Claude Sonnet 4.6
The previous Sonnet flagship — still active, 50% more expensive than Sonnet 5, and the last Sonnet to use the pre-4.7 tokenizer.
Last verified August 21, 2026 · 2 sources
Claude Sonnet 4.6 is legacy, and unusually easy to argue yourself off. Claude Sonnet 5 is a third cheaper on both input and output, has the same 1M-token context window and the same 128K max output, and Anthropic describes it as a substantial improvement across reasoning, tool use, coding and knowledge work.
The one genuine reason to stay is the tokenizer. Sonnet 4.6 uses the pre-4.7 tokenizer; every Claude 5 model uses the newer one, which produces roughly 30% more tokens for the same text. If you have hard-coded token budgets, context-packing logic or per-request cost ceilings calibrated against the old tokenizer, moving to Sonnet 5 changes their arithmetic even though the nominal context window is unchanged.
It is also the last Sonnet where thinking.type: "enabled" appears at all — listed as
“Yes (deprecated)”.
Specifications
| Property | Value |
|---|---|
| Claude API ID | claude-sonnet-4-6 |
| Context window | 1M tokens (~750K words on the older tokenizer) |
| Max output | 128K tokens |
| Adaptive thinking | Yes |
| Extended thinking parameter | Yes (deprecated) |
| Reliable knowledge cutoff | August 2025 |
| Pricing | $3 in / $15 out per million tokens |
Lifecycle
Active but legacy. Tentative retirement not sooner than 17 February 2027.
It is also the named replacement for a long line of retired models: Claude Sonnet 4, Claude Sonnet 3.7, both Claude Sonnet 3.5 snapshots and Claude Sonnet 3.
Sources
- 01Models overview — Claude Platform DocsOfficialplatform.claude.com
- 02Pricing — Claude Platform DocsOfficialplatform.claude.com