DeepSeek, Moonshot, Z.ai and other open-weight models handled 56% of Vercel AI Gateway tokens in August, up from 7% in December, yet Anthropic still took 64% of spend, per Vercel's September AI Gateway Production Index. Moving routine calls to cheap open weights shrinks the token line of your budget; your frontier-vendor invoice will probably stay put, because the high-value calls stay on the most capable model. See where the 64% of spend comes from.

Stanford and Caltech's HomeBody lets GPT-6 Astra call a five-skill library directly, and a Unitree G1 tidied an unfamiliar kitchen with its perception stack on a single RTX 4090 laptop, per the project page. If the approach generalizes, robotics teams maintain one skill API instead of training a policy per robot, and upgrading the reasoning model upgrades the robot. Stanford lists the current costs itself: pauses between skills from Astra's latency, and finger servos that overheat on long runs. See the five skills Astra calls.

Naive-N0.5-Flash needs roughly 315GB of FP8 GPU memory just to hold the weights, per NaiveAI's model card, so most teams will call its hosted API instead of self-hosting it. If you're standing up the app or internal dashboard that calls that API, Cloudways handles the managed server, patching, and backups so you are not babysitting a VPS. Spin up a Cloudways server.

A bug-bounty researcher found a flaw that could expose a user's dedicated Muse VM, emails and files, per Reuters. Exploiting it took a malicious link and one click on "Allow." Meta cut the rating from SEV-2 to SEV-3, then built an isolated VM anyway and added an approval gate before Muse sends email or makes purchases, per BigGo Finance. If your team routes email or payments through any agent, that Allow prompt is the attack surface, so train people to read it. Read how the exploit and the new approval gate work.