Costs & Subscriptions
What the tools cost, and how to use them efficiently
Claude Code (Anthropic)
Claude Code is included with all paid Anthropic subscriptions. There is no separate charge for using it.
| Plan | Price | Context | Best for |
|---|---|---|---|
| Pro | $20/month | 200K tokens | Getting started, occasional use |
| Max | $100/month | up to 1M tokens | Regular daily use, longer sessions |
| Max 20x | $200/month | up to 1M tokens | Heavy use, parallel sessions, teams |
The plans differ primarily in usage limits — how many messages and tokens you can consume per day or per hour before hitting a rate limit. Pro is deliberately limited to keep costs manageable; Max is designed for sustained professional use.
All plans use Claude Sonnet as the default model. Pro users occasionally get Claude Opus access; Max users get more consistent access to the most capable models.
Existing Claude.ai subscribers can use Claude Code immediately — just install it via npm install -g @anthropic-ai/claude-code or follow the setup guide. Your existing subscription already covers it.
What a “Message” Costs in Practice
Subscription plans don’t charge per token — they operate on a usage limit measured in “messages” per day or hour. What counts as heavy usage?
A rough guide based on typical research workflows:
| Activity | Approximate messages |
|---|---|
| Short prompt + response | 1 message |
| Asking Claude to write a full R script | 1–3 messages |
| Debugging a regression until it works | 5–15 messages |
| One focused 1-hour coding session | 20–40 messages |
| A full day of active use | 50–150 messages |
Pro hits its limits at roughly 50–100 heavy messages per day. If you’re doing exploratory work — one session in the morning, not all day — Pro is usually fine. If you’re using Claude Code as your primary programming interface throughout the working day, you’ll hit Pro limits within a few hours and need Max.
Usage limits reset daily (and partially hourly). If you hit a limit, waiting an hour often restores some capacity. The limits are designed to prevent runaway usage, not to frustrate normal work.
How Quickly You Burn Tokens
Even though subscription plans don’t bill per token, understanding token consumption helps you understand why you hit limits faster on some days.
A rough calibration:
- One focused 1-hour session of active coding: roughly 50,000–100,000 tokens
- A session where you read a large CSV: add 50,000–100,000 tokens instantly
- A session with many iterations on a long script: 100,000–200,000 tokens
What burns tokens fast: - Reading large files (data previews, long scripts) - Many short back-and-forth exchanges (each turn carries full history) - Asking Claude to write long documents or comprehensive reports - Sessions that stay open for hours without /compact
What burns tokens slowly: - Short, targeted questions - Using /compact regularly - Fresh sessions for new tasks - Asking Claude to “look at lines 40–60” instead of “read the whole script”
The context window guide covers strategies for using tokens efficiently.
When Pro Is Enough vs. When to Upgrade
Pro is enough if you:
- Are trying Claude Code for the first time
- Use it for occasional tasks — a few times a week
- Work in short, focused sessions
- Are primarily using it for writing and editing, not intensive coding
Consider upgrading to Max if you:
- Use Claude Code daily as part of your research workflow
- Run sessions longer than 1 hour regularly
- Work on large codebases with many files
- Find yourself hitting rate limits and waiting
- Are using Claude Code for a time-sensitive project (paper deadline, conference)
Max 20x is for:
- Teams sharing an account
- Parallel sessions (running multiple Claude Code instances simultaneously)
- Research groups with centralized subscription management
Start with Pro. You’ll know within a week whether you need Max. The upgrade is immediate — no new account needed.
Codex (OpenAI)
OpenAI’s Codex is a competing tool, accessible through ChatGPT and the API. For comparison:
| Plan | Price | Notes |
|---|---|---|
| ChatGPT Free | $0 | Limited model access, no file system integration |
| ChatGPT Plus | $20/month | GPT-4o access, limited coding tools |
| ChatGPT Pro | $200/month | Extended limits, o1 access |
| API pay-per-use | Varies | Codex-specific pricing; can be more expensive than subscription for heavy use |
Codex as a terminal-based coding assistant comparable to Claude Code is a newer offering from OpenAI. As of April 2026 it requires an API key and is typically billed per token at API rates, which can exceed subscription costs for heavy use. The workshop covers both tools — see the exercises for a direct comparison.
Other Tools in the Ecosystem
| Tool | Price | What it does | Best for |
|---|---|---|---|
| Cursor | Free / $20/month | VS Code fork with AI built in — write and edit code in a full IDE | Those who prefer a traditional code editor with AI assist |
| GitHub Copilot | $10/month | In-editor code completion (autocomplete) | Developers in VS Code or JetBrains who want real-time suggestions |
| ChatGPT Plus | $20/month | Browser-based chat — no direct file system access | Writing, brainstorming, explaining concepts |
| Gemini Advanced | $20/month | Google’s offering — strong at multimodal tasks | Document analysis, PDF reading |
How these compare to Claude Code:
Cursor and GitHub Copilot are primarily code editors with AI assistance — they help you write code interactively but don’t operate autonomously on your project the way Claude Code does. Claude Code can read files, run code, navigate a project, and handle multi-step tasks with minimal hand-holding.
