Core API capabilities
Messages APIConversational completions with system prompts, streaming, temperature control, and stop sequences.
Tool UseDefine tools as JSON schemas. Claude decides when to call them. Auto, any, and specific tool modes.
VisionSend images and PDFs directly. Analyze charts, read documents, describe photos, extract data.
StreamingServer sent events for real time token by token output. Essential for responsive UIs.
Web SearchServer side tool. Claude searches the web and synthesizes results. $10 per 1,000 searches.
CitationsSource citations referencing specific passages in your documents for verifiable responses.
| Opus 4.6 | Sonnet 4.6 | Haiku 4.5 | |
|---|---|---|---|
| Input per MTok | $5 | $3 | $1 |
| Output per MTok | $25 | $15 | $5 |
| Context window | 1M tokens (GA) | 1M tokens (GA) | 200K |
| Max output | 128K tokens | 64K tokens | 64K tokens |
| Extended thinking | |||
| Adaptive thinking | |||
| Fast mode | Yes (2.5x faster, 6x pricing) | ||
| Relative speed | Moderate | Fast | Fastest |
Choosing the right model
Start with Sonnet 4.6 for most tasks. Upgrade to Opus 4.6 when you need the deepest reasoning, best coding output, or most reliable agent behavior. Use Haiku 4.5 for high volume, latency sensitive tasks like classification, routing, and summarization where speed and cost matter more than peak intelligence.