Baseten launches Grid, a router that sends coding tasks to open models to cut costs by about 70%
What's new? Baseten has launched Grid, a gateway, router and guardrails layer for companies running AI coding agents, in public preview from 4 November 2026. Its core is Baseten Router, which reads each coding task and sends it either to a frontier closed model or to a cheaper open-weight one, so teams pay frontier prices only when a task needs it. Grid works with models from OpenAI, Anthropic, Google and Amazon Bedrock as well as Baseten's own, and plugs into Claude Code, Codex, OpenCode and Pi.
The numbers: In Baseten's tests, routing between Claude Opus 5.5 and the open model GLM 5.3 Flash came within 0.8 percentage points of Opus's pass rate (69.6% vs 70.4%) at 72% lower cost, $0.38 per task against $1.34, on a held-out set spanning DeepSWE, SWE Atlas QnA and Terminal-Bench. Baseten says it was the only router it tested, against offerings from OpenRouter, Fireworks, Together AI and Applied Compute, that beat random routing by a statistically significant margin. The Grid launch post gives a more conservative "nearly 70%" saving with a 0.9-point quality cost, and notes savings vary by workload. These are Baseten's own benchmarks; no independent evaluation has been published.
Why it matters: Baseten argues that spending on AI coding is growing fast while developers default to the most capable model for every task, not the one best suited to it. A router that can quietly hand routine tasks to open models turns the shrinking gap between open and closed models into savings, and gives inference platforms like Baseten a way to sit between companies and the model labs. Grid also lets admins set team budgets, approved models and safety policies in one place.
Read more: Baseten (Router), Baseten (Grid), Unite.AI, Sarah Guo (X)
Image: Baseten Router cost vs pass-rate chart (Baseten)