ChatGPT Plus is useful for writing and explaining but has no access to your file system unless you use the Artifacts feature or upload files manually. For research workflows involving code and data, it’s less capable than Claude Code in practice.
The tools are not mutually exclusive. Some researchers use Claude Code for autonomous tasks and Cursor for line-by-line coding. The workshop focuses on Claude Code because it’s the most capable for the workflows economists care about.
For a running account of real-world usage patterns and costs, see Scott Cunningham’s Claude Code series — 36+ posts documenting daily use for empirical econometrics.
Institutional Options
Do not share subscription credentials. One account per person is required by Anthropic’s terms of service. Shared accounts can be suspended. If multiple people at your institute need access, they each need their own account.
Group subscriptions and volume licensing:
Anthropic offers team and enterprise plans for organizations. These allow centralized billing, admin controls, and (for enterprise) additional data processing agreements. If your department wants to provide access to multiple researchers, contact Anthropic directly through anthropic.com.
Anthropic for Research:
Anthropic maintains a research access program for academic projects. This typically provides API credits rather than subscription access — useful for large-scale computational work but different from Claude Code’s interactive use case. Check Anthropic’s website for current terms; the program changes.
Kiel Institute:
At time of writing, the Kiel Institute does not have a centralized Claude Code subscription. Individual researchers are responsible for their own accounts. Ask your group leader if this has changed. For expense reimbursement, see the FAQ below.
API access:
An alternative to subscriptions is direct API access, billed per token. For heavy users who run long automated tasks or batch jobs, API pricing can be more cost-effective. For interactive daily use, subscriptions are almost always cheaper. Claude Code works with API keys (set the ANTHROPIC_API_KEY environment variable), but you lose the subscription’s built-in usage management.
Cost-Saving Tips
The biggest factor in how quickly you burn through limits is context window management. See the context window guide for detailed strategies. The short version:
- Short, focused sessions use far fewer tokens than long meandering ones
/compactbetween tasks compresses accumulated history- Fresh sessions for new tasks give you a clean context budget
- Read only what you need — ask for column names, not the whole CSV
FAQ
Can I use my API key instead of a subscription?
Yes. Set ANTHROPIC_API_KEY in your environment and Claude Code will use API billing instead of your subscription. Check current per-token pricing at anthropic.com/pricing. One heavy day of use might cost $5–15 via API. For regular use, a subscription is almost always cheaper. For sporadic use, API might be more economical.
Is there a free tier?
Claude.ai has a free tier for browser-based use, but Claude Code (the terminal tool) requires a paid plan. There is no free tier for Claude Code as of April 2026.
How do I expense this?
A $20/month Claude Pro subscription is analogous to other professional software subscriptions. Most German research institutions can reimburse software needed for research work through your project budget or department budget. Keep the receipt from Anthropic (they email a PDF invoice monthly). Check with your Sachbearbeiter or institute administrator about the right cost center and whether pre-approval is needed.
Can two people share an account to split the cost?
No. Anthropic’s terms of service require one account per person. Beyond the terms issue, shared accounts create practical problems: usage limits are shared, conversation history is shared, and you lose the personal context that makes the tool useful.
What happens if I go over my limit mid-project?
You’ll see a rate limit message. Options: wait (limits partially reset hourly), /compact to reduce context and retry, or upgrade your plan. Claude Code will not lose your work — it just can’t generate more responses until the limit resets.
Pricing Reference
Current pricing (subject to change): anthropic.com/pricing
A Real-World Cost Stack
Chris Blattman (UChicago) publishes his actual monthly stack at ~$140/month. It’s the most honest and well-reasoned cost breakdown we’ve seen for an academic user:
| Tool | Purpose | Cost |
|---|---|---|
| Claude Max (Opus + Sonnet) | Primary AI + Claude Code | ~$100 |
| ChatGPT Plus | Deep Research, web search, images | ~$20 |
| Wispr Flow | System-wide voice dictation | ~$10 |
| Granola | Meeting transcription | ~$10 |
His ROI framing: at 5–15 hours saved per month and a notional $40–80/hour rate, the stack pays for itself many times over for senior researchers. For grad students without institutional support, he suggests starting much smaller — a single Claude Pro subscription (~$20/month) is enough to learn whether the tool fits your workflow.
→ See also: claudeblattman.com.