Anthropic API Billing and Extra Credits for Third-Party Tools

If you use Claude through Cursor, Windsurf, Soulforge, or any other third-party coding tool, you have probably noticed that your billing works differently than chatting on claude.ai. Anthropic separates "plan usage" from "extra usage," and third-party tools pull from a different pool than you might expect. Here is how it actually works, what each tool draws from, and how to avoid surprise charges.

How Anthropic Splits Billing Into Two Pools

Anthropic runs two separate credit pools for every Pro and Team subscriber:

Plan usage covers direct interactions on claude.ai, the Claude mobile app, and first-party surfaces. This is what your $20/month Pro subscription pays for.

Extra usage credits cover everything else. When a third-party tool authenticates through your Anthropic account via OAuth, every request it sends draws from this extra usage balance instead of your plan allocation.

This split happened because third-party tools tend to send far more API calls than a human typing in a chat window. A single Cursor autocomplete session can generate dozens of requests per minute. Anthropic needed a way to meter that separately so heavy tool usage would not burn through plan limits that were designed for conversational use.

Which Tools Draw From Extra Credits

The billing pool a tool draws from depends entirely on how it authenticates with Anthropic. Tools using OAuth (signing in with your Anthropic account) hit your extra usage credits. Tools using their own API key or a BYOK setup may bill differently.

| Tool | Auth Method | Billing Pool | Notes | |---|---|---|---| | claude.ai (web) | Direct login | Plan usage | Your $20/mo subscription | | Claude mobile app | Direct login | Plan usage | Same as web | | Cursor | OAuth via Anthropic | Extra credits | Autocomplete and chat both draw | | Windsurf | OAuth via Anthropic | Extra credits | All Claude requests count | | Soulforge | OAuth via Anthropic | Extra credits | Agentic workflows consume more tokens | | Claude Code (CLI) | OAuth via Anthropic | Extra credits | Heavy usage during long sessions | | Cline | OAuth via Anthropic | Extra credits | VS Code extension | | Continue | OAuth or BYOK | Extra credits (OAuth) or your API key | Depends on config | | Custom API apps | API key | Direct API billing | Billed per token to your API account |

How Extra Credits Get Consumed

Token pricing for extra credits follows the same rates as the Anthropic API. The cost depends on which model the tool requests and how many input/output tokens each request uses.

| Model | Input (per 1M tokens) | Output (per 1M tokens) | Typical cost per request | |---|---|---|---| | Claude Sonnet 4.6 | $3.00 | $15.00 | $0.01 to $0.05 | | Claude Opus 4.6 | $15.00 | $75.00 | $0.05 to $0.30 | | Claude Haiku 4.5 | $0.80 | $4.00 | $0.001 to $0.01 |

The "typical cost per request" varies wildly based on context length. A Cursor autocomplete suggestion with 500 tokens of context costs a fraction of a cent. A Soulforge agentic workflow that feeds an entire codebase into context and generates multi-file edits can cost $0.50 or more per turn.

Real-world spend examples

Here is what we have seen across different usage patterns:

• Light Cursor usage (autocomplete, occasional chat): $5 to $15/month in extra credits
• Heavy Cursor with Opus: $30 to $80/month, especially with large codebases
• Windsurf daily coding: $10 to $40/month depending on model selection
• Soulforge agentic sessions: $20 to $100+/month since agentic loops run multiple turns autonomously
• Claude Code for a full project: $50 to $200/month during active development sprints

Checking Your Balance and Spend

You can see your current extra usage balance and spending history at console.anthropic.com under Settings > Billing.

To check which tools are consuming the most credits:

1. Go to your Anthropic dashboard
2. Navigate to Usage
3. Filter by "OAuth applications" to see per-tool breakdowns
4. Look at the daily and hourly graphs to find spikes

If you see unexpected spikes, check whether a tool left a background process running. Cursor's indexing, Windsurf's workspace analysis, and Soulforge's autonomous agents can all generate requests when you are not actively coding.

