What Is Baseten Grid? The Beginner's Guide to Routing and Governing Enterprise AI Coding
CodexFuse Team · 2026-11-05T11:15:53.339Z · 17 min read min read
What Is Baseten Grid? The Beginner's Guide to Routing and Governing Enterprise AI Coding
Baseten Grid is Baseten's gateway, routing, and guardrails layer for enterprise AI coding, and this guide answers the question what is Baseten Grid in plain terms. Baseten opened Grid in public preview on November 4, 2026, one day before this guide was written, and access is handled through a waitlist. Instead of letting every developer pick any model with any company key, Grid puts one managed layer between an organization's coding agents and the models they call, so platform teams choose the models, set the budgets, and see the spend.
The idea matters because coding agents turn model choice into a recurring bill. Every agent run burns tokens, and when dozens of developers each run Claude Code, Codex, OpenCode, or Pi with their own settings, the cost spreads across teams, tools, and credentials with no single owner. This guide explains what Baseten Grid does, how its routing works, what its guardrails check, what the headline saving claim actually means, and how to join the waitlist while Grid is still in preview.
What Is Baseten Grid, Exactly?
Baseten Grid is an organization-wide gateway that connects a company's coding agents to models on Baseten and connected providers, with admins choosing each team's models and budgets. A gateway is a single control point that every request passes through, and Grid uses that position to route each request to an approved model, enforce spend limits, and apply safety checks. Developers keep using the coding agents they already know, and they pick from routes their team admin set up instead of configuring models and keys themselves.
Baseten Grid explained in plain terms: the Baseten Grid gateway is a traffic controller for AI coding. Without it, every developer is a driver choosing their own road and paying their own toll. With Grid, the organization sets the roads, the speed limits, and the budget, while developers just drive.
If you run models yourself, you will know the other half of the job: getting the model file onto your machine safely. Before loading a downloaded checkpoint, compare its hash against the publisher's value, and the SHA-256 hash generator can produce the hash you need to compare. For a machine to run it on, VPS hosting gives you a server you can rebuild in minutes if an experiment goes wrong.
One honest note up front: Grid is one day old as of this guide, and no independent reviewer has used it hands-on yet. Everything here comes from Baseten's own docs, its launch announcement, and the first news coverage, and every saving figure is a vendor claim, not a measured result.
How Does Baseten Grid Work?
Grid works in four steps, and the vendor docs describe them as a loop an admin sets up once and developers then use every day. The admin chooses the models and the budgets, developers connect their agents, and everyone watches the usage together.
- Choose each team's models from Baseten Model APIs, frontier providers, or both, and add an auto route when a classifier should choose per request.
- Set a monthly spend limit for each developer and for the team as a whole, and raise one developer's limit when they need more.
- Connect each developer's coding agent with one setup command that adds the team's routes to the agent's model picker.
- Track spend, tokens, and cache hit rate by team, developer, model, and provider, then adjust models and limits as needs change.
A route is a named entry a coding agent can call. A direct route calls one model, while an auto route chooses among several direct routes. A connection is a set of provider credentials stored on a team, so the team's routes can call that provider's models.
What Is Baseten Grid Routing?
Baseten Grid routing is the part of Grid that decides which model answers each coding request, and it is the feature the launch coverage talks about most. For every request, Baseten Router evaluates the organization's approved open- and closed-source options and weighs likely task performance against cost. Instead of ranking models by token price alone, the router also weighs the expense of retries, failed tool calls, and incomplete answers.
An auto route can emphasize cost, intelligence, or a balance of both, according to the launch coverage. That choice matters because the cheapest model per token is not always the cheapest per finished task: a weak model that retries three times can cost more than a strong model that gets it right once. The router's job is to make that tradeoff per request instead of leaving it to each developer's guess.
| Route type | What it does | Best for | Who chooses the model |
|---|---|---|---|
| Direct route | Calls one model every time | Predictable work where one model is known to fit | The admin, when creating the route |
| Auto route | A managed classifier picks among several allowed models per request | Mixed work where the best model varies | The classifier, inside the admin's allowed set |
When Baseten serves the selected model, long multi-turn agent sessions can be directed to a warm key-value cache to reduce their cost, according to the launch coverage. A warm cache means the model reuses attention state from earlier turns instead of recomputing it, which matters for coding agents that hold long conversations with a codebase.
