Anthropic's Sonnet 5 rate holds at $2 in and $10 out per million

The workload

A founder embedding Claude in a product faces the same task as with any model vendor: matching the tier named on the pricing page to the model actually called in code, because Anthropic prices several current tiers at once rather than a single flagship rate. Verified: Anthropic's pricing page, retrieved 16 September 2026, lists Claude Opus 5 at $5 per million input tokens and $25 per million output tokens, Claude Sonnet 5 at $2 per million input and $10 per million output, and Claude Haiku 4.5 at $1 per million input and $5 per million output.

What the documents show

Verified: Anthropic's pricing documentation, also retrieved 16 September 2026, confirms the same three rates and states the caching mechanism: a 5-minute prompt-cache write costs 1.25 times the base input price, a 1-hour write costs 2 times the base input price, and a cache read costs a tenth of the base input price, for example $0.20 per million tokens on Sonnet 5 against its $2 base rate. The documentation states these multipliers stack with other modifiers, including a separately confirmed 50% discount on both input and output tokens for asynchronous batch processing. The two documents agree on every figure checked, which is expected since the pricing page and the documentation are meant to describe the same live rate card rather than independent estimates.

The operating cost

Verified: on Sonnet 5, a request of 1 million input tokens and 1 million output tokens costs $2.00 plus $10.00, or $12.00, at standard rates. This is an estimated figure applying the published rate to a round token count; a cache hit on that same input would cost $0.20 instead of $2.00, and routing the same request through the Batch API would cost $1.00 for input and $5.00 for output, or $6.00 total, per the documentation's batch table. Actual cost for a given product depends on cache-hit rate and whether latency requirements permit batch processing, neither of which the pricing page can determine on its own.

The stop condition

This is editorial: the documentation notes that Sonnet 5's launch pricing, originally introductory through 31 August 2026, was made the standing rate rather than increasing as originally scheduled, showing that list prices for a given tier can change even without a new model release. A cost model built on today's rate should be re-checked whenever Anthropic issues a pricing note like that one, not treated as fixed for the product's lifetime.

• Which named model tier is the product actually calling, and does the founder's cost model cite that tier's rate specifically?
• Would caching or batch processing fit the product's latency requirements enough to capture the discounted rates?
• Has the founder checked for a pricing note superseding the currently listed rate, the way Sonnet 5's introductory rate was made standard?

Anthropic's two documents agree closely enough to be read as one source in practice, but the agreement is worth confirming rather than assuming, since a living pricing page can diverge from its own documentation after either is updated.