AI news #3:
Anthropic shipped Opus 5 — frontier at half the price

MCP breaks compatibility today — some connectors will drop out

MCP is Anthropic’s standard for connecting AI agents to outside services: it is what every seller-facing MCP server we covered in earlier issues runs on (Helium 10 and the rest). Today, July 28, the largest revision of the protocol since launch is being finalised — and it breaks backward compatibility.

• The protocol goes stateless: sessions and the handshake are gone and the housekeeping data rides along with every request, so ordinary network infrastructure can route agent traffic without special handling.
• MCP Apps arrive: a tool can hand the chat an interactive panel with buttons instead of a wall of text. Long-running operations move into a separate Tasks extension.
• Sampling, Roots and Logging are deprecated — you get 12 months to move off them.

The Gemini 3.5 Pro flagship slips again — Alphabet stock dips

Bloomberg reported that Google is months behind schedule on Gemini 3.5 Pro (it was promised back in June). The model is stalling on coding after a botched training-data update. Alphabet stock fell about 4.4% — the market fears it is falling behind OpenAI and Anthropic.

• Google’s entire shopping bet — Universal Cart, agentic checkout, AI Mode with purchases — rides on Gemini.
• With the flagship stalling, the pecking order among agentic storefronts is still unsettled.

Anthropic and Blackstone launch Ode — a Claude rollout squad, aimed at retail

Who is Blackstone? The world’s largest private equity firm: it buys and grows real businesses and real estate with pension and insurance money. Roughly $1.35T under management — more than any rival — and 250+ companies in its portfolio. Together with Anthropic and Hellman & Friedman it launched Ode (a deal of about $1.5B): a separate firm that does not build new models but puts Claude into specific business processes.

• The model is a landing party: ~100 engineers sit inside the client company and wire Claude into its processes instead of writing endless consulting reports.
• The idea grew out of Blackstone’s own pain: rolling AI out across its 250+ companies through consultants kept going sideways. The announcement’s conclusion: the bottleneck is not access to a model, it is turning the model into a result.
• Retail is named outright as a priority industry — and where retail goes, e-commerce follows.

ChatGPT Work lands on the desktop — one app for Chat and Work

OpenAI shipped a desktop update: a global switch between Chat and Work modes, projects inside the app, and cloud Work tasks synced across web, phone and desktop. Work is a GPT-5.6 agent that spends hours assembling finished artefacts — spreadsheets, reports, dashboards — and connects to Slack, Gmail, Drive and Salesforce.

FLUX 3: the model behind product-photo services now makes video with sound

Black Forest Labs is the German lab whose FLUX models sit under most product-photo generators. It has released FLUX 3: a single model trained on images, video, audio and action prediction at once. The video version is in early access by request; the image version follows in the coming weeks.

• The announcement names e-commerce as a target use case, in the words “keeping the product and its materials consistent in motion”.
• The model outputs video with sound up to 20 seconds long. Testing partners include Canva, Picsart and Krea.

Stripe is in talks to buy OpenRouter for about $10B

According to the WSJ, payments giant Stripe is negotiating to buy OpenRouter — the service that acts as a single access point to hundreds of AI models from different companies, a kind of power socket for developers. In May OpenRouter was valued at $1.3B; the talk now is around $10B. In parallel, Stripe has a $53B bid in for PayPal.

• OpenRouter offers access to 400+ models from 60+ companies (OpenAI, Anthropic, Google, Meta) and has around 8M developers.
• For Stripe this is the second AI acquisition in under a year — model routing plus usage billing in one bundle.

Amazon is shutting down almost the whole Nova line — image and video generators included

Reuters, citing Business Insider: Amazon is winding down most of the flagship Nova models — the top-end Premier and Omni, the Reel video model and the Canvas image generator. Resources are being moved to a new top-tier model, a direct competitor to GPT and Claude, expected to debut at re:Invent later this year.

• Only three survive: Nova 2 Lite, Nova 2 Sonic (voice) and Nova Forge, the service for building your own models.
• Canvas and Reel are exactly the models Amazon built its AI photo and video generation for listings and ads on, and the ones a number of seller services run on through Bedrock.
• It comes shortly after Amazon closed its own AGI lab. The Nova brand is not being buried — the new model may ship under it.

AI Overviews now appear in 43% of Google searches, AI Mode is up to 279M visits

Similarweb (the web-traffic measurement service) published a fresh breakdown: Google’s AI answer now shows on 43% of queries, against 15% a year ago. Visits to AI Mode — the separate chat mode of search — grew from 126M in June 2025 to 279M in May 2026. Queries got longer and more conversational: Google is turning from a gateway to websites into the endpoint where the answer is given.

