3 Things AI: The “Retest the Work, Not the Hype” Edition

1. Reprice one real task before changing your model

A service company already uses a capable model to turn messy project notes into a weekly client summary. The output is good, but the team limits use because every long run feels expensive. A cheaper new model sounds like an easy switch. The useful question is whether it lowers the cost of an accepted summary, not merely the price of a token.

Anthropic introduced Claude Opus 5.5 on September 22. The company says typical workloads cost 40% less than Opus 5 at default settings, with listed prices of US$4 per million input tokens and US$20 per million output tokens. It also reports output more than 30% faster and says the model performs at the level of Claude Fable 5.1 on most work.

The benefit is room to retry work that previously looked too costly or slow. The tradeoff is that lower unit prices do not guarantee a lower business cost. A model that needs more review, takes extra tool steps or produces unusable answers can erase the saving.

Pick one recurring task with a known good result. Replay 20 representative examples through the current model and the new option with the same instructions, tools and review rubric. Record total input and output usage, elapsed time, reviewer edits and how many results were accepted. Compare cost per accepted result rather than cost per call.

Anthropic’s performance and cost figures are vendor-reported and depend on its test settings. Your documents, connected tools and review standard may produce a different result, so this is a reason to test a route, not a reason to replace one.

2. Give one browser workflow a company-approved version

An office manager watches staff copy information between a CRM, a scheduling tool and a customer portal. People naturally try public AI helpers to reduce the repetition, but nobody can say which instructions are approved or what happens when sensitive text reaches the clipboard. Blocking every experiment would preserve the manual work without resolving the demand.

Google described new enterprise browser controls on September 23. In a trusted-tester program, IT teams can publish preconfigured, vetted AI Skills for managed users. Google also says Chrome Enterprise Premium can report AI and software use, redirect people to approved alternatives and will add rules that detect and block sensitive records at the copy trigger before they reach the clipboard.

The benefit is a shared way to distribute one repeatable workflow and put data rules close to where the work happens. The tradeoff is dependence on a managed browser, licensing and administration. A sanctioned skill can still encode a bad process, and a browser control does not govern work on every personal device or native application.

Choose one low-risk task, such as moving public event details into an internal draft. Write an approved skill with the permitted sites, fields and stopping point. Test it using invented records on a managed device, then try the same copy action with a field marked sensitive. Confirm what is blocked, what is logged and where a person must approve the result.

Several capabilities are in trusted testing, described as coming soon or limited by region and Workspace eligibility. Google’s announcement describes the intended controls; it does not show that they cover your device mix or satisfy your privacy obligations.

3. Test whether old media becomes useful, not merely sharper

A small manufacturer has years of product photos and demonstration clips, but many are too soft or low-resolution for a new catalogue. Reshooting everything would take time and money. AI enhancement may recover useful material, yet a polished image that changes a label, surface or product detail could create a different problem.

Adobe announced the completed acquisition of Topaz Labs on September 23. Adobe says Topaz technology for image and video enhancement is already available in Firefly and Photoshop, can run on a professional’s device or through a Topaz cloud option, and will reach more Creative Cloud workflows. Topaz will also remain available as a standalone brand.

The benefit is a chance to reuse valuable media without moving every file through a separate specialist workflow. The tradeoff is that enhancement can make an asset look more convincing without making it more accurate. Local processing may help with file handling, but it does not settle ownership, consent or truthfulness.

Select ten assets that have clear originals and modest business value. Define what may improve—noise, resolution or low-light clarity—and what must not change, including text, product geometry, colour and identifying features. Process copies, keep the originals untouched, and have the person responsible for the product compare both at full size before approving any public use.

Adobe’s announcement does not establish how well the tools will handle your specific archive, and some broader integrations are still described in future terms. A small visual review is more informative than assuming an acquisition has already solved the workflow.

The bigger pattern

A cheaper model, an approved browser skill and an enhancement tool all make an existing task easier to revisit. None removes the need to define what a good result looks like. The balanced move is to make the comparison small enough to measure, keep the original or current route available and let the observed result decide what changes next.

Which recurring task could your team retest this week with a clear before-and-after measure?