Nothing broke this week. That's the story.
Three separate systems that touch your catalog, your ad account or your creative pipeline changed their default behaviour between Monday and Friday, and in all three cases the change was designed to require nothing from you. No error. No approval screen. No email you'd have opened. If you did nothing at all, everything proceeded — which is precisely the version of a change that costs money, because there is no moment where you notice it happening.
I've written some version of that sentence about Amazon four times this year. This week is the first time the most important instance was a tool sitting on my own machine, where the fix is a settings file rather than a calendar reminder. Here are the five things worth your attention.
1. Claude Code's auto mode became the default on Thursday
Anthropic announced on August 7 that starting August 14, new Claude Code sessions on Pro, Max and Team plans run in auto mode. Instead of prompting for approval at each step, it proceeds unless an action is judged "irreversible, destructive, or aimed outside your environment." TechCrunch covered it on August 9.
The permission model didn't get weaker so much as it got inverted. It used to be that the human approved and the tool waited. Now the tool proceeds and a classifier decides which actions were serious enough to interrupt for. That is a real, well-engineered classifier — Anthropic describes hard deny rules that can't be approved even on request, git repository visibility checks before pushes, and git status checks before destructive operations. Anthropic also says plainly that for high-stakes changes to production infrastructure you should still be reviewing the actions yourself.
Here's the operator part, and it's the same split I wrote about in July when Opus 5 became a subscription default: auto mode is the new default on Pro, Max and Team, and remains opt-in on Enterprise, the API, Bedrock, Google's Agent Platform and Microsoft Foundry (with Anthropic signalling that will change within a month). If your automations call the API with an explicit model string and explicit permissions, Thursday changed nothing for you. If a VA, a contractor, a freelance developer or an agency does work for you inside a Claude Code subscription — running a flat-file transformation, a bulk copy pass, a catalog export cleanup — the approval step you assumed was sitting between them and your files stopped being the default on Thursday and nobody sent you a memo.
The move this week is not to panic and it is definitely not to write "always ask me before you change anything" into a prompt. A prompt is not a permission setting. Managed settings support defaultMode to pin an organisation-wide default and disableAutoMode to switch it off entirely. Deny rules are configuration; instructions are a request. If someone else's Claude Code session can touch a file that eventually becomes a listing, you want the first kind.
2. Sponsored Products started serving inside creator content on Monday
August 10 was the go-live date for Sponsored Products appearing inside Amazon Influencer Program content, as PPC Land reported ahead of the change. Campaigns were enrolled by default at existing bids and budgets. The advertiser notice said no action was required, which was true.
I covered the setup for Velocity Sellers ten days ago, so I'll keep this to the part that only becomes real now that it's live: your search term report is beginning to accumulate terms nobody typed. Some creator placements have no query behind them — a review page is not a search — so what shows up in the report is Amazon's inferred keyword. Every search-term workflow you run assumes a search term is something a person typed. That assumption is what makes harvest-and-negate work.
The practical consequence is that the clean pre-change baseline is now gone. If you didn't pull a placement report before Monday, you can't have one. Pull one this week anyway, write the date on it, and treat it as a during-change baseline rather than a before. And be careful with negatives for the next few weeks: a negative keyword added off an unfamiliar inferred term is easy to add, easy to forget you added, and quietly suppresses on-Amazon delivery for a phrase real humans do search.
3. Amazon DSP accounts upgraded themselves into a single global seat
Amazon Ads told advertisers on July 30 that existing DSP accounts would be automatically upgraded into a single advertiser account spanning 34 countries — no re-registration, no action required, legacy identifiers left running — and PPC Land's coverage has been working through the implications since. It landed in the trade press again this week.
For most brands doing $200K/mo on Amazon this is genuinely not urgent, and I'd rather say that than manufacture alarm. Multi-country buying from one login, unified billing and permissions on one page, and clean-room access provisioned by the upgrade are all straightforwardly good if you're running DSP across markets. The item belongs on this list for one reason: it is the third change in one week that arrived as an automatic upgrade requiring nothing from you.
If you do have DSP exposure, the ten-minute job is permissions. Unified permissions on one page means whoever had access to one market now sits in a console that reaches 34. Go and look at who's on the account, confirm every name is someone who still works with you, and confirm the billing owner is your entity and not an agency's. Consolidation events are the cheapest possible moment to find a stale login, and the worst possible moment is October.
