Every feature below is a capability of the Unified API — you opt in per request with a parameter or header, no configuration step required.
| Feature | What it does | Opt in with |
|---|---|---|
| Prompt caching | Reuse a stable prompt prefix across requests at up to 10× lower cost (Anthropic routes) | cache_control on a message |
| Vision (image input) | Send images to vision-capable models alongside text | content parts in messages |
| Tool calling | Let the model call functions your app defines and executes | tools |
| Custom metadata | Attach key-value pairs to a request, persisted with the thread and echoed back | metadata |
| Message compression | Summarise older turns in long threads to reduce token usage | "compression": "true" in metadata |
| Thread compaction | Fold older turns into a rolling summary checkpoint so long-thread cost stays flat | context_management |
| Skill instructions | Inject caller-composed skill prompts without replacing the agent’s identity | skill_instructions |
Features compose freely — a single request can cache its system prompt, send an image, define tools, and carry metadata at the same time.