SDK (Software Development Kit) — AI Glossary

Reference details

Related terms

• OpenAI Python SDK
• Anthropic Python SDK
• API Key

How it works

1. A client wraps the provider's HTTP API in language-native types, so request construction and response access are checked at the call site.
2. Authentication, serialisation and header handling are managed, which removes the most common integration mistakes.
3. Streaming is parsed from the wire protocol into typed events, so callers consume objects rather than reassembling chunks.
4. Transient failures are retried with exponential backoff by default, which means one logical call may be several requests.
5. Because the client owns retries and timeouts, its configuration interacts with the caller's own — a timeout shorter than the backoff ladder cancels work the SDK would have completed.

Example: The configuration interaction that looks like flakiness

A client retries at 1s, 2s and 4s, so a call that fails twice before succeeding takes about 3 seconds of waiting plus three requests. Wrap it in a 2-second application timeout and it is cancelled mid-ladder, surfacing as a failure on a request the SDK was going to complete. The integration looks unreliable and the provider is fine. Reading the client's default retry and timeout values is the first diagnostic for intermittent failures that correlate with load.

Commonly confused with

Used loosely for anything that calls a model, which blurs three layers. An SDK is the provider's own typed client for one API. A gateway or multi-provider library normalises several providers behind one interface and adds routing and fallback. A framework goes further and structures the application — chains, agents, retrievers — with model calls as one part. They stack, and the usual mistake is reaching for a framework when an SDK would do, which adds abstraction over an API that was already simple.

When to use it

Reach for it when:

• Calling one provider directly with typed requests, streaming and structured output.
• Transport concerns — auth, retries, serialisation — should be handled rather than written.
• Minimal abstraction over the API is wanted.

Reach for something else when:

• Several providers need one interface with routing and fallback, which is a gateway.
• Application timeouts are set below the client's retry ladder, which cancels recoverable work.
• Assuming OpenAI-compatible endpoints support every parameter — compatibility is usually partial.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to SDK (Software Development Kit) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path