Claude API for Education

AI Coding with Claude Code

Claude Code is Anthropic's AI-powered software engineering agent designed to translate natural language into code, capable of reading, writing, and executing code across languages like Python, JavaScript, and Go. It serves as an intelligent coding partner to accelerate development, debug, and assist in managing software projects via the terminal, desktop, or IDE. Claude Code is currently available only with an API key.

To learn more about Claude Code, click here.

Learn More

To deepen your understanding of the Claude API and enhance your skills through comprehensive training designed for all levels—from beginner to advanced—explore Anthropic's documentation and learning resources. These offer expertly curated guides, hands-on activities, and resources directly from Anthropic, empowering you to leverage the Claude API effectively and confidently.

Data Privacy

Claude API for Education is approved for use with certain classes of Sensitive Data, but is NOT approved for use with Protected Health Information (PHI) as defined in the Columbia University Data Classification Policy. For users at CUIMC, please refer to the AI and Generative Technology Use at CUIMC page for additional information regarding AI tools approved for use with Sensitive Data.

Per CUIT's AI Data Classification Table (as of 5/2026)

Updates

View the latest Model and Feature Updates here.

Pricing:

We use the same pricing guide as Anthropic in regards to tokens. For more information about Anthropic's pricing for the API, check out their site here.

• For more details about pricing per model, you can view that information here.

When requesting a Claude API key, you can set your budget amount, anywhere from $5 to $5,000. A departmental chart string is required when requesting an API key. If you are unsure where to find your chart string, reach out to your department or Finance team.

Claude API Token Best Practices:

• Set a hard output cap every time. The Claude API requires max_tokens on every request, so use it to cap output length rather than relying on a model default.
• Preflight big prompts. Count input tokens before you call the model using the token counting endpoint (client.messages.count_tokens). If the input is too large, chunk or summarize first. Note that the returned count is an estimate and the actual tokens used may differ by a small amount.
• Log usage per request. Read the usage fields (input_tokens and output_tokens) on each response and alert when a request crosses your threshold.

Python Example

from anthropic import Anthropic
client = Anthropic()

message = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=200,
    messages=[
        {"role": "user", "content": "Summarize this in 5 bullets: ..."}
    ],
)

print(message.usage)  # input_tokens / output_tokens / total_tokens

Office Hours

Office hours will be taking a break for the summer. Office hours will resume in September of 2026.

AI Consultation

We provide consultations to understand your needs and ensure our AI services align with your requirements. To discuss how our AI services can support your specific use cases and workflows, please request a consultation at [email protected].

Please do not use this email for general questions pertaining to our AI Services. For those questions, please email us at [email protected].

Guidelines for Using Anthropic's API at Columbia University

Anthropic's API platform supports generative AI features inside Columbia tools and workflows. You can use it to build apps, automate tasks, and run AI services at scale.

These guidelines cover best practices and key responsibilities. Also follow Columbia’s Generative AI Guidelines and all university privacy, security, and intellectual property policies.

1. Be Careful Who You Share API Keys With

   Sharing API keys should be treated with caution, and that caution should also be communicated to those you share keys with. Risks of unauthorized use include budget theft, data insecurity, and lack of accountability.

    

2. Use the API responsibly.

• Validate outputs before you act on them. Example: Don't auto-send emails, publish web content, or change records based only on model output.
• Keep a human in the loop for high-impact work. Example: grading, HR decisions, disciplinary actions, clinical workflows, and official communications.
• Design for reliability
  • Log prompts, responses, and errors in a secure way
  • Add guardrails like input validation, allowlists, and rate limits
  • Use test cases that reflect real scenarios in your unit
• Respect intellectual property
  • Don't paste copyrighted or licensed content unless you have rights to do so
  • Don't assume model output is free of IP risk
     

2. Data and Privacy

Handling Sensitive Data

Columbia's Claude API access is covered under the same Education program controls as Claude for Education.

If your work involves PII or research-sensitive data:

• Follow the minimum necessary principle. Only send what the task needs.
• De-identify whenever you can. Remove direct identifiers like names, MRNs, addresses, phone numbers, SSNs.
• Follow Columbia's Data Classification Policy. Match your use to the right data class and required controls.
• Keep data flows explicit
  • Know what data your app sends to the API
  • Know what you store after the response returns
  • Set retention rules for logs and transcripts

CUIMC Users

The Claude API is NOT approved for PHI. CUIMC members are by default restricted to AI models approved for PHI Data (currently OpenAI models within CHAT). Before using any AI tool for PHI or other sensitive clinical data, consult CUIMC IT at [email protected]

• Use your [email protected] email address for account and service updates tied to the Education program.
• If you do not receive emails to a [email protected] email, you can request Dual Affiliation here
   

Security Controls You Must Apply in Your App

The API is only part of the system. Your application also needs controls.

• Secrets management
  • Store API keys in a secrets manager or environment variables
  • Never hard code keys in code, scripts, or shared docs
  • Never commit keys to GitHub or shared repositories
• Access controls
  • Restrict who can use your internal tool
  • Use least-privilege permissions for systems that call the API
• Logging hygiene
  • Don't log sensitive data by default
  • If logging is required, redact identifiers and secure the logs
• Data retention
  • Set clear retention windows for stored prompts, outputs, and telemetry
  • Delete data you no longer need
• Disclose AI assistance when required by course, journal, or unit policy
  • Example: "Drafted with AI assistance and reviewed by the author."

If you build tools for students or faculty:

• Make attribution easy
  • Example: include an exportable "AI assistance" note or citation snippet.
     
     

3. Ethical Use and Limitations

You are responsible for outcomes.

• Expect errors and bias. Test for failure modes relevant to your use case.
• Avoid automated decisions about people. Example: admissions screening, hiring rankings, performance reviews, patient triage. If you must support these areas, keep the API role advisory and require human review.
• Use trusted sources for facts. Example: when summarizing policy, link back to the official policy text.
   

Questions to ask before you ship

• What’s the worst realistic harm if the model is wrong?
• What data leaves Columbia systems, and what stays?
• Who can access logs, outputs, and stored prompts?
• How will a user report issues or appeal a result?