Claude Code Plans and Billing: What Actually Controls a Session

Last updated: August 17, 2026

Claude Code is Anthropic’s agentic coding client for the terminal, IDE extensions, a desktop app, and the web. A paid Claude plan includes it. The same client can also run on an Anthropic API key, a cloud provider credential, or an organization gateway, and that choice changes which meter the session draws from and which agreement applies.

Verdict in one line: a Claude subscription covers Claude Code cleanly only while the session is actually authenticated with that subscription.

Sources: Anthropic, Claude Code overview, checked 17 August 2026 · Anthropic, Claude Code authentication, checked 17 August 2026

What a paid plan includes

Anthropic’s plan comparison marks Claude Code as available on Pro, Max 5x, and Max 20x, and unavailable on the free plan. The pricing FAQ says the same thing and adds that Claude Code shares the plan’s usage pool, so terminal work and chat draw from one allowance.

That settles access. It does not settle the operational questions a buyer needs next: which credential a given session will use, which billing account receives the usage, and which agreement covers the code that goes into the prompt.

Prices are worth reading on your own screen. Opened from Japan on 17 August 2026, the page showed Pro at $22 billed monthly, $18 per month on annual billing at $220 up front, and Max from $110, under a footer line stating that prices include 10 percent JCT. We recorded that as one observation from one location and are not generalizing it into a rule about how the page behaves elsewhere.

Sources: Anthropic, Plans and pricing, observed 17 August 2026

The credential order in local sessions

For terminal and IDE sessions, Anthropic publishes the order in which Claude Code picks a credential when more than one is present. Cloud provider variables come first. Then ANTHROPIC_AUTH_TOKEN, then ANTHROPIC_API_KEY, then an apiKeyHelper script, then CLAUDE_CODE_OAUTH_TOKEN, then Anthropic profile and federation credentials. Subscription OAuth from /login, described as the default for Pro, Max, Team, and Enterprise, is seventh.

The documentation is direct about the consequence. If you hold an active subscription and also have ANTHROPIC_API_KEY set in your environment, the key takes precedence once approved. Anthropic’s Pro and Max support article states that usage then bills at standard API rates, distinct from plan pricing.

Approval is asked once in an interactive session and remembered. A non-interactive run started with -p uses the key whenever it is present, with no prompt.

The narrow claim is worth stating carefully. Such a session does not draw from the subscription’s included allowance. The subscription itself remains active and can still be used elsewhere. What diverges is the plan badge on the account and the meter behind a particular session.

Sources: Anthropic, Claude Code authentication, checked 17 August 2026 · Anthropic, Use Claude Code with your Pro or Max plan, 11 June 2026

Surfaces do not share one behavior

Not every place Claude Code runs reads the same credentials, which means a route confirmed in one surface does not transfer to another.

The environment variables above apply to the CLI and to the surfaces that wrap it, which Anthropic lists as the VS Code extension, the Agent SDK, and GitHub Actions. Claude Desktop and cloud sessions do not call apiKeyHelper or read those variables at all; they use OAuth, with one exception for desktop sessions running a third-party inference configuration. Claude Code on the web always uses subscription credentials, and Anthropic states that setting a key in the sandbox environment does not override them.

A signed-in Claude apps gateway session sits outside the precedence list entirely. When one exists, the CLI authenticates with the gateway token even where a cloud provider variable is set, and the bearer token, API key, helper, and profile sources are not used.

Administrators can constrain this. Managed settings can pin the login method and the permitted organization, and Anthropic states that environment credentials are blocked at startup when that pin is in force, because organization membership cannot be verified for them.

Sources: Anthropic, Claude Code authentication, checked 17 August 2026

Terms and data by account and route

Anthropic’s Claude Code data usage page groups accounts in two. Free, Pro, and Max are consumer users. Team and Enterprise plans, API, third-party platforms, and Claude Gov are commercial users, described as maintaining existing policies.

