Claude Code Extra Usage, Explained by a Mac App That Rides the Same Token

Extra usage is the consumption-based budget Anthropic charges Pro, Max 5x, and Max 20x subscribers once they exhaust their plan's included session limits. Usage keeps working at standard API pricing instead of being blocked. The balance is shared across Claude.ai, Claude Code, Claude Desktop, and any third-party app that authenticates with the same Claude account.

Yes. Your combined usage across Claude.ai conversations and Claude Code terminal sessions counts toward the same plan limits, and once you exceed them, both interfaces draw from the same extra usage balance. Usage bundles bought in advance also work across Claude.ai, Claude Code, Claude Cowork, and third-party products.

Any app that authenticates with Anthropic's OAuth client ID for Claude Code (9d1c250a-e61b-44d9-88ed-5944d1962f5e) and requests the user:inference scope can obtain tokens that route inference through the same subscription billing path. Fazm, for example, runs the exact same OAuth flow as the Claude Code CLI: same client ID, same scope, same Keychain service name (Claude Code-credentials). When you log in, Fazm's queries consume extra usage from your plan just like a terminal session would.

Screenshot agents send full images to the model on every action. Image tokens cost roughly 1.15 tokens per 1.15K pixels for Claude, so a single 1440x900 screen capture can run 1,700 to 3,000 tokens per turn before the model even sees text. Accessibility-based agents read a structured text representation of the UI (every button, label, and text field as text with coordinates), which typically runs under 200 tokens for the same screen. Over a 20-step workflow the difference compounds into 10x to 30x cost deltas.

Run `security find-generic-password -s "Claude Code-credentials"` in Terminal. A single record exists whether you logged in through the claude CLI's /login command or through Fazm's Login with Claude button. The record stores the OAuth access token, refresh token, and expiration under a single JSON blob keyed `claudeAiOauth`.

Fazm's ACP bridge sets MAX_IMAGE_TURNS = 20 per session (acp-bridge/src/index.ts:676). After 20 turns that include screenshots, the bridge stops sending images on subsequent turns to avoid tripping Claude's per-message image limit and to keep the token bill bounded. Accessibility-tree context is still sent, so the agent continues to work from structured text.

The token exchange fails with a 403 or a rate limit error, and Fazm surfaces the error back to the chat. You can switch Fazm to its bundled built-in key (Fazm's own credits) or enable auto-reload on extra usage in your Claude account settings. The two modes are selectable in Settings and do not share billing pools, so you can fall back to Fazm's built-in key without touching your plan balance.

Yes. Fazm reads total_cost_usd from every Claude session result and computes a per-turn delta (patched-acp-entry.mjs:45-47: _sessionCostUsd = item.value.total_cost_usd; _lastCostUsd = item.value.total_cost_usd - prevSessionCost). That value is logged per query and rolled up into session cost analytics, so you can see exactly which automation step was expensive.