BYOK vs OAuth: Two Different Billing Paths

There are two ways third-party tools can access Claude, and they result in completely different billing:

If cost control matters to you, BYOK is almost always the better path. You can set a hard monthly spend limit on your API key, and the tool stops working when you hit it rather than silently running up charges.

Setting Spend Limits

Anthropic lets you set a monthly limit on extra usage credits:

1. Go to console.anthropic.com/settings/billing
2. Find "Extra usage limit"
3. Set a monthly cap (for example, $50)
4. When you hit the cap, all OAuth-connected tools stop until the next billing cycle

For API keys (BYOK), you can set separate limits:

# Check your current API usage via the Anthropic API
curl https://api.anthropic.com/v1/usage \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01"

Tool-Specific Billing Quirks

Cursor

Cursor sends requests for both autocomplete (Tab) and chat (Cmd+K / sidebar). Autocomplete uses shorter context windows and cheaper models by default. Chat uses the full context of your open files. If you select "Use Claude Opus" in Cursor's model picker, your per-request cost jumps roughly 5x compared to Sonnet.

Cursor also has its own Pro plan ($20/month) that includes some Claude requests bundled. If you are on Cursor Pro and using their bundled Claude, those requests do not hit your Anthropic extra credits at all. The Anthropic billing only applies when Cursor routes through your Anthropic OAuth.

Windsurf

Windsurf (formerly Codeium) uses Claude for its Cascade feature. Each Cascade step generates one or more Claude API calls. Multi-file edits with Cascade can send 5 to 10 requests in a single operation because it needs to understand each file's context separately.

Soulforge

Soulforge runs agentic coding workflows where the AI plans, executes, and iterates autonomously. A single task can generate 20 to 50+ API calls as the agent loops through planning, coding, testing, and fixing. This makes Soulforge the most credit-intensive of the three for complex tasks.

Common Pitfalls

• Leaving background processes running. Cursor indexes your workspace on startup and can send requests even when you are not typing. Windsurf's Cascade can get stuck in a retry loop. Close tools you are not using.
• Not realizing OAuth tools share one pool. If you use Cursor AND Windsurf AND Soulforge, they all draw from the same extra credits balance. One heavy session in Soulforge can drain credits you expected to have for Cursor.
• Ignoring model selection. Opus costs 5x more than Sonnet for input and 5x more for output. Use Sonnet as your default and only switch to Opus for tasks that genuinely need it.
• Confusing API billing with extra credits. If you also have an Anthropic API account for custom apps, that is a separate balance. API key usage and OAuth usage are metered independently.
• Assuming your Pro plan covers tools. The $20/month Pro plan covers claude.ai and the mobile app only. Every third-party tool needs extra credits or its own API key.

Quick Setup Checklist

Here is how to get your billing under control across all your tools:

1. Log into console.anthropic.com and check your extra usage balance
2. Set a monthly spend limit that matches your budget
3. For each tool, decide whether to use OAuth or BYOK:
   • Cursor: supports both. Switch to BYOK if you want per-tool cost tracking
   • Windsurf: supports both. BYOK gives you API-level usage logs
   • Soulforge: check current auth options in settings
   • Claude Code: supports both OAuth and API key via ANTHROPIC_API_KEY env var
4. If using BYOK, create a separate API key per tool so you can see exactly where spend comes from
5. Monitor usage weekly for the first month to calibrate your limits

# Set up BYOK for Claude Code
export ANTHROPIC_API_KEY="sk-ant-your-key-here"

# Verify it works
claude --version

Wrapping Up

Anthropic's billing split puts third-party tools on a separate meter from your chat usage. Whether you use Cursor, Windsurf, Soulforge, or all three, knowing which pool each tool draws from lets you budget accurately and avoid the "request rejected" wall mid-session. Use BYOK where possible for granular control, set spend limits, and check your dashboard weekly until you know your baseline.

Fazm is an open source macOS AI agent. Open source on GitHub.