• Links inside AI answers are becoming rare: in May, ChatGPT cited sources on just 6.8% of US queries.
• Long conversational queries need a different listing and feed structure — built for the whole question, not for a two-word keyword.

From August 2 labelling AI creatives is mandatory in the EU — fines up to €15M

Google signed the European code of practice on AI content transparency, and on August 2 Article 50 of the AI Act comes into force. From that day AI content has to be marked twice: machine-readably, with invisible marks such as SynthID and the C2PA standard (provenance data inside the file), and explicitly, with a visible icon no later than the first time a user sees it.

• The rule covers everything generated or materially altered by AI: watermarks for text longer than 200 tokens, layered marking for images, video and audio.
• Assets created inside Google Ads tools are marked automatically — but responsibility for the visible label still sits with the advertiser.

Amazon Ads: ten new AI features — you can now steer the model with your own audiences

Brand+ and Performance+ are the autopilot campaign types in Amazon DSP, where the algorithm decides who sees the ad and where. Amazon shipped ten updates to them at once, and the main one is Audience Signals: you hand the model your own audience data as a hint. There is no extra charge — the announcement says “no additional fees” outright.

• Model upgrades: optimisation for brand outcomes, Prime Video signals plugged in, asynchronous remarketing in Performance+.
• New inventory: audio ads for Brand+ and Performance+, rolling out across 16 locales.

Manage Products has been rebuilt — with an AI assistant sitting inside it

Amazon quietly updated Manage Products in Seller Central, without a press release. The day used to look like this: a product falls out of search — one screen; fix the bullets — a second; check stock — a third. Now it is one page. Verbatim from the announcement: “Edit listings, fix suppressed ASINs, manage inventory and track product health all from one page”.

• AI right on the working screen. Amazon’s wording: “Ask Seller Assistant directly in Manage Products to share insights into your business or take selected actions on your behalf”. So you do not leave for a separate chat — you ask on the page with your own products: “why is this ASIN not showing?”, “what is about to run out?”. And it performs some of the actions for you.
• A configurable view. You can pick which columns to see and which to hide and save it as your own preset — the listing manager and the supply person get different screens instead of one 40-column sheet.
• Manage Shipments and Manage FBA Inventory were updated too. Amazon stresses separately: “Your existing data and workflows remain unchanged” — nothing breaks, this is an interface rebuild.

Study: Amazon’s AI (Alexa for Shopping) does not look at search results

Marketplace Pulse ran about 2,000 queries and compared what Alexa for Shopping recommends with ordinary search results. The outcome: 63.9% of recommended products were outside the organic top 10, and 40.9% do not show in normal search at all. Only 14.3% of the recommendations were sponsored.

• The AI picks the product from context and shopper intent, not from search position or ad bid.

Amazon closed its AGI lab and stepped out of the race for a superbrain of its own

Amazon confirmed layoffs in its artificial general intelligence unit and closed the AGI Lab in San Francisco — the lab set up in December 2024 to build AGI and business agents. The company’s wording: “sharpening our focus on the initiatives that matter most to customers”. The unit’s leadership had already left earlier.

Alibaba launched an agent that finds suppliers and haggles for you around the clock

Alibaba.com released the Accio Sourcing Toolkit. You describe what you need and the agent picks factories itself, writes to several at once, and negotiates price, lead times, minimum order quantity and shipping terms — around the clock, without you pinging anyone between negotiation rounds.

• The agent runs several conversations in parallel, compares offers, cites competitors’ better terms inside the negotiation, and follows up with whoever went quiet.
• A hard stop: the agent not close the deal, not confirm the order or not make the payment — the decision stays with you.

For the first time shoppers picked AI over the brand’s own site

Bloomreach (a personalisation platform for online stores) surveyed 4,040 adults in the US and the UK. 41.4% said they would rather buy through an AI tool, and only 38% directly on the brand’s site. A year ago it was exactly the other way round: 58.9% for the brand site against 41.1% for AI. A parallel CI&T study shows the same picture: 74% of Americans have already used an AI agent while shopping, and 90% have used one or are willing to try.

• 75.4% have already used AI for purchase decisions, and 80.4% said the experience met or beat their expectations.
• 37.5% started spending more, 39.5% buying more often.
• Top use cases per CI&T: comparing brands (55%), finding the lowest price (55%), finding where to buy (49%).

Makers are fleeing Etsy over a flood of AI listings — an AI-banning rival opens on August 1

Handmade sellers complain that cheaper AI content ranks above their work in Etsy search. An artist from Boulder describes a 30% drop in sales; a British illustrator, minus 98% over a year. Etsy has allowed AI art since 2024 provided the seller discloses it, but a company VP admitted that “enforcement is a tricky thing”.

• The rival Fybe — a marketplace that bans AI content and dropshipping outright and moderates by hand — opens seller registration on August 1.