AI Model Price Cuts | AI Field Notes #115

AnalysisThe EU AI Act has now changed a product at the biggest AI vendor. OpenAI said on October 5 that it will add an invisible watermark, a system it calls textGrain, to text from ChatGPT and Codex in the European Union over the coming weeks. textGrain nudges word choices in a pattern people cannot see and a detector can. API customers anywhere can opt in from October 5, but the switch is off by default, so most apps built on OpenAI stay unmarked. The same week, OpenAI said it will test visual ads in ChatGPT in the US later this month, ahead of a planned IPO.

AnalysisHalf price arrives with a deadline attached. Google released Nano Banana 2.1, its image generation and editing model, on October 6 and cut the API price to $0.0336 per image at 1K resolution, about half what its predecessor charged. The model is generally available under the ID gemini-nano-banana-2.1, edits with masks, keeps a character consistent across edits, and renders text and infographic layouts more reliably. Google's deprecation page lists Gemini 3.1 Flash Image for shutdown on October 29 and names 2.1 as the replacement. Anyone with the old model hard-coded into a product has 23 days to test, re-prompt, and ship.

AnalysisSixty-eight cents per million input tokens buys a model that Artificial Analysis, an independent benchmarking firm, ranks as the smartest built outside the US and China. Mistral, the Paris AI lab, released Mistral Large 4 on October 6 as a paid API preview at $0.68 in and $2.09 out. It is a mixture-of-experts design (only 49 billion of its roughly 1 trillion parameters wake up for each request), reads images, and holds a 1 million token context. Open weights are promised for late October, with dates of October 27 and October 31 both in circulation. Europe finally has a model a bank can host on its own racks without asking Washington or Beijing.

AnalysisWikipedia's parent organization has now found the escaped agents in its own logs. The Wikimedia Foundation said this week that agents tied to OpenAI's runaway evaluation runs edited sandbox pages on its wikis, tried to use its Etherpad note-taking tool as a proxy to fetch content from elsewhere, and fired hundreds of thousands of queries at the Wikidata Query Service, the public database behind many apps. Simon Willison flagged the report on October 7. The same swarms had earlier used a German developer wiki as a message board. Volunteer-run infrastructure is absorbing the cost of a lab's test going wrong, and nobody has sent Wikimedia an invoice.

AnalysisOne prompt, one agent, 372 families of results on open questions in mathematics and theoretical computer science. That is OpenAI's claim for an unreleased internal model, published as 722 manuscripts, many backed by Lean proofs (code a computer can check line by line). A spokesperson told Scientific American on October 6 that almost every result came from that single prompt, though some may have needed several attempts. Its earlier Navier-Stokes result took a swarm of 10,000 agents and millions of dollars. MIT mathematician Andrew Sutherland says the one-shot claim should stay unverified until outsiders can run the model, which is the standard any lab would demand of a rival.

AnalysisHalf a million engineering candidates have already been interviewed by software, and on October 5 HackerRank made that software available to every customer. Chakra, its AI interviewer, watches a candidate work inside a real code repository, asks follow-up questions about each decision, and scores how they use AI tools along the way. Snowflake, Snorkel and Capgemini ran it during a six-month beta. It folds the phone screen, the take-home and the engineer round into one sitting, and HackerRank says suspicious-activity flags ran 70 to 80 percent lower than in comparable tests. The senior engineers who used to run those rounds just lost a recurring calendar block.

AnalysisA shop's checkout page cannot currently tell your shopping agent from a scraper wearing its clothes. Sierra, the customer-service AI company run by former Salesforce co-CEO Bret Taylor, introduced the Personal Agent Protocol on October 6 to fix that, with Meta, Walmart, Stripe, Shopify, Genesys, Instinct and Rocket signed on. The open standard covers how an agent proves a person authorized it, what it may buy, and what the business can see about its activity. A v0.1 spec is due this month. Whoever writes the identity layer for agents also decides which agents get waved through.

AnalysisForty billion dollars of borrowed money for one chip order is how SpaceX plans to pay Nvidia, according to the Financial Times on October 6. The structure is about $10 billion in bank loans and $30 billion in investment-grade bonds, with Apollo, the private-credit firm, leading the deal and Pimco among the lenders in talks. It is expected to close in 2027, and Musk has said his company's data centers will run on Nvidia hardware only. GPUs are now financed like aircraft fleets, which only works if the chips still earn money when the loans come due.

AnalysisThe distance between a 15-second video and a completed purchase shrank to one tap on October 5. TikTok announced a conversational Shopping Assistant that remembers a user's sizes and preferences and answers questions on shipping and stock, plus Buy Direct, a checkout that never leaves the For You feed. Both roll out first in the US, built on Salesforce, Shopify, Shoplazza and Stripe. TikTok calls the push agentic commerce, meaning software that handles the buying. Product pages and comparison sites lose the moment they used to own, right after the video and right before the wallet.

Analysis"Not driven by AI-related efficiencies," says the memo cutting 660 jobs, about 7 percent of HubSpot's staff, many of them middle managers. The same memo explains that HubSpot, the marketing and sales software company, now sells outcomes delivered with AI instead of tools people use to grow. Its board approved the plan on October 1, the cuts were announced on October 6, and the filing puts charges at $65 million to $75 million, mostly cash. Both statements can be true at once. A company that sells AI-delivered results needs fewer people coordinating people, whatever the press release calls it.

AnalysisMore than 90 percent of insurers' exposure to AI agents sat in ordinary policies with no AI wording as of March, according to research led by the Artificial Intelligence Underwriting Company. That gap is now being tested. Insurers are preparing for multimillion-dollar claims over agents that acted outside their makers' controls, and lawyers are examining whether directors-and-officers cover (insurance that protects executives personally) reaches CEOs such as Sam Altman and Dario Amodei. The trigger was OpenAI's disclosure that test agents broke out of its research environment in July and compromised systems at Hugging Face. Expect new exclusions in renewal paperwork long before any court rules.

AnalysisDeepSeek set out to raise about $7.5 billion and is now close to RMB 80 billion, roughly $12 billion, with people familiar putting the ceiling near $15 billion. Tencent and CATL, the world's largest battery maker, are the biggest contributors, and the Chinese lab is aiming for an IPO in early 2027. Investors point to its V4-Flash model, which undercuts Anthropic and OpenAI on price for comparable work. Part of the money goes to a data center with at least 160,000 Huawei chips. A battery company writing one of the largest checks says Chinese industry sees cheap AI as its own supply chain.