JetBrains' 10× AI Bill Needs a Traffic Cop • The Weighted Average

JetBrains says its AI development expenses rose roughly 10× in six months, while most developers used three to five tools per month. Its Central CLI is the company’s answer: route Claude Code, Codex, and Gemini CLI through one governed layer without forcing engineers onto one model. The near-term lesson is not “buy JetBrains”; it is “put the traffic on a ledger before capping the tools.”

The bill rose because choice became infrastructure

JetBrains’ account of its first spending controls is unusually candid about the source of the problem. Developers kept using the IDEs, then added terminal agents, agentic development environments, and integrated tools. Most used 3–5 AI products in a month. Starting in January, usage nearly doubled month by month; JetBrains reached 150 Claude Code seats and moved to an Enterprise plan priced against API usage.

The company considered narrowing the approved list to one or two tools, then rejected that as a false economy. A developer may prefer Claude Code on Opus one week, Codex the next, and a different mix after a new model release. Standardization would simplify procurement, but it could also freeze the organization below the frontier of its own best workflow.

The first control attempt was manual. JetBrains says staff spent four days opening provider consoles, downloading data, and grouping usage by department and business unit. Dashboards improved visibility but did not conveniently enforce spending limits. That distinction is the operator gap: knowing that a team spent more is retrospective accounting; routing requests through a policy point can change the next request.

JetBrains repurposed a developer’s CLI wrapper into Central CLI and says it rolled the tool out internally in a couple of months. More than 1,000 developers switched within a few weeks, exposing Windows terminals, remote machines, authentication, and other edge cases. The product is therefore not just a billing page. It is a proxy in the execution path, with the reliability and availability obligations of infrastructure.

The Central CLI quickstart names three supported agents—Claude Code, Codex, and Gemini CLI—and says requests are routed through JetBrains Central. That is a meaningful coverage boundary, not proof that every agent or personal subscription is governable. The company says four more agents were in internal beta and niche setups remained out of scope.

Today’s local-agent lead and the adjacent Claude Code control-plane analysis make the same architectural argument from the safety side. A prompt or a promise is not a policy until the runtime can observe and enforce it. For AI spend, the policy question is not which model developers like; it is which traffic the organization can attribute, cap, and review.

A proxy is useful only if the ledger is honest

JetBrains’ management docs define each AI Credit as $1 in the billing currency. Combine that unit with the company’s 10× six-month increase and an illustrative baseline becomes legible: if a developer’s earlier workload cost one credit, the comparable current workload would cost roughly ten credits, or 9 additional credits above the original dollar. This is a normalization, not JetBrains’ per-user bill; the source does not publish the underlying dollar baseline or a causal decomposition of the tenfold rise.

The AI management documentation says organizations can assign licenses, control AI Credit limits and spending, and allocate top-up credits with monthly limits for individual users. Quota is consumed first; top-up credits begin after it is exhausted. That gives managers a control surface without asking them to predict every model’s token price.

The accounting is not complete by default. JetBrains’ consumption documentation includes Junie, Claude Agent, Codex, AI Assistant, and uncategorized credit-consuming tools, but excludes in-editor features that do not consume credits and BYOK usage billed directly by a provider. A dashboard can therefore be accurate about its own credit rail while missing the largest spend on a developer’s personal key or a tool outside the proxy.

The July Teams and Organizations announcement frames Central as vendor-agnostic governance: shared context, managed execution, policy, analytics, and cost attribution around the tools developers already prefer. That is the right product shape for a fast-moving model market. It also means procurement should ask exactly which providers, endpoints, prompts, logs, and data flows are covered before treating “vendor-agnostic” as a security property.

The strongest counterpoint is economic. JetBrains says the CLI is intended for teams with heavy API-based usage across multiple tools. If a team is still spending modestly, adding a proxy, account, network dependency, and credit system can cost more in friction than it saves. The company also has an incentive to make its own credit layer the center of the workflow, even while it promises open integrations.

The verdict should therefore be conditional. Teams with three or more active agents, rising API bills, and no trustworthy per-user attribution should run a bounded Central CLI trial alongside their existing path. Record request volume, provider, model, token or credit cost, latency, failure rate, developer wait time, and BYOK spend. Teams with one stable tool and low usage should first instrument that provider directly.

Evidence that would change the call is concrete: a complete reconciliation between Central credits and provider invoices, a measured reduction in cost per validated change, no material increase in agent latency or outage exposure, and coverage for the tools developers actually use. If the proxy merely moves spend into a prettier dashboard, it is administration. If it turns a tooling zoo into an enforceable, model-neutral budget without slowing the work, it is a control plane worth keeping.