How Do Monthly Spend Limits Work?
A spend limit is a monthly cap, in US dollars, on a developer's or a team's Grid spend, and limits reset at the start of each UTC calendar month. Admins set a limit for each member and one for the team as a whole, and they can raise a single developer's limit when a task needs more room. Developers receive short-lived Grid tokens during setup, and spend limits apply to those tokens.
Grid meters spend every 15 minutes, so spend can pass a limit before new requests are blocked. That metering window is the maximum overshoot: a team burning 100 dollars per hour could run up to about 25 dollars past its limit in that window, since 100 times 15 over 60 equals 25. The docs also warn that a coding agent can stop getting responses after working fine earlier, and the first thing to check is whether the developer or the team hit a monthly limit.
| Limit | Set by | Resets | What happens at the cap |
|---|---|---|---|
| Per-developer limit | Admin | Start of each UTC month | That developer's new requests are blocked |
| Team limit | Admin | Start of each UTC month | The whole team's new requests are blocked |
| Metering interval | Grid | Every 15 minutes | Usage can take up to 15 minutes to appear, and spend can overshoot briefly |
Usage can take up to 15 minutes to appear in the usage view, so a dashboard can lag the real spend by one metering interval. Admins who watch a fast-burning team should treat the dashboard as near-real-time, not instant, and set limits with that 15-minute window in mind.
What Are Baseten Grid Guardrails?
Grid places configurable safety policies at the gateway, applying checks to requests and responses centrally, according to the launch coverage. Policies can vary by workload, and records identify what was flagged and how it was handled. The risks Baseten names 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, and Baseten says enterprises need both layers. Safety checks examine risky requests and responses, while execution controls determine what an agent can access and do. A guardrail can flag a prompt injection hidden in a document, while an execution control decides whether the agent may run a shell command at all.
| Layer | What it checks | Example |
|---|---|---|
| Guardrails (safety policies) | Requests and responses at the gateway | A document with hidden instructions that tries to redirect the agent |
| Execution controls | What an agent can access and do | Whether the agent may read a file, run a command, or call an outside service |
Organizations can opt in to sharing traces, which can be used to train custom routers for Pareto efficiency on their specific workloads, according to the launch coverage. A Pareto-efficient router is one that finds the best tradeoff between cost and quality for that organization's own traffic, rather than using a generic balance.
Baseten Grid vs Frontier Gateway
Baseten Grid vs Frontier Gateway is the comparison that confuses new readers most, because both are gateways from the same company, but they face opposite directions. Grid governs an organization's own developers and their coding agents. Frontier Gateway lets a company offer an OpenAI-compatible inference API to its own customers, with the company managing customer access, usage limits, and billing while Baseten operates the gateway.
The vendor docs draw the line directly: to give customers or applications access to models across providers with federated API keys, use Frontier Gateway. Grid is for organizations that want central control over coding agent models and spend, and it uses short-lived Grid tokens for developers, not federated customer keys. A model lab selling inference to its customers needs Frontier Gateway, while a company whose engineers write code with AI agents needs Grid.
| Product | Who it serves | Control point | Keys |
|---|---|---|---|
| Baseten Grid | An organization's own developers and coding agents | Admin-chosen models, team access, monthly spend limits | Short-lived Grid tokens per developer machine |
| Frontier Gateway | A company's customers and applications | Customer access, usage limits, billing events | Federated API keys under groups |
| Baseten Switch | One developer on one machine | Local routing for Claude Code and Codex CLI | The developer's own setup on that machine |
Baseten Grid vs Switch: Which One Is for a Single Developer?
Baseten Switch is the single-developer version of the same idea, and the docs put the difference in one line: Switch runs on one developer's machine. It is a local gateway for Claude Code and Codex CLI that maps model families to Baseten Model APIs slugs, compares Baseten spend with estimated costs from Anthropic or OpenAI, and can route some requests back to those providers.
Switch requires macOS 13 or newer on Apple Silicon or Intel, and it starts with a command that also installs a menu bar app. Grid, by contrast, is organization-wide: admins choose models per team, set monthly spend limits, and watch usage across teams, while developers join with one setup command. An individual developer who wants cheaper models in Claude Code starts with Switch or a direct provider setup, while a team that needs shared budgets and audit records needs Grid.
