Baseten launches Grid to route and govern enterprise AI coding – Unite.AI

Baseten announced Grid on November 4, 2026, opening a public preview of its routing, gateway, and guardrail layer for enterprise AI coding. The company reported that the Baseten Router achieved nearly 70% cost-per-task savings in initial benchmarks, with a quality trade-off of 0.9 percentage points compared to Opus 5.5; early access is managed via a waitlist.

Routing beyond token prices

Grid automates a choice that developers would otherwise make themselves: which approved model should handle each coding task. For each request, Baseten Router evaluates the open source and closed source options approved by the organization and evaluates the likely performance of the business against costs.

Instead of ranking models only by token price, the router incorporates the cost of attempts, failed instrument calls, and incomplete responses. Its goal is to reduce the cost of completing a task by meeting the team’s quality and performance requirements. Automated routing can emphasize cost, intelligence, or a balance of both.

Organizations can also choose to share traces, which can be used to train custom routers for Pareto efficiency on their specific workloads. Baseten said the exact savings will vary based on the shape of the workload and the client’s desired balance between cost and quality, making the benchmark an indication of potential rather than a one-size-fits-all reduction.

Chintan Parikh, chief technology officer at Opendoor, said Grid consolidated visibility into internal spending on AI and automatically reduced spending with minimal loss of quality. Almost everyone at Opendoor uses AI, and most employees don’t want to decide which model to choose.

Central access without changing development tools

Grid Gateway provides the administrative layer. Platform teams can approve templates, grant team-level access, set budgets, and use company-managed credentials by tracking usage across teams, applications, and agents. A dashboard breaks down spending by model and displays total spend, token volume, cache hit rate, and costs saved.

At launch, Grid supports models from OpenAI, Anthropic, Gemini Enterprise Agent Platform, Amazon Bedrock, and Baseten. Developers can connect Claude Code, Codex, OpenCode and Pi with a single configuration command, while administrators can automate configuration through mobile device management.

Internal applications and agents that use the Anthropic or OpenAI API can connect by changing a base URL and model ID. This brings existing software under the same endpoint without requiring organizations to replace their established development tools.

Platform teams can also assign each harness a default coding template and a smaller background template for subtasks like summaries. Developers receive these defaults during setup and can modify the templates later.

Shared guardrails and inference infrastructure

Grid places configurable security policies at the gateway, applying controls to requests and responses centrally. Policies can vary based on workload, and records identify what was flagged and how it was handled. Risks identified include documents containing hidden instructions that redirect agents, unauthorized actions, and customer information surfacing where it shouldn’t.

These guardrails are distinct from execution controls. Security controls examine risky requests and responses, while execution controls determine what an agent can access and do. Baseten argues that businesses need both levels.

Below the gateway, Baseten’s inference platform handles routing, translation between cabling models and protocols, caching, scaling, and optimization. When Baseten serves the selected model, extended multi-shift agent sessions can be directed to the hot key-value cache to reduce its costs. Its model performance, post-training, and infrastructure research teams also work on custom kernels, speculative decoding, and evaluations that each model must pass before shipping.

Beyond the preview, Baseten plans to support delegation of long-running work to cloud-hosted agents and coordination of multiple agents on complex tasks. There are also paths tailored to customer data and use cases, richer security signals, and support for customer guardrails.

The economic test is therefore the cost of an activity completed with the required quality, with the symbolic price of only one input. Since the claimed savings depend on the shape of the workload and several agent management features remain planned, the preview leaves open how consistently the benchmark results will translate into production coding work.