The pattern this week was lopsided in a way worth naming before the items: everything that happened on the demand side of AI shopping was about scale — a billion people just lost their rate limit, and Amazon put numbers on how much more its AI-assisted shoppers spend — while everything on the tooling side was about control: inspection layers, security scanning, spend caps, and a stack of dated shutdowns. Nobody shipped operators a new capability this week. They shipped deadlines and governors. Both matter more than whatever the model-of-the-week discourse was arguing about.
Here are the five things, and what I'd actually do about each.
1. OpenAI removed the limit on text chats for every ChatGPT user
On August 7, OpenAI began rolling out unlimited text-based chats for all ChatGPT users — a platform that just passed 1 billion weekly users — with the new GPT-5.6 Luna model becoming the default for Free and Go tiers and GPT-5.6 Sol going to Plus and Pro (per OpenAI's release notes, with rollout reporting from The AI Insider, Aug 7). OpenAI's internal testing claims factual errors dropped 62% for Luna and 68% for Sol versus GPT-5.5-Instant — their number about their own product, so hold it loosely, but the direction is the point.
The operator read: the rate limiter just came off your customer's product research. The marginal shopper — not the AI enthusiast, the normal person on the free tier — can now interrogate a purchase for as long as they want, with a model that's meaningfully better at facts. Every claim on your listing is now read by an assistant with infinite patience and improving skepticism, at zero cost to the shopper. The brands exposed here aren't the ones with bad products; they're the ones whose titles, bullets, and A+ make claims their attributes don't back up. An assistant that's 62% better at facts is 62% better at noticing your bullet says "fits all models" while your compatibility attributes are blank. The attribute-completeness work I keep banging on about just got a billion more readers.
2. Amazon's AI-assisted shopper is measurably its best customer
The number everyone missed: more than 350 million shoppers have used Alexa for Shopping in the past 12 months, active users nearly doubled year-over-year in Q2, interactions are up 5x — and US customers who use it spend about 40% more per order than non-users, with Alexa+ users joining Prime at nearly 25% higher rates (Amazon Q2 2026 earnings call, with follow-up coverage from CX Dive updated Aug 3). Last Sunday's synthesis took the advertising line from that same call; this is the other half, and it surfaced properly in this week's coverage.
The operator read: stop modeling the AI shopping layer as an early-adopter slice you'll deal with later. The 40%-more-per-order figure means the shoppers arriving through the AI layer are, right now, your highest-value traffic — and whether you're in their consideration set is decided by structured attributes, honest natural-language content, and machine-legible creative, not by your bid. If you've been doing that work, this is the receipt that it serves your best customers today, not a future cohort. If you haven't, you're thin on exactly the surface where the money already moved. The dumb take is "voice and chat shopping is still tiny" — 350 million people with a 40% basket premium is not tiny, it's just invisible in your reporting, which is a different problem.
3. Anthropic shipped inference hooks — the write-path gate becomes a product
On August 5, Anthropic released inference hooks in beta for Claude Enterprise: a policy layer that inspects every prompt and every tool call before it reaches the model, with enforcement, shadow mode, and role-based exclusions, across chat, Claude Code, and Cowork (per Anthropic's release notes). In plain terms: a compliance gate that sits in front of the model on every single call, that someone other than the person running the prompt controls.
The operator read: I've spent months telling operators to hand-build exactly this — constraint sentences in skill files, approval stops, "the human approves the write." That discipline now has an enforcement layer you can buy instead of improvise, and it changes the vendor conversation. The question for your agency or your SaaS tools upgrades from "do you pin your model version" to "what inspects your tool calls before they execute — and who set that policy?" A vendor running agents against your catalog with no inspection layer isn't reckless by 2024 standards. By this week's standards, they're running without equipment that exists. Ask the question in writing; the shape of the answer is the answer.
4. Skills and plugins became a supply chain, and Anthropic started scanning it
On August 6, Anthropic released skill and plugin security scanning in beta for Enterprise plans — automated detection of malicious content in the skills and plugins your team runs — alongside a public beta of self-hosted environments that keep repos, artifacts, and secrets on your own infrastructure (per Anthropic's release notes).
The operator read: the fact that this product needs to exist is the news. Skills — the small instruction files that drive agent workflows, including the eleven I run my own creative operation on — are now a supply chain, and a supply chain can carry something hostile. A skill file has whatever access you gave the agent that runs it: your catalog, your ad console, your inbox. Most operators I know have installed at least one community skill or plugin they never read. This week's move: actually read them. It's a ten-minute job — they're short text files, that's the entire point of them — and you should be able to say out loud what every installed skill is permitted to touch. If your team can't produce the list of what's installed, that's the finding.
5. Four models got shutdown dates — availability expires too, not just pricing
The quiet cluster: Anthropic retired Claude Opus 4.1 from its platform on August 5. OpenAI retires o3 from ChatGPT on August 26 and the DALL·E GPT on August 30 (per OpenAI's release notes). Google shuts down gemini-robotics-er-1.6-preview on August 31 (per the Gemini API changelog), and the Sora API is reported to sunset September 24 — that last one I've only seen secondhand, so verify against OpenAI's own notes before acting on it.
The operator read: Thursday's post argued that model pricing now behaves like a SaaS trial — promotional rates with expiry dates you have to diary. This week is the other column in the same spreadsheet: availability expires too. Pinning your model string protects you from silent behavior drift; it does nothing when the string itself gets deleted, and a pinned string with a published retirement date isn't a contract, it's a countdown. The fix is the same inventory I keep prescribing, with one more column: every automation, the model string it runs on, what it costs, when the price changes, and now — whether that string has a retirement date anywhere in the vendor's notes. Fifteen minutes, quarterly. The operators who get burned by retirements are never the ones running messy experiments; they're the ones who built something reliable eighteen months ago and stopped reading release notes because it kept working.
What I'd ignore this week
The GPT-Live-1 voice rollout and the donut-speaker hardware leak. The new voice stack is a consumer experience; the $300–400 screenless speaker is a 2027 device. I said it in July about the last hardware leak and it stands: calendar, not roadmap. Nothing about a device that doesn't ship for a year changes a decision you make this quarter.
The "factual errors down 62%" number as a benchmark event. It's directionally real and I used it above — but it's OpenAI grading OpenAI. The only eval that bills you when it's wrong is the golden set you run against your own SKUs, and no lab will ever publish that one for you.
"AI bubble" discourse, round forty. A week of governance features and retirement dates is what a maturing infrastructure looks like. Zero decisions in the takes, either direction.
Any panic about the retirements themselves. A dated shutdown with weeks of notice is the good version of this problem — it's the silent swaps that cost money. The failure mode isn't that models retire; it's that nobody in your operation is assigned to notice.
The week's summary, honestly: the AI shopper got unlimited and turned out to be the best customer in the store, and the tooling world spent the week installing brakes and posting closing dates. If your listings are legible to machines and your automations are written down in a list somebody owns, this was a good week for you. If neither is true, everything above is homework.