Model comparison · 2026-09-22
Claude Sonnet 4.6: strong on existing systems, Sonnet 5 for new work
Covers that anthropic/claude-sonnet-4.6 is the previous Sonnet flagship, that its catalog description says it remains strong for existing systems but recommends Sonnet 5 for new integrations, and what that distinction means in practice.
The description boils down to two words: 'existing' and 'new'
anthropic/claude-sonnet-4.6's catalog description makes this distinction directly: it's the previous Sonnet flagship with balanced quality and cost across coding, agents, computer use, and professional workflows, remaining strong for existing systems, but Sonnet 5 is the more current choice for new integrations. This resembles the 'older generation still in the catalog' pattern seen in the Grok 4.5/4.6 pair, but here the difference is not only recency: the newer model is cheaper. Sonnet 4.6 is priced at $3 input / $15 output per million tokens, Sonnet 5 at $2 / $10, so Sonnet 5 is about one third cheaper on both input and output.
The pricing card ($3/$15 per million tokens) and capability set (function calling, image input, document input, 1M context) are solid and sufficient for critical production work; this model isn't 'weak' or 'lacking,' it's just not the family's most current member.
What this distinction translates into as a concrete decision
If you already have a production system needing coding, agent flows, or computer use, already tested against Claude Sonnet 4.6 with predictable behavior, staying on this identifier is a reasonable choice; the system works, its behavior is known, and moving without a functional gain carries an unnecessary regression risk.
When deciding to stay on the existing system, include the price difference: the same traffic costs about one third less on Sonnet 5. If your monthly token volume is large, that difference can quickly pay for a migration test; if volume is small, keeping tested behavior may weigh more. If you're building an integration from scratch, starting directly with Sonnet 5 makes more sense, since it is both more current and cheaper.
- An existing, tested production system: staying on 4.6 is reasonable — don't move without a functional need.
- A new integration: the catalog description directly recommends Sonnet 5.
- Price difference: 4.6 at $3/$15, Sonnet 5 at $2/$10 (per million tokens); factor it into the migration decision with your monthly volume.
Function calling and document input are consistent here too
Claude Sonnet 4.6 supports function calling, image input, and document input; that provides a consistent base interface with other models in the Anthropic family (Opus 5, Fable 5, Haiku 4.5). Most of your integration code doesn't change when moving from one model in the family to another.
Frequently asked questions
Is Claude Sonnet 4.6 much cheaper than Sonnet 5?
No, the opposite: Sonnet 4.6 is priced at $3 input / $15 output per million tokens, Sonnet 5 at $2 / $10. The newer model is about one third cheaper on both input and output; the reason to stay on 4.6 is the existing system's tested behavior, not price.
Is there a removal plan for Sonnet 4.6?
The model page is currently open and stays isPublic; this repo follows a retirement/migration policy rather than removal for models — follow the current state on the model page.