Recompute Your AI Cost Line After a Model Price Cut

The price cut starts the audit

When a model maker slashes the price of the small model you run in production, the tempting move is to switch the endpoint and call it a win. The better move is to open the cost sheet.

Anthropic launched Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for the under-100,000-token bucket, a 90% reduction from Haiku 4.5; that bucket covered roughly 90% of requests to the previous Haiku model.

Anthropic estimated average workloads would run about 75% cheaper than Haiku 4.5, with the figure shifting by request size and token use.

AWS's blog pointed to a similar task-level result. The model is available through Amazon Bedrock and Claude Platform on AWS.

Do not paste that average into your own P&L. The discount is smaller for long prompts, so the headline number can miss your mix.

Anthropic also cut Claude Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per million tokens and expects agent workloads to come in about 20% cheaper, but the real figure tracks how often your app reuses cached context.

A price cut is useful only when it changes the cost of a task you actually run. The audit turns the headline into a number you can defend in the P&L.

Before shifting volume, check that the smaller model can finish a narrow job consistently enough to support repeated use.

The worksheet beats the headline number

The formula is: input tokens times input price, plus output tokens times output price, plus cache or batch adjustments, divided by tasks. The hard part is getting the token counts from your own logs, not from a marketing page.

Your P&L needs a completed-task price. A lower API bill helps cash, but only a lower per-task cost that still preserves quality helps margin.

Keep the worksheet next to the P&L. Each row is a task, and each column is a number you can check against your logs. If a task has no stable definition, split it until it does.

• Count input and output tokens for each task in a recent sample.
• Split tasks into prompts at or below 100,000 tokens and prompts above that line.
• Price the lower bucket at $0.10 per million input tokens and $0.50 per million output tokens.
• Price the long-prompt bucket at $0.50 per million input tokens and $2.50 per million output tokens, a 50% reduction from Haiku 4.5.
• Add cache-read cost where you reuse stored context, using the new $0.10 per million rate for Claude Sonnet 5.5.
• Multiply by retries and expected accuracy, because a cheaper call that fails repeatedly can cost more.
• Divide total cost by completed tasks to get the per-task number.

The worksheet forces the question that headlines skip: what does a completed task cost now, not what does a token cost?

The go/no-go test protects margin

Use the per-task number to decide. Shift high-volume work when the cheaper tier lowers cost after retries and accuracy and the output still passes your quality bar.

Keep the task on its current model when it needs more tokens, more retries, or more human review.

Do not move work just because the unit price fell. A lower rate can still leave you with a worse margin if the task becomes harder to finish cleanly.

For a small team, the first candidate is often a support queue, a triage step, or a draft reply. A smaller model that handles that slice well can reduce support automation costs without touching the core product.

If it does not, the cheaper rate is irrelevant. You are paying for a task that still needs a human to finish, and the human time is the real burn.

Slower migration is fine. Move the narrow task first, then recheck the per-task number before expanding.