dream labs' robot model API: $3 an hour. each call returns 1.6 seconds of actions, so about $0.0013 per action. their own launch page: a research beta that "has not been validated for safety-critical use." the cheapest worker on the floor. still needs someone watching it.

Anthropic: Claude writes ~90% of Claude Code's code. June to September, merged PRs more than doubled. The share reverted fell 40%+. Their goal isn't "a human read every line". It's less human time per PR, with reverts as the guardrail. If reverts rise, you only moved the work.

Holo, an AI marketing autopilot for small businesses: $5M ARR in 252 days, 20,000+ paying customers. That's under $250 a year each. About $20 a month. Small businesses don't buy "AI". They buy marketing done, with an approve button. At $20 a month it has to be boring.

Founders: what's the most expensive copy-paste in your company? Not the most annoying one. The one where a typo costs real money. Automate that one first. Not the one that would look best in a demo.

evan you: frontier models estimate a task at 2 weeks, then do it in 2 hours. a study this summer: coding agents overestimate runtime ~3-7x. given a tool that reports elapsed time, they got it right almost every time. give the agent a clock. don't let it quote your project.

inception's mercury 3: $0.10 in / $0.35 out per million tokens. that's 50% off at launch. the regular price is $0.20 / $0.70. no end date published. budget at the regular price. treat the discount as a bonus, not your margin.

beacon, an open-source memory layer for coding agents, raised $5.5M. it records sessions across 20+ tools and turns fixes into reusable skills, after you review them. the catch: the setup wizard preselects their cloud. local is the opt-out. read the install screen.

The first feature of any automation: an alert for when it stops. The second: a name next to that alert. Not a team. A person. Automations rarely fail loudly. They just stop, and nobody notices until month-end.

Salesforce's Agentforce chief on personal AI agents: a person checks a delayed order once. Their agent "can call back every few hours until the order ships." At 2 a.m. too. Support volume used to be capped by customer patience. Agents don't have any.

make's new task mining records how your team works on the desktop and in the browser, then ranks what's worth automating. you pick which apps it sees. employees can see it running and pause it. fine. just tell your team before you switch it on, not after they find it.

Steve Huynh, ~20 years at Amazon: a senior engineer he knows now tags docs "AI;DR". AI; didn't read. "If a model wrote it, a model can summarize it." His rule: AI for decisions you can undo. A human writes and a human reads anything you can't. Works for automations too.

dhh on agent QA for omarchy: "we can finally replace all that toil with agents." the design detail I'd copy: the agent running the tests only gets a keyboard, a mouse and screenshots. it never reads or changes code. whoever checks the work shouldn't be able to change it.

notion's new AI credits: 1,500 a month per seat if notion picks the model. 500 when you pick it yourself. no rollover. run out and the work pauses until reset, unless an admin pays. meeting notes start using credits dec 1. the default isn't just easier. it's the bigger bucket.

Bob McGrew: when a design changes, "you can retrain a human in a day — but it takes three to six months" to change the automation. His startup Arda says complex workflows at Anduril went from up to 30 hours to ~2.5. The cost of automation isn't building it. It's changing it.

claude code's wait before the first token: 4.07s down to 1.33s (p90, model time excluded). the changes anthropic named: gzip the conversation every turn, start cloud sessions while you type. plumbing, not a smarter model. time everything around the model before you blame it.

OpenAI's new COBOL test: GPT-6.1 Sol migrates 86.6% of programs correctly at low effort. The previous model: 5.4%. Their warning: COBOL drops extra decimals unless told to round. A rewrite that rounds runs fine and moves money wrong. Done means same output, not "it runs".

elevenlabs' new voice agent models: $0.08 a minute for both, plus LLM usage. on tool-calling tasks, max succeeded 87.1% of the time. mini: 60.3%. for support calls, the cheap one gets expensive fast. the per-minute price isn't the bill.

If someone has to check every output, your automation is a draft generator. That's fine. Drafts save real time. Just don't price it, staff it or sell it like it runs on its own.

Gartner, 2,002 CIOs: only 31% have rules for when an AI pilot should not go to production. 17% have rules for shutting down low-value AI that's already live. Those that do are 2.3x and 2.6x more likely to be in Gartner's top group. Write the kill criteria before the kickoff.

minimax m3.1 flash: $0.15 in / $0.60 out per million tokens. deepseek's off-peak rate. artificial analysis: when it didn't answer right, 93% of the time it answered wrong instead of saying it didn't know. cheap tokens. confident mistakes. check it before it writes to a CRM.