For the consumer group, the page states that Anthropic will train new models on Free, Pro, and Max data when the account setting is on, and says this includes Claude Code used from those accounts. For the commercial group, it states that Anthropic does not train on code or prompts sent to Claude Code under commercial terms unless the customer has chosen to provide data for model improvement, such as through the Development Partner Program. That program is described as available only for the Anthropic first-party API, not for Amazon Bedrock or Google Cloud’s Agent Platform.

Retention follows the same split. Consumer accounts with model improvement enabled are given a five-year period and those without it thirty days. Commercial standard retention is thirty days. Zero data retention is described as available to qualified accounts on Claude for Enterprise, not included in the standard Enterprise plan, and enabled one organization at a time after an account team confirms eligibility.

Three paths sit outside that split and apply whatever the plan is. Transcripts submitted through /feedback, /bug, or /share are retained for five years. If you answer Yes to the optional session-quality follow-up, Claude Code uploads the conversation, any subagent transcripts, and the raw session log; the page states that known key and token patterns are redacted while source code and file contents are uploaded as-is, and that shared transcripts are kept up to six months. Claude Code also writes session transcripts locally in plaintext under ~/.claude/projects/ for thirty days by default, adjustable through cleanupPeriodDays.

None of this ranks the agreements against each other. These are stated defaults and retention periods. Ownership, confidentiality, liability, indemnity, and termination sit in the terms themselves and require a separate reading, which this briefing did not perform.

Sources: Anthropic, Claude Code data usage, checked 17 August 2026 · Anthropic, About the Development Partner Program, checked 17 August 2026

Which model a session resolves to

Default is a configuration result rather than a fixed model name, and at least four inputs decide it.

Account type is the first. Anthropic’s model configuration page gives Opus 5 as the default for Max, Team Premium, Enterprise pay-as-you-go, and the Anthropic API, and Sonnet 5 for Pro, Team Standard, and Enterprise subscription seats. Provider is the second: the same page shows the opus and sonnet aliases resolving to different versions on Amazon Bedrock, Google Cloud’s Agent Platform, and Microsoft Foundry than on the Anthropic API. Enterprise administrators supply the third and fourth, through an organization default model and through an availableModels allowlist that can substitute an older permitted version for a newer one.

Fable 5 is handled separately. Anthropic’s plan page describes it as included on Max, Team Premium, and Enterprise premium seats, capped at half the weekly allowance and drawn from the existing pool rather than added to it, and as running on pay-as-you-go usage credits from the first request on Pro and Team Standard. The pricing table shows the same split, with Fable reading “Usage credits” under Pro and “50% of weekly limits” under both Max tiers.

One row in that table sits oddly against the documentation. Context window reads 200k for Free, Pro, Max 5x, and Max 20x alike, while the model configuration page describes automatic upgrades to a one million token window on Max, Team, and Enterprise plans. The two pages may be describing different configurations. We are printing the mismatch rather than resolving it.

Sources: Anthropic, Model configuration, checked 17 August 2026 · Anthropic, Claude Fable 5 on your plan, checked 17 August 2026 · Anthropic, Compare features across plans, checked 17 August 2026

The approval default on new sessions

Anthropic’s permission modes page states that on Pro, Max, and Team plans the built-in starting mode for new sessions is auto, in which a classifier model reviews each action rather than the user. The change requires Claude Code v2.1.228 or later on macOS, Linux, and WSL, and v2.1.233 or later on native Windows.

The exceptions are listed on the same page and matter as much as the rule. Sessions started with -p, Agent SDK sessions, Enterprise plans, Console API keys, and sessions on Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry, Claude Platform on AWS, or a signed-in gateway all still start in manual. So do sessions where a settings file disables auto mode, and the first session after an install or an upgrade.

Anthropic set out its reasoning in an engineering post dated 25 March 2026. Users accept 93 percent of permission prompts, which the company presents as evidence that the prompt had stopped functioning as review. The same post reports that on a curated set of 52 real overeager actions, the two-stage classifier let 17 percent through, and states that for someone manually approving every action carefully this is arguably a regression, and that auto mode is not a drop-in replacement for careful human review on high-stakes infrastructure.

