Anthropic says the $2-per-million-input-token and $10-per-million-output-token price for Claude Sonnet 5 is now permanent rather than introductory. AWS, Google Cloud and Microsoft Foundry documentation show the model is broadly available through enterprise cloud channels, but teams still need to measure total workflow cost.

A launch discount is nice. A permanent price is more useful.

Anthropic says Claude Sonnet 5’s introductory API pricing is now permanent: $2 per million input tokens and $10 per million output tokens. The August 10 update removed a previously planned increase to $3 input and $15 output tokens.

That matters because teams can now estimate a pilot without wondering whether its economics disappear in a month.

Sonnet 5 is also available through multiple enterprise channels. AWS lists it as an active Bedrock model with a one-million-token context window. Google Cloud lists it as generally available, with text, image and PDF inputs and text output. Microsoft Foundry lists Sonnet 5 as generally available both on Azure-hosted infrastructure and Anthropic-hosted infrastructure.

In plain English: a company does not have to choose between trying Claude only in a consumer chat window or building a completely separate vendor relationship. It may be able to test the model through a cloud environment it already manages.

But “$2 in and $10 out” is not the number that determines whether a workflow is affordable.

The number that matters is cost per verified completed task.

Suppose a model creates a first draft of a report cheaply, but an employee spends 20 minutes correcting missing facts, poor structure and unsupported claims. The real cost includes the model, the human review and any downstream errors. A more expensive model can be cheaper if it reliably produces usable work. A low token rate can become costly if a tool repeatedly retries, uses a million-token context unnecessarily or forces reviewers to rebuild outputs.

The practical first test is straightforward. Pick one contained job, such as drafting a research summary from approved documents or extracting fields from a fixed form. Track four things: model cost, human review time, accuracy against known examples and the failure types. Do that across enough examples to see a pattern, not just one lucky result.

The one-million-token context window is powerful, but it is not an instruction to load every document the company owns. More context can mean more cost, more sensitive data exposure and more places for irrelevant material to distract the model. Start with the smallest relevant data set.

There is also a control point worth remembering. Microsoft’s documentation says Foundry does not provide built-in content filtering for Claude models at deployment time. That does not mean the model has no safeguards. It means a deployment team must understand the controls around its own application rather than assuming the cloud marketplace did all the safety work.

What to watch: cloud-specific pricing, quotas, regional availability and tool support. The permanent Sonnet 5 price makes pilot planning easier. It does not eliminate the need to measure real-world quality before scaling.

Bottom Line

Permanent Sonnet 5 pricing makes pilots easier to model, but teams should still calculate total cost per verified completed task, including context, reasoning, retries, and human review.

Sources