Yesterday was the unsigned Ox Alpha preview and a custody card that still keeps your prompts. Today the signed product is selling inventory under the answer.

OpenAI said last Tuesday that ChatGPT Ads would expand to 31 European countries. Monday is the switch. Notebookcheck is the clean day-of read: Free and Go see a labeled unit under the reply. Plus, Pro, Business, Enterprise, and Edu stay dark. Temporary Chats stay dark. Accounts that look under 18 stay dark.

The interesting part is what is not on yet. OpenAI’s own FAQ says personalized ads are not initially available in the EEA or Switzerland. There is nothing to consent to on day one. What picks the unit is the conversation you are having, plus rough location and language. Earlier chats, saved memories, and ad history stay out of it. Consent becomes a later product.

That is a more honest launch than the “opt-in personalization” headlines. It is also a more fragile one. Conversation-context ads are still ads attached to a goal. Planning a trip, comparing laptops, asking how to start a hobby — those are the moments OpenAI is selling. The company says ads stay labeled, sit apart from the answer, and do not steer the model. Treat that as a policy, not a theorem. Policies get product managers.

The help-center table is the 31-market list OpenAI’s blog did not print: the 27 EU states plus Iceland, Liechtenstein, Norway, and Switzerland. Advertisers buy through the Ads Solutions team and the usual agency groups first. Self-serve Ads Manager is “later this summer.” Plus stays the off switch if you do not want the unit. The free plan still has a quieter off switch that costs you messages and tools.

If you actually run agents against ChatGPT instead of chatting, this does not change your API bill. It changes the default surface your non-technical teammates live on. The free box now has a sponsor. Paid still does not.

A used ThinkPad and a USB-C dock remain the boring way to keep the noisy consumer client off the machine that holds the keys.

Secondary: DeepSeek’s vision Flash is an experiment with two loud benches

DeepSeek put V4-Flash-Vision-Exp on the paid developer platform on August 21. No free tier yet. No architecture card for the vision variant. The parent is still V4 Flash: a 284 billion parameter mixture of experts that lights 13 billion per prompt, trained on 32 trillion tokens, with HCA and CSA compressing the KV cache. DeepSeek says that pair cuts the compute for a million-token prompt by 73 percent.

Vendor benches, so keep the salt:

Opus 4.8 is last generation. Fable 5 and Opus 5 were not in the table. Independent runs on Artificial Analysis and LMArena have not landed yet. Still, a cheap MoE that can look at a screenshot and keep going is the actual product shape people will try this week. If you are going to throw UI dumps at it, a decent webcam and a second monitor beat photographing a laptop with a phone.

Secondary: Faraday is a 27B science agent with a vendor scoreboard

Inherent, a London lab of DeepMind alumni, came out of stealth with a 50 million seed and shipped Faraday. The claim: independently reproduce published papers without being handed the answer. The stack is the part that matters. Core reasoning sits on Alibaba’s Qwen 3.6 at 27 billion parameters. Code goes to OpenAI’s GPT-5.5 Codex. They did not train a new frontier model. They trained a loop, then rented the coder.

Cofounder Edward Hughes told TechCrunch the interesting result was not beating Claude Opus 4.8 and GPT-5.5. It was the training method: reinforcement learning for “research taste” rather than cloning how science is written down. The bar they set is higher than accuracy. They want the agent to pick experiments worth running. Inherent’s own numbers. A dozen people in King’s Cross. Treat the leaderboard as a demo until someone else reruns it.

That is the same thesis as last week’s harness stories, pointed at papers instead of games. A small team wrapping an open-weight reasoner plus a rented coder is a shape more labs will copy than another unsigned frontier checkpoint. If you want to try the cheap half of that stack locally, Qwen-class weights still fit a workstation with enough RAM.

The bill under the models is DRAM

Bloomberg via Fortune, Saturday: Nvidia’s contract server builders told Microsoft, Google, and Oracle to expect more than 15 percent on AI server systems shipping early 2027. Vera Rubin and Grace Blackwell configs are in the pile. Nvidia did not comment. The driver is not the GPU sticker. It is DRAM from Samsung, SK Hynix, and Micron. Nvidia still prints a 75 percent gross margin and still cannot eat the memory bill.

Token prices are moving the other way. Anthropic made Sonnet 5’s 2 / 10 introductory rate permanent and canceled the September 1 step-up to 3 / 15. OpenAI already cut GPT-5.6 Sol about 20 percent last week. Software is on sale. The rack is not. That gap closes later, not in the blog post.

Monday’s takeaway is narrower than another model chart. The free ChatGPT box in Europe now has a labeled unit under the answer. A cheap Chinese vision Flash is claiming two hard visual benches against last year’s Opus. A 27B science agent is claiming it can rebuild a paper. And the people who buy the racks were told the 2027 invoice goes up.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.