Playbooks · Teams tasks
Baseten Grid routes coding-agent traffic
Written November 8, 2026
Baseten's Grid is a routing, gateway, and guardrails layer for AI coding-agent traffic, in public preview. What it controls, and the savings claim behind it.
What it is
Baseten announced Grid on November 4, 2026, opening a public preview of a routing, gateway, and guardrails layer for enterprise AI coding. It sits between developers and the model layer, and it answers a question teams currently answer by hand, badly, and repeatedly: which approved model should handle this coding task. Early access is handled through a waitlist, and Opendoor is the named customer using it.
Grid has three parts. The gateway is the administrative layer: platform teams approve models, grant team-level access, set budgets, and use company-managed credentials, while a dashboard breaks spending down by model and shows total spend, token volume, cache hit rate, and cost saved. The router picks a model per request: for every request, Baseten Router evaluates the organization's approved open- and closed-source options and weighs likely task performance against cost. The guardrails put configurable safety policies at the gateway, applying checks to requests and responses centrally. At launch, Grid supports models from OpenAI, Anthropic, Gemini Enterprise Agent Platform, Amazon Bedrock, and Baseten, and it is built to work with Pi, OpenCode, Codex, and Claude Code without changing how developers work.
The savings claim, and why it is approximate
Baseten reports that its Router achieved nearly 70% savings in cost per task in initial benchmarks, with a 0.9-percentage-point quality tradeoff against Opus 5.5. That is Baseten's number, from its own benchmarks, and it should be read as one: Baseten itself says exact savings vary with workload shape and the customer's desired balance between cost and quality. The figure also moves between Baseten's own accounts of it — a company post cited 72% savings against Opus 5.5 — so treat the claim as "around 70%, on Baseten's setup" rather than a precise number. The honest test is cost per finished task at the quality the team requires, with token price only one input. A router that picks a cheaper model for a task the cheap model then fails is not saving anything.
What it is not
- It is not a sandbox. Grid routes and governs traffic between models. The isolated execution environment is Baseten Sandboxes, a separate product — do not conflate the two.
- Not a model. Grid picks between models; it does not run one. The quality of the work still depends on the model the router chooses and the task it is given.
- Not a finished product. Several agent-management capabilities remain planned, and access runs through a waitlist. A preview that routes production traffic is a test, not a migration.
What to do first
- Measure cost per finished task now. Before routing anything, record what a representative set of coding tasks costs on the team's current default model, and what quality bar each task must clear. Without that baseline, no router claim can be checked.
- Send a slice of traffic through the preview. One team, one harness, a fixed task set. Compare cost per finished task and quality against the baseline — not token price, which is only one input.
- Set the budgets and approvals before the traffic. The gateway's value is the control: approved models, team access, budgets. Decide those in advance, or the router is just a cheaper default.
- Watch the tradeoff, not the headline. A 0.9-point quality drop is small on a benchmark and can be large on the one task type a team cares about. Check quality per task type, not in aggregate.
Limits
- This page reads Baseten's announcement through its launch coverage. No Grid run was performed here, so it says nothing firsthand about routing quality, latency, or cost on a real workload.
- The 70% savings figure is Baseten's, from its own benchmarks, and Baseten's own accounts disagree on whether it is 70% or 72%. No independent measurement of it exists in the material behind this page.
- No price for Grid is published in the launch coverage, and access runs through a waitlist. This page states that rather than guessing a price or a timeline.
- Several agent-management capabilities remain planned. This page treats them as planned, not as shipped.
Related playbooks
- Copilot Agent Merge: watch a PR and fix its blockers
- Currents Sessions: one link per agent task
- Draft a grant report an agent checked against your own numbers
- Audit agent blueprint creation in Entra
Sources
- Baseten Launches Grid to Route and Govern Enterprise AI Coding — Unite.AI
- Request access to Baseten Grid — Baseten
Back to: Small-team playbooks hub · Also: Baseten Sandboxes keep agents running for days · Model prices: what the major models cost · OpenRouter Auto Code: Opus quality, 74% cheaper?