Baseten launches Grid to cut AI coding costs
Baseten has put Grid into public preview, combining model routing, a multi-provider gateway, and guardrails into one layer it says cut cost per coding task by up to 70% in early benchmarks.
Baseten announced Baseten Grid, an enterprise layer for routing AI requests, tracking usage, and controlling access across model providers. It is in public preview starting today.
Grid builds on Baseten Router, which we covered earlier. The router picks a model from an organization's approved list for each request. It weighs expected task performance and total cost, including retries, failed tool calls, and incomplete answers, not just token price. In Baseten's initial benchmarks, it saved nearly 70% on cost with a 0.9-percentage-point drop in quality compared with Opus 5.5. The company says real savings will depend on workload and on the cost-quality tradeoff each team picks. Teams can opt in to share traces, which Baseten can use to train custom routers for their workloads.
The Gateway covers models from OpenAI, Anthropic, Gemini Enterprise Agent Platform, Amazon Bedrock, and Baseten. Admins can set model access, budgets, and company-managed credentials for each team. Developers connect Claude Code, Codex, OpenCode, or Pi with a single command, and admins can push that setup through MDM. In-house apps that call the Anthropic or OpenAI APIs can connect by changing the base URL and model ID.
Guardrails apply configurable safety policies at the gateway and log what was flagged and how it was handled. Baseten says it is also working with Goodfire on Project Beacon, which uses signals from inside the model to catch unsafe behavior before output reaches a user or tool. The post does not give pricing or a date for general availability.