Claude Sonnet 5.5, Shopify checkout: Sept. 29 dispatch

Claude's latest update gives you a reason to retest routine writing costs, but changing the model name alone can break an automation. Shopify is extending AI shopping into checkout, while NVIDIA is adding controls around automated workers. Today's build list: test one replacement, check the purchase handoff, and separate instructions from permissions.

Claude Sonnet 5.5 changes the upgrade checklist

Anthropic released Sonnet 5.5 on September 28, keeping Sonnet 5's unit prices while reporting lower costs per task. Its launch announcement attributes savings to using fewer tokens, the small pieces of text AI systems read and produce. Treat the saving as a supplier claim until your own repeat tasks show it. Start with summaries and draft copy, where you can compare answers before customers see them.

The published rates below were checked September 29. They apply to the API, the software connection that lets another app request an AI answer, rather than a chat subscription.

Charge per million tokens	Sonnet 5	Sonnet 5.5
Text sent in	$2	$2
Text generated	$10	$10

The part that breaks is an old setting carried into the new connection. Anthropic's migration instructions say thinking.type: "disabled" now causes an error. The replacement, between_tools, turns off reasoning before the first answer but permits progress updates between actions. Forced tool selection also stops working: an automation cannot carry over the old instruction that makes Claude call one particular connected function. If that function writes a customer record, an apparently successful answer without the action is a business failure.

Keep the review record from the content approval build. If you followed last week's Opus 5.5 dispatch, this is a separate test for routine jobs. A cheaper draft loses its advantage when someone spends longer repairing it.

Shopify lets browser agents finish checkout

Shopify's September 28 checkout expansion lets an AI shopping assistant proceed beyond filling a cart on eligible stores. The checkout documentation describes WebMCP, a way for a browser to expose named actions that an assistant can use directly. The customer stays in the same browser tab and must authorize the order. Store owners should test that handoff before commissioning a custom shopping assistant.

The important wiring change is where the assistant looks for actions. When it moves from your storefront into checkout, it must refresh the available action list. The checkout provides get_checkout to read the order, update_checkout to change details, and complete_checkout to submit it. Updating details requires sending the complete desired checkout state; leaving fields out can clear them. Hand that requirement to your developer before testing address changes or discount codes.

Use an eligible test checkout and ask your implementer to demonstrate this sequence: read the order, change delivery details, read it again, show the current total, and request buyer approval. Only then submit. If the total changes, approval must be requested again. Only a returned status of completed confirms the order. When checkout actions are unavailable, hand the page back to the buyer instead of trying to click around the restriction. Your acceptance test is a correct, authorized purchase, not a convincing conversation.

NVIDIA puts controls outside the AI worker

NVIDIA announced its Open Agent Safety Platform on September 28. The release combines OpenShell, software that limits what an AI worker can access while running, with Sentry, a hardware monitoring design that operates separately. An agent is an AI system allowed to take actions across software. This matters when yours can reach customer files or change business records.

If you only use a chat window to draft copy, leave infrastructure purchasing alone. If you run custom agents, ask your implementer to demonstrate an enforced boundary: give the test worker access to one sample folder, then ask it to read another. The second request should be blocked by the surrounding system, regardless of what the AI says. Record the blocked action before expanding access. Use the guarded lead-qualification build to define which business actions require review.

On the bench

• Claude Haiku 5.5: Anthropic says it is coming in the next few weeks. Save routine classification examples for a later cost comparison.
• Slack approvals: NVIDIA says its OpenShell integration supports permission requests in Slack. Test who receives a request and whether rejection actually blocks the action.

Get the next build from the automation runbooks.