Introducing Together Code Sandbox & Together Code Interpreter: SOTA code execution for AI

Together AI wants to solve a problem that's been quietly annoying every developer building AI coding tools: once your model spits out code, where does it actually run? Today the company shipped two answers. Together Code Sandbox is a customizable microVM environment for full development workflows, and Together Code Interpreter is a leaner, API-driven tool for just executing Python and getting results back.

The numbers Together is touting are the headline here. Sandbox VMs boot from a snapshot in 500 milliseconds at the 95th percentile, and cloning one takes under a second. You can hot-swap VM sizes from 2 to 64 vCPUs and 1 to 128 GB of RAM, and storage persists through a git-versioned filesystem. HeroUI, which built an AI chat product on top of it, says the platform took their VM startup time from roughly 2 minutes down to under 2 seconds, and shrank total build time for HeroUI Chat from five months to two weeks. Founder Junior Garcia says they had a working MVP in a couple of hours instead of waiting months to see whether an idea even worked.

Code Interpreter is the simpler sibling: one API call runs Python code inside an isolated sandbox, sessions last 60 minutes and can be reused, and pricing is a flat 3 cents per session. It's aimed squarely at agentic apps and, more interestingly, at reinforcement learning pipelines where generated code needs to be tested at scale to compute rewards. Agentica, the Berkeley AI Research and Sky Computing Lab project behind DeepCoder-14B-Preview, used it to run over 100 concurrent sandboxes and thousands of code evaluations per minute during training — a workload that would be brutal to manage with hand-rolled infrastructure.

The pitch, stripped of marketing language, is that Together is trying to become the execution layer underneath the current wave of AI coding products, the unglamorous plumbing that AI IDEs, SaaS platforms and RL researchers all need but don't want to build themselves. Sandbox pricing runs at $0.0446 per vCPU-hour plus $0.0149 per GiB RAM-hour, competitive enough that it's clearly meant to undercut teams currently duct-taping together their own container fleets.

Whether this becomes a real category or just a feature that gets absorbed into bigger platforms is the open question. But the RL training use case in particular — running thousands of parallel code evaluations to train coding models — feels like exactly the kind of infrastructure need that's about to get a lot bigger, not smaller.