Baseten Grid and Router: What the 72% Coding Cost Claim Measures

Baseten opened Grid in public preview on November 4, bringing model access, routing and spending controls into one place for enterprise coding teams. Its separate Router study reports 72% lower cost than sending every task to Claude Opus 5.5, with a pass rate 0.8 percentage points lower.

That result addresses a familiar problem: a team can buy an expensive model for difficult work and then use it for every small task too. Routing tries to make that choice automatically. The useful question is whether it keeps the savings after the work leaves a benchmark and enters a real repository.

What Baseten Grid adds

The Grid announcement describes three parts:

• A gateway for model access, team budgets and company-managed credentials.
• A router that selects among approved models using expected task quality and cost.
• Central guardrails for requests and responses.

Baseten says developers can connect Claude Code, Codex, OpenCode and Pi. This gives a platform team a place to govern several tools without asking everyone to adopt the same editor.

The operational benefit is broader than a lower token bill. If model access and spending live in one place, a team can see which workflows consume the budget and decide which ones justify the most expensive option. That visibility is useful even before automatic routing proves its value.

What the 72% benchmark result means

The Router technical report evaluates routing between Opus 5.5 and GLM 5.3 Flash. Its held-out tasks come from DeepSWE 1.1, SWE Atlas QnA and Terminal-Bench 2.1, with pass rates averaged over repeated runs.

The study reports the 72% saving and 0.8-point pass-rate difference. The Grid product announcement gives slightly different figures: nearly 70% savings with a 0.9-point tradeoff. These are Baseten's benchmark reports, not independently reproduced results or a promised reduction in a customer's bill.

Baseten offers Cost, Balanced and Intelligence thresholds. It also notes that the competing routers in its comparison do not all have identical model pools. That makes the result a comparison of the tested services and settings, rather than an isolated test of routing algorithms.

A small average quality difference can still matter if failures cluster in the work a team cares about most. A router that handles routine edits well but struggles with migrations could look attractive across a mixed test set while creating expensive review work in that team's release process.

Preview access and security limits

Grid's public-preview entry point is a waitlist. The launch also puts cloud-hosted long-running agents and multi-agent coordination on the roadmap. Those should remain separate from what a team can evaluate in the current preview.

Before budgeting a rollout, confirm the commercial terms, model availability, request limits and data-processing conditions. A benchmark percentage is not a price sheet, and moving requests through a gateway adds another service to the application's data path.

Gateway checks also need to sit alongside controls on what an agent can actually access. A clean model response does not establish that a later shell command, file upload or infrastructure change is authorized. Keep credentials scoped and test the execution environment as well as the model layer.

How to test routing on your own work

Start with a fixed set of recent tasks and compare the routed workflow with the team's existing setup. Include routine changes, ambiguous bug reports and tasks that previously required human intervention.

Measure the cost of a completed, accepted change. Count retries, failed tool calls, review time and latency, alongside whether the tests pass. Keep model versions and routing settings in the record so that a later change in results can be explained.

The strongest outcome would be fewer expensive model calls without moving the cost into developer cleanup. That is what would make routing useful as part of an engineering workflow, beyond a good-looking benchmark chart.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI infrastructure teams turn technical launches and benchmark results into clear stories developers can use.

Book a 30-minute call.