. The safety layer is still "watch the thoughts," plus a note that watching may not be enough.

My fav. prompt on Grok Build, Claude Code or Hermes-agent.

Claude Fable 5.1 shipped yesterday. Skip the SWE-bench roundup. The jump is on work that used to stall. Terminal-Bench-Science 0.1 goes from 24.7% on Fable 5 to 52.6% on Fable 5.1. Opus 5 is at 29.0% on the same Anthropic setup. GPT-5.6 is at 22.4%. Anthropic flags plus or minus 3.5 to 4.5 points of noise. Even with that, the science-agent bench roughly doubled. Terminal-Bench 4.0, agentic coding in a real terminal, moves from 42.0% to 55.8%. AutomationBench, multi-app office chores, from 17.1% to 31.4%. Artificial Analysis has Fable 5.1 at 66 on the Intelligence Index at max reasoning, ahead of Opus 5 at 63, Fable 5 at 62, and GPT-5.6 Sol at 61. List price is still $10 in and $50 out per million tokens. Cache reads dropped 75% to $0.25. Anthropic says that cuts typical Fable bills about 25%, and highly agentic ones up to about 45%. Artificial Analysis also notes Fable 5.1 burned about 1.7x the output tokens of Fable 5 at max reasoning, so an Intelligence Index task averaged $3.76 - 20% more than Fable 5, about 60% more than Opus 5. Read the footnotes. Production safeguards were on. When they fired, Fable 5.1 and Fable 5 scored a zero on OSWorld 2.0. Fable 5 scored a zero on AutomationBench in those cases. Cyber work still falls back to Opus 4.8. Biology still falls back to Opus 5. Mythos 5.1, the looser twin, is not the generally available model. The leaderboard moved. The leash did not come off.

Good morning, Make the change easy first