That evaluation ran on a Sonnet 4.6 classifier. The model configuration page states the classifier now defaults to Sonnet 5. We located no published evaluation of the current classifier using the same method, so the 17 percent belongs to March 2026 and to a model that is no longer the default.

Sources: Anthropic, Choose a permission mode, checked 17 August 2026 · Anthropic Engineering, How we built Claude Code auto mode, 25 March 2026 · Anthropic, Model configuration, checked 17 August 2026

Other routes to the same models

Anthropic is not the only vendor selling access to Claude models, and the entitlements differ by storefront.

GitHub Copilot meters with AI Credits at one cent per credit. Pro at $10 per month carries $15 in monthly credits, Pro+ at $39 carries $70, and Max at $100 carries $200. Its model table lists Haiku 4.5 and Sonnet 4, 4.5, 4.6, and 5 as available on Pro, while every Opus version and Fable 5 is marked unavailable until Pro+. Copilot Pro+ and Max can also delegate work to third-party coding agents including Claude Code, so the two products are not a clean substitute pair.

Cursor sells a free Hobby tier, an Individual tier from $20 with Pro, Pro+, and Ultra options, Teams at $40 per user, and custom Enterprise pricing. Its Pro description leads with extended agent limits and generous limits for Grok alongside frontier model access.

Neither is compared here on output quality. The point for a buyer is narrower: the same model family can arrive through several purchase structures, and the plan tier decides which models are reachable at all.

Sources: GitHub, Copilot plans, checked 17 August 2026 · Cursor, Pricing, checked 17 August 2026

Preflight before you approve a route

Anthropic documents a check that takes one command. Running /status shows the active login method, and an API key row appears in place of it when a key is in use. That is the fastest available confirmation of which credential a session holds.

It is a preflight, not an audit. Before a company repository is opened in Claude Code, five facts are worth recording for the surface a developer will actually use.

A screenshot showing Pro or Max is not sufficient evidence for this decision. It shows the plan, and the plan is one input among several.

Sources: Anthropic, Claude Code authentication, checked 17 August 2026 · Anthropic, Claude Code data usage, checked 17 August 2026

What this briefing does not establish

FAQ

Methodology

This is a Tier C briefing. No account was purchased, no session was run, no invoice was inspected, and no checkout or order form was captured. Every page linked in the sections above was opened on 17 August 2026, and each claim is attached to the section where its source appears.

Two limits on the evidence are worth naming. The price figures in section 01 are one observation from one location; we did not test how the page renders elsewhere, and we removed a second reading from an earlier draft because it could not be traced to a named observer with an independent capture. The classifier evaluation in section 06 is Anthropic’s own, published in March 2026 against a model that is no longer the default.

Two conflicts are preserved rather than resolved. Context window on the pricing table against extended context in the model configuration documentation. And Anthropic’s data usage grouping of third-party platforms as commercial users against the fact that a Bedrock, Google Cloud, or Foundry route runs under an agreement with that provider, which we did not open.

Corrections are published with a numbered log. If a figure here does not match your screen, send the capture and the date.

Full source list: Anthropic pricing · Authentication · Data usage · Permission modes · Model configuration · Pro and Max support · Fable 5 on your plan · Development Partner Program · Auto mode engineering post · GitHub Copilot plans · Cursor pricing

Verdict

Approve a route, not a plan.

For a personal project, subscription login is the simplest route, and it behaves the way the pricing page implies as long as no higher-ranked credential is present on the machine. Check /status once and the question is settled.

For a company repository, the plan badge is not enough on its own. An organization-owned Team, Enterprise, API, or approved cloud-provider route puts credentials, billing, and data settings under one owner, and managed settings can pin the login method so an environment variable cannot quietly redirect a session. A personal Pro account can run Claude Code, and it can also run it under the consumer data grouping without anyone deciding that it should.

For an assessment of the code Claude Code writes, this briefing is the wrong document. That needs a hands-on test, and it is not what Tier C evidence supports.

The controls described here are distributed across Anthropic’s authentication, pricing, data usage, model configuration, and permission documentation. Assembling them is the work this article did. Confirming them against your own configuration is the work it cannot do for you.