claude-sonnet-4-6 — pricing, context window and limits

The previous-generation Sonnet, now the pricier one per token

Sonnet 4.6 costs more per token than both Sonnet 5 and Sonnet 5.5 — the newer Sonnets launched cheaper, and Sonnet 5's launch price became permanent rather than rising. That makes this the one Sonnet migration where the sticker price points clearly toward moving. It still isn't the whole answer, because this model keeps the older tokenizer: identical text produces fewer tokens here, so the newer models' per-token saving is partly spent on counting more tokens. Measure a real request on both with count_tokens before quoting a saving.

Thinking here is opt-in, not default

This model only thinks when you ask it to. Both newer Sonnets think by default, and on Sonnet 5.5 you can no longer switch thinking off with disabled at all. Code that treats these generations as interchangeable — a router, a fallback, a shared request builder — has to branch on that difference rather than assume one default across all three.

Still a reasonable choice for a tuned workload

For a workload already tuned and performing well here, there's no deadline forcing a move: the model remains active, and its earliest possible retirement is shown in the lifecycle table above. A deliberate migration tested against real traffic beats a rushed one.

Related

See migrating: Sonnet 4.6 to Sonnet 5.5 for the direct move, and what actually changed from Sonnet 4.6 to Sonnet 5 for the tokenizer and thinking-default changes in practical terms.