4. Google shuts down the Imagen 4 API models tomorrow
The Imagen models are deprecated and reach shutdown on August 17 across both the Gemini Developer API and the Agent Platform Gemini API, with Google's own migration guidance pointing at the Gemini image models (gemini-2.5-flash-image) and a generate_content call instead.
Last week's synthesis had a retirement cluster in it — Opus 4.1, o3, the DALL·E GPT, a robotics preview model — and I made the point that a dated shutdown is the good version of a model change, because it announces itself. This one is worth its own entry because it's the first in that cluster that lands on a creative pipeline rather than a text one.
If any step in how your listing imagery gets made runs against an Imagen model string — background generation, lifestyle scene fill, concept iteration, a vendor's tool you never asked about — that step stops working tomorrow. The failure will be loud, which is fine. The thing worth doing is smaller and more durable: most operators who built an automation inventory this summer put a pricing review date on it and never added a retirement column, and the ones who did add it were thinking about text models. Your image and video generators are dependencies with end-of-life dates too, and they're the ones with a client deliverable attached.
While you're in there: ask any creative vendor which image model they run on and whether it's pinned. A vendor who answers with a model name has thought about it. One who answers with a paragraph about their AI-powered workflow has not, and that reply is the answer.
5. Gemini 3.7 Flash launched Thursday with a price that expires on a date
Google released Gemini 3.7 Flash on August 13 at an introductory $0.75 / $3.75 per million input/output tokens through December 31, 2026, moving to $1.50 / $7.50 standard on January 1, 2027. The same promotional rate was applied retroactively to 3.6 Flash, so through year-end the two cost the same despite launching three weeks apart.
I got a price wrong in public nine days ago — I told operators to diary Anthropic's September 1 Sonnet increase and Anthropic then cancelled it — so I'll be precise about what I think this means. It is not "AI is getting cheaper." It's not a trend at all. It's the fourth repricing event I've had to write about in eleven weeks, in both directions, and the second one carrying a published expiry date attached to a cheap high-volume tier.
The tier matters. This is the price point that decides whether catalog-scale work pencils: bulk copy passes, review mining across forty ASINs, attribute completeness audits, A+ module copy for the long tail, search-term classification. High token volume, low value per individual call, run across everything. Those jobs get shelved and un-shelved on arithmetic somebody did once.
So: January 1, 2027 goes in the same column as every other expiry you're tracking. Not the price — the date. The lesson from getting it wrong last week isn't that I should have found a better number, it's that a cost model built around a number is fragile and one built around a re-check date isn't.
What I'd ignore
The DeepMind leadership change. Koray Kavukcuoglu taking over day-to-day at DeepMind is a real story and it contains zero decisions for anyone selling physical products. Org charts at frontier labs are not an input to your Q4.
GPT-5.6-Cyber and the Daybreak security programs. Genuinely interesting, entirely irrelevant to a listing.
Benchmark tables on 3.7 Flash. DeepSWE up 16.7 points, AutomationBench up 13.4. Real evals, none of which measure whether a model writes a bullet that won't get your listing suppressed. That's still the only benchmark that bills you when it's wrong, and you still have to build it yourself from twenty of your own real SKUs.
Any advice to "audit your AI spend before September 1." That genre sprang up three weeks ago around an increase that isn't happening. Most of it hasn't been corrected. Mine has.
The urge to swap creative model providers because of the Imagen shutdown. Migrating is a config change. Re-validating output against a fixed set of your own products is the expensive part, and a forced migration is a bad week to discover you never had that set.
The through-line
Four times this year I've written about a default that moved underneath an operator who did nothing wrong. Every previous instance was Amazon's, and the honest answer was always a calendar: check more often than the platform's clock runs.
This week's most important instance was different. Claude Code's default is one you own. There is a settings file, an admin control, and a deny-rule syntax, and every one of those beats the best-intentioned instruction you could write into a prompt. That's a rare thing on this list — a default you can actually pin.
Most operators will pin nothing, because nothing broke on Thursday and nothing will break next Thursday either. That's how these things work. The exposed ones are the operators who built something good enough that they stopped watching it.
Sources: Anthropic's auto mode announcement (Aug 7), TechCrunch (Aug 9), PPC Land on creator placements, PPC Land on the DSP global account, Google's Imagen migration guidance, Gemini API release notes. Dates and availability change; check your own console and your own Seller Central notifications rather than a blog post, mine included.