Vercel Sandbox Drives: agents get persistent disk across runs

Mount up to 16 TiB into a sandbox so agents stop rebuilding each run. Plus an open 7B robot model, cheaper agent microVMs, and Aion 3.5's doubled context.

Nowline SEP 24 7:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Persistent disk you mount into a sandbox

    Vercel Sandbox took Drives into public beta: persistent volumes you mount as a directory, 1 TiB by default and up to 16 TiB, so an agent reuses its workspace, models, and dependency trees instead of rebuilding them every run. Storage is $0.05/GB-month (Hobby includes 15 GB); create one with Drive.getOrCreate() and mount it at startup, up to 4 per sandbox.

  • An open 7B world-action model for robots

    Black Forest Labs released FLUX 3 Action, an open-weights 7B model that turns a camera feed, robot state, and a text instruction into physical actions. It tops the RoboLab-120 benchmark at 42.92% — 6.1 points over the best prior open model — at roughly half the parameters and up to ~3.95x faster. Weights, code, the fine-tuning recipe, and a worked example on the low-cost SO-101 arm ship through Hugging Face LeRobot; about 200 teleop episodes fine-tunes it to your own robot.

  • Agent microVMs at about a third of the going rate

    Prime Intellect opened Prime Sandboxes — full Linux microVMs (own kernel, Docker Compose, background services) built for agentic RL and long-running agents. Promo pricing through Dec 22 is $0.02/vCPU-hour and $0.0125/GiB-hour, so a 1 vCPU / 1 GiB / 32 GiB box runs about $0.039/hour, roughly a third of rival sandboxes. It ships with 365K+ prebuilt environments, the largest such catalog.

  • Aion 3.5 doubles context, holds the price

    AionLabs shipped Aion 3.5 and 3.5 Mini on OpenRouter, doubling context to 262K tokens at the same rates as 3.0 — $3/$6 per M for 3.5, $0.70/$1.40 for Mini. Both are GLM-based multi-model systems tuned for roleplay and storytelling, reason on every call at three effort levels, and take tools and JSON; no image input, and no published benchmarks yet.