Together Code Interpreter

How Together Code Interpreter works

TCI exposes code execution as a simple API call: you create a session, an isolated environment with a 60-minute lifespan, and call it to run Python, streaming back stdout, stderr and evaluated output. Sessions are reusable and can build on prior executions, and the design targets high concurrency (hundreds of simultaneous executions), which is why it shows up in reinforcement-learning and agent-evaluation pipelines. It is Python-only today. For longer-lived or non-Python workloads, Together offers a separate Code Sandbox with usage-based vCPU and RAM pricing.

Pricing in practice & watch-outs

Billing is simple: $0.03 per session, where a session is a 60-minute reusable environment, so cost scales with how many execution environments you spin up rather than raw compute time. That makes budgeting easy for batch and RL workloads. The constraints are scope: Python only, no GPU in the interpreter, and no published cold-start figure, so latency-sensitive interactive use should be validated first. If you are already on the Together stack for inference, running generated code in the same place is the natural fit.