Which Coding Agents Work With Baseten Grid?
Grid supports Claude Code, Codex, OpenCode, and Pi at launch, and developers connect any of them with the same setup command. In-house applications and agents that already use the Anthropic or OpenAI API can connect by changing a base URL and model ID, according to the launch coverage. Pi, Droid, and OpenCode can also connect directly to the OpenAI-compatible Model APIs endpoint with a Baseten API key, without Grid, for teams that do not need governance.
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, which keeps routine background work on a cheaper model while the main coding work uses the stronger one.
| Agent | Connection | Credentials |
|---|---|---|
| Claude Code | baseten grid setup adds team routes to the model picker | Short-lived Grid tokens from the Baseten CLI |
| Codex | baseten grid setup adds team routes to the model picker | Short-lived Grid tokens from the Baseten CLI |
| OpenCode | baseten grid setup adds team routes to the model picker | Grid API key named for the machine and team |
| Pi | baseten grid setup adds team routes to the model picker | Short-lived Grid tokens from the Baseten CLI |
If you use Claude Code daily, the setup is worth understanding before you run it, and how to use Claude Code covers the agent's own workflow in detail. Setup replaces each coding agent's model and endpoint settings with the team's routes and does not save the previous values, so check the dry-run preview first. The same setup keeps MCP servers, plugins, and other settings untouched, which matters for teams whose agents rely on custom tools, and what MCP over MOQT is explains the protocol those tool connections can use.
How Do Developers Connect to Baseten Grid?
A developer connects by running baseten grid setup once, after an admin adds them to a team that has at least one route. Setup finds the supported coding agents installed on the machine, adds the team's routes to each agent's model picker, and configures Grid credentials for the machine. The prerequisites are simple: the organization uses Grid, the developer is on a team with a route, and the developer uses a supported coding agent on macOS or Linux.
Setup replaces each agent's model and endpoint settings with the team's routes and does not save the previous values, so a developer who customized their model setup should preview the change first with the dry-run flag. The command also keeps MCP servers, plugins, and other settings, so existing tool connections survive the switch. After setup, baseten grid status checks the connection, baseten grid sync picks up new or changed routes, and baseten grid teardown removes Grid from the machine.
One limitation is worth knowing before setup: Claude account connectors do not work while Claude Code points at Grid. The docs suggest configuring those tools as MCP servers in Claude Code instead, which keeps the same capabilities reachable through a different connection.
What Does the 70 Percent Saving Claim Actually Mean?
Baseten says its router achieved nearly 70 percent savings in cost per task in initial benchmarks, with a 0.9-percentage-point quality tradeoff against Opus 5.5, and that figure is a vendor claim, not a measured result. A 70 percent saving means the routed cost per task would be roughly 30 percent of the un-routed baseline, about one-third the cost, if the benchmark held for a given workload. The 0.9-point gap is a difference in benchmark points, not a 0.9 percent relative drop: if Opus 5.5 scored 80.0 percent, the router's mix would score about 79.1 percent.
Baseten itself 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. One news roundup also noted a 72 percent figure circulating in coverage with no source behind it, which is a reminder to check any saving number against the vendor's own wording. Until an independent team measures Grid on real workloads, the honest reading is that routing can cut coding costs a lot on some tasks while giving up a small amount of quality, and the size of both effects depends on the workload.
The supported model sources at launch are OpenAI, Anthropic, Gemini Enterprise Agent Platform, Amazon Bedrock, and Baseten, with other providers coming soon, according to the launch coverage. Billing follows the source: Baseten bills Model API usage, while connected providers bill usage to the account behind the team's provider key.
How Do You Join the Baseten Grid Waitlist?
Grid opened in public preview on November 4, 2026, and early access is handled through a waitlist, so no reader can sign up and use it immediately as of November 5, 2026. Baseten's announcement post points interested teams to its request-access page, and administrators can automate configuration through mobile device management once their organization is admitted. The waitlist is the only public route in: there is no self-serve signup, no public pricing page for Grid, and no free tier announced.
