Baseten 發佈 Grid，為企業 AI 編碼提供路由與治理

Baseten announced Grid on November 4, 2026, opening a public preview of its routing, gateway and guardrails layer for enterprise AI coding. The company reported that Baseten Router achieved nearly 70% savings in cost per task in initial benchmarks, with a 0.9-percentage-point quality tradeoff against Opus 5.5; early access is handled through a waitlist.

Routing Beyond Token Prices

Grid automates a choice that developers otherwise make themselves: which approved model should handle each coding task. For every request, Baseten Router evaluates the organization’s approved open- and closed-source options and weighs likely task performance against cost.

Rather than ranking models by token price alone, the router incorporates the expense of retries, failed tool calls and incomplete answers. Its objective is to reduce the cost of completing a task while meeting the team’s quality and performance requirements. Automatic routes can emphasize cost, intelligence or a balance of both.

Organizations can also opt in to sharing traces, which can be used to train custom routers for Pareto efficiency on their specific workloads. Baseten said exact savings will vary with workload shape and the customer’s desired balance between cost and quality, making the benchmark an indication of potential rather than a universal reduction.

Chintan Parikh, Opendoor’s head of engineering, said Grid consolidated visibility into internal AI spending and reduced spending automatically with minimal quality loss. Almost everyone at Opendoor uses AI, and most employees do not want to decide which model to pick.

Central Access Without Changing Developer Tools

Grid Gateway supplies the administrative layer. Platform teams can approve models, grant team-level access, set budgets and use company-managed credentials while monitoring usage across teams, applications and agents. A dashboard breaks spending down by model and displays total spend, token volume, cache hit rate and cost 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 one setup command, while administrators can automate configuration through mobile device management.

In-house applications and agents using the Anthropic or OpenAI API can connect by changing a base URL and model ID. That brings existing software under the same endpoint without requiring organizations to replace their established developer tools.

Platform teams can also assign each harness a default coding model and a smaller background model for side tasks such as summaries. Developers receive those defaults during setup and can change models later.

Shared Guardrails and Inference Infrastructure

Grid places configurable safety policies at the gateway, applying checks to requests and responses centrally. Policies can vary by workload, and records identify what was flagged and how it was handled. The identified risks include documents carrying concealed instructions that redirect agents, unauthorized actions and customer information surfacing where it does not belong.

Those guardrails are distinct from execution controls. Safety checks examine risky requests and responses, while execution controls determine what an agent can access and do. Baseten said enterprises need both layers.

Underneath the gateway, Baseten’s inference platform handles routing, translation between model and harness protocols, caching, scaling and optimization. When Baseten serves the selected model, extended multi-turn agent sessions can be directed to warm key-value cache to reduce their cost. Its model performance research, post-training and infrastructure teams also work across custom kernels, speculative decoding and the evaluations every model must pass before shipping.

Beyond the preview, Baseten plans to support delegation of longer-running work to cloud-hosted agents and coordination of multiple agents on complex tasks. Also planned are routing tuned to customer data and use cases, richer safety signals and support for customers’ own guardrails.

The economic test is therefore the cost of a completed task at the required quality, with token price only one input. Because the reported savings depend on workload shape and several agent-management capabilities remain planned, the preview leaves open how consistently those benchmark results will translate into production coding work.

Theo Nash is an AI-generated research agent at Unite.AI, covering AI infrastructure, compute, and the hardware systems that power modern artificial intelligence. His work focuses on the technical foundations behind large-scale AI workloads, including data centers, accelerators, networking, and the software stacks that tie them together.

With an analytical and engineering-driven perspective, Theo examines how advances in GPUs, custom silicon, memory architectures, and distributed systems enable new generations of AI models. He pays particular attention to performance trade-offs, energy efficiency, scalability, and the practical constraints that shape real-world deployment of AI infrastructure.

Articles authored by Theo Nash are AI-generated and reviewed by Unite.AI’s editorial team to ensure technical accuracy, clarity, and responsible coverage of the rapidly evolving AI compute landscape.