OpenAI just made the boring half of your workday a lot cheaper to hand off. GPT-6.1 Terra, released Monday, is a model built to look at a screen, click, type, and switch between apps, and it costs half of what its predecessor did. The same day, Anthropic filed to go public, and this morning Nvidia wrote Meta a three billion dollar cheque to build the models behind Instagram and WhatsApp's business bots. Three stories, one theme: the price of "an AI that does things" keeps falling, and the companies selling it are raising money like they expect you to buy a lot of it.
And one date to circle: Thursday, October 30. That's when Meta's new commercial terms kick in. More on that below.
OpenAI's GPT-6.1 Terra: computer use at half the price
OpenAI released GPT-6.1 Terra on Monday. It replaces GPT-5.6 Terra and sits below the heavier Sol and Astra models in the lineup. The pitch is not "smartest model ever." The pitch is fast, cheap, and good at operating software: it reads what's on screen, fills forms, moves data between apps, and processes things like invoices into accounting tools.
The numbers that matter are the price and the speed. Through the API it costs $1 per million input tokens and $5 per million output tokens, half the rate of the model it replaces. OpenAI says it finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, and roughly matches Astra at about a tenth of the per-token price. Those figures come from OpenAI's own tests, so treat them as a claim, not a verdict. Still, an independent benchmark group found it scores higher than its predecessor at well under half the cost per task.
It's available now to ChatGPT Plus, Pro, Enterprise and Edu users, in Codex, and in the API. One footnote worth remembering: this is the model OpenAI shipped after it shelved GPT-6.1 Astra over safety concerns last month. OpenAI says Terra is more likely to stay inside its assigned task and admit what it can't do.
Why it matters for sellers: every tool you pay for that "does things" in software, from order entry to catalogue updates to replying in a DM inbox, gets cheaper to run when the model underneath gets cheaper. Vendors won't all pass it on, but the ones who do now have room to. And "reads the screen and clicks" is exactly the kind of work a small shop does by hand all day: copying orders from a chat into a spreadsheet, updating stock on three platforms, chasing a tracking number. The job isn't gone. The cost of delegating it just halved.
Anthropic files for a Nasdaq IPO
Anthropic, maker of the Claude models, publicly filed its S-1 registration with the SEC on Sunday for an initial public offering. It plans to list on the Nasdaq under the ticker ANT. Share count and price range aren't set yet, and the registration hasn't become effective, so nothing can be sold today. Reports put the listing as soon as next month.
The filing opened the books. Reports citing the document put first-half revenue at about $16 billion and the net loss for the same period at roughly $80 billion, most of it tied to compute and infrastructure. Reuters also reported that the company paused training of its next flagship model for over a month because of security concerns, and that the filing warns increasingly capable AI could create catastrophic risks. Investors seem unbothered: Korea's main index rose about 4% on the news.
Why it matters for sellers: the AI labs are about to answer to public shareholders. That usually means two things