The launch was the first of three Baseten launches that week, with Sandboxes following on Thursday and Loops after that, according to the company's announcement. That pace matters for buyers, because Grid's planned features, including delegation of longer-running work to cloud-hosted agents, coordination of multiple agents on complex tasks, routing tuned to customer data, richer safety signals, and support for customers' own guardrails, are described as coming after the preview, not as available now. A team joining the waitlist should judge Grid on what the preview offers today and treat the rest as a roadmap.
The only customer named in the launch coverage is Opendoor, whose head of engineering said Grid consolidated visibility into internal AI spending and reduced spending automatically with minimal quality loss. Opendoor's engineers had been using Grid before the public preview, and Baseten's own post says its own teams use it too. No other customer names, customer counts, or adoption numbers were published with the launch.
Who Should Use Baseten Grid, and Who Should Not?
Grid is for organizations that want central control over coding agent models and spend, in the vendor docs' own words, and the fit is clearest for platform teams running many developers on several coding agents. Those teams get one place to approve models, grant team-level access, set budgets, and watch usage across teams, applications, and agents, with a dashboard that breaks spending down by model and shows total spend, token volume, cache hit rate, and cost saved. A company where almost everyone uses AI, and most employees do not want to decide which model to pick, is the customer Baseten describes.
Solo developers and small teams probably will not need this level of governance infrastructure today, in the words of the first commentary on the launch, and the docs agree by pointing individuals to Switch or a direct provider setup. A single developer has no team budgets to enforce and no audit trail to consolidate, so a local gateway or a direct API key is simpler and cheaper to run. Grid earns its place when model choice is a shared, recurring expense that no single developer should decide alone.
Teams that run their coding agents inside sandboxed environments face a related question about where agent work executes, and how to use E2B explains the sandbox approach for running agent code safely. For those teams, Grid answers who pays for the model calls, while a sandbox answers where the agent's code is allowed to run.
Sources
- Baseten Docs, Baseten Grid, updated November 4, 2026.
- Baseten Docs, Use models in coding agents, updated November 3, 2026.
- Baseten Docs, Connect your coding agent to Grid, updated November 4, 2026.
- RuntimeWire, Baseten launches Grid to route coding agent traffic and control enterprise AI spend, November 4, 2026.
- Unite.AI, Baseten Launches Grid to Route and Govern Enterprise AI Coding, November 4, 2026.
- Baseten on X, Introducing Baseten Grid, November 4, 2026.
- Explainx.ai, AI News Today, November 5, 2026, November 5, 2026.
- Baseten, Company summary, reviewed September 3, 2026.
Frequently Asked Questions
What is Baseten Grid in simple terms?
Baseten Grid is a gateway that sits between an organization's coding agents and the AI models they call, letting admins choose each team's models, set monthly spend limits, and watch usage in one dashboard. Developers keep using Claude Code, Codex, OpenCode, or Pi, and they pick from routes their admin set up. It opened in public preview on November 4, 2026, behind a waitlist.
Is Baseten Grid the same as Frontier Gateway?
No. Grid governs an organization's own developers and their coding agents, with admin-chosen models and monthly spend limits, while Frontier Gateway is a customer-facing gateway that lets a company sell inference to its own customers through federated API keys, groups, and billing events. Switch is the third product, a local gateway for one developer on one machine.
How much does Baseten Grid cost?
Baseten has not published Grid pricing as of November 5, 2026, and there is no per-seat fee or free tier announced. The billable quantities the vendor docs name are Model API usage, billed by Baseten, and connected-provider usage, billed by that provider. Access is through a waitlist, not a self-serve signup.
Can I use Baseten Grid today?
Only if your organization is admitted from the waitlist. Grid is in public preview as of November 5, 2026, one day after launch, and Baseten's announcement points interested teams to its request-access page. Individual developers who want cheaper models now should look at Switch or a direct provider setup instead.
Which models can Baseten Grid route to?
At launch, Grid routes to models from OpenAI, Anthropic, Gemini Enterprise Agent Platform, Amazon Bedrock, and Baseten, with other providers coming soon. Admins add open-weight models from Baseten Model APIs, frontier models through their own provider API keys, or both, and they can add an auto route that lets a classifier choose among several models per request.
