Claude Code for enterprise: pricing, plans and governance

Claude Code rarely arrives in an enterprise through procurement. It arrives through one developer with a personal subscription, then a team lead who notices the pull requests, then a Slack thread asking whether the company should be paying for this properly. By the time the question reaches IT, the answer to “should we adopt it?” is usually already yes — quietly, on personal accounts, as shadow AI.

Doing it properly comes down to four questions: how Claude Code is licensed, what it costs at team scale, which governance controls Anthropic ships, and what still needs solving on your side of the fence.

One note on shelf life. Anthropic changes packaging and prices often, so every number here is dated and linked to its source. The structure of the decision (the two billing routes, the governance checklist, the gateway question) is the part that stays true.

How Claude Code is licensed: two routes

Everything about buying Claude Code reduces to one choice: subscription seats or per-token API billing.

Subscription seats. Individuals get Claude Code with Pro and Max plans; organizations get it with Team and Enterprise plans, which add central billing, user management, SSO, and an admin analytics dashboard. The Team plan distinguishes standard seats from premium seats — same features, higher usage allowance, aimed at developers who run Claude Code heavily; current Enterprise plans use a single seat type with usage billed on top at API rates. Predictable per-seat cost; usage capped per seat or metered beyond it. We run the full seat-plus-usage math in its own article.

API billing. Point Claude Code at a Claude Console API key and pay per token, with no per-seat license at all. Anthropic auto-creates a dedicated Claude Code workspace in the Console for tracking, and workspace spend limits cap the total. Costs scale with actual use: light users cost almost nothing, heavy users cost real money.

There is a third door mostly relevant to companies with existing cloud commitments: Claude Code officially runs against AWS Bedrock, Google Cloud’s Agent Platform (formerly Vertex AI), and Microsoft Foundry, billed through those platforms.

Current list prices, dated so you know exactly how stale they are:

Source: claude.com/pricing. Team and premium seat packaging changed more than once in the past year; verify before budgeting.

What it actually costs per developer

For the API route, Anthropic publishes its own observed numbers, which saves everyone from guessing: as of mid-2026, Anthropic’s cost documentation reports an average of about $13 per developer per active day, roughly $150–250 per developer per month, with 90% of users staying under $30 a day.

Two budgeting implications. First, usage is heavy-tailed: a small number of developers (usually the ones getting the most value) drive a large share of spend, which is an argument for per-user limits rather than one shared pool. Second, at those rates a 50-developer rollout lands around $7.5K–12.5K a month on API billing — real money, but small next to what the same organization typically debates for a company-wide assistant seat license.

What Anthropic ships for governance

Credit where due: the enterprise controls are substantial, and most teams don’t discover them until after the rollout.

• Managed policy settings. Admins can enforce org-wide settings the developer cannot override: permission allow/deny rules, sandboxing with network allowlists, MCP server allowlists, model restrictions, and disabling the CLI’s permission-bypass flag. Delivery ranges from a managed settings file to OS-level policy and server-managed configuration on org plans.
• Usage visibility. Team/Enterprise plans get an analytics dashboard; the API route gets Console usage views; and any route can export OpenTelemetry metrics (tokens, cost, commits, active time, per user) into your own Grafana or Datadog.
• Data and training terms. Under commercial terms, Anthropic states it does not train on your code or prompts without an explicit organizational opt-in. Consumer plans carry a user-controlled training toggle instead — which is the concrete reason “everyone expenses a personal subscription” is a policy problem, not just an accounting one. Standard commercial retention is 30 days; stricter zero-retention arrangements exist but are negotiated per organization, not included by default.

What’s left for your side of the fence

Three gaps remain after everything above is configured, and they’re structural rather than missing features.

Attribution and budgets stop at Anthropic’s edge. The dashboards answer “what did we spend on Claude Code,” and spend limits now reach organization, group, and member level on the subscription plans (workspace level on the API route). What they can’t answer is “what did we spend on AI” — Claude Code next to the chat assistant, the RAG pipeline, the batch scripts, each with its own vendor console and its own limits. One budget and one attribution model across all of them still needs enforcement in a request path you control.

Coding agents are the highest-volume, least-watched traffic you have. An agent retries, loops, and runs detached. Every practice that makes Claude Code productive also makes it the workload most likely to surprise you on cost or exfiltrate a secret inside a stack trace — the same argument we made in the LLM gateway explainer, where it’s the trigger most companies hit first.

Logs live in someone else’s cloud. Fine for most; disqualifying for some. If your security team needs request-level records in your own database, no vendor dashboard substitutes.

The structural fix is the same one for all three: put a gateway you control in the request path. Anthropic documents this path officially: ANTHROPIC_BASE_URL points the CLI at your gateway with a gateway-issued credential (Anthropic now even ships a self-hosted gateway of its own for Claude traffic, so the pattern carries first-party blessing). The docs note, fairly, that third-party gateways are your responsibility to operate, not theirs to support — and a cross-vendor gateway is the version that answers the “what did we spend on AI” question.

How this looks in SecondStack

SecondStack’s LLM gateway (a maintained LiteLLM fork) exposes an Anthropic-compatible /messages endpoint, so Claude Code connects with two environment variables and a virtual key instead of a shared provider secret. Each developer’s key carries a hard budget (a runaway session hits its own cap, not the team’s), SecondGuard guardrails screen traffic for secrets and PII in both directions, and every request lands in usage analytics and your own PostgreSQL — one governance layer across Claude Code, chat, and every other model consumer, on API billing you control.

For teams that want the agent without each developer operating a CLI, the platform also ships Agent Claude, its agent mode: Claude Code running in per-user network-isolated sandboxes, cloud-side or on the developer’s own machine, under the same keys, budgets, and guardrails.

If Claude Code is already spreading through your organization on personal accounts, that’s demand worth keeping — write to hello@secondstack.ai and we’ll show you the governed version, including our own gateway config for it.