Baseten launches Grid, a gateway that routes enterprise coding-agent requests across models to cut costs

Baseten, the San Francisco company best known for hosting and serving AI models, has moved into the control layer of enterprise software development. On November 4 it opened a preview of Grid, a product that sits between a company's coding agents and the models they call, combining request routing, a governance gateway and safety guardrails. Early access is handled through a waitlist.

Routing on cost per task

The core component, which coverage refers to as Baseten Router, assesses each request against the set of open and closed models an organization has approved and weighs expected task performance against cost. Baseten says it looks beyond token prices to include the cost of retries, failed tool calls and incomplete answers. Teams can tilt routing towards cost, capability or a balance of the two, and can choose to share traces to train routers tuned to their own workloads.

According to Baseten, its initial benchmarks showed cost per task nearly 70% lower than with Claude Opus 5.5, for a quality trade-off of 0.9 percentage points. The company itself cautions that savings depend on the shape of the workload and the balance a customer picks. Neither the benchmark tasks nor the methodology have been made public, and some headlines circulating with a 72% figure do not match what Baseten has stated.

Gateway, guardrails and integrations

Platform teams can approve models, give teams access, set budgets, use company-held credentials and follow spending by team, application and agent. At launch, Grid works with models from OpenAI, Anthropic, Google's Gemini Enterprise Agent Platform, Amazon Bedrock and Baseten itself. Claude Code, Codex, OpenCode and Pi can be connected with a single setup command, and internal apps built on the Anthropic or OpenAI APIs can switch by changing the base URL and model ID. Guardrail policies run on requests and responses at the gateway, targeting risks such as documents with hidden instructions or customer data appearing where it should not.

What is still missing

Baseten has not disclosed pricing. Several announced features, including delegating long-running work to cloud agents, coordinating multiple agents and support for customers' own guardrails, are described as planned rather than available. Opendoor is the only customer named so far. How well the benchmark savings hold up in day-to-day production coding remains an open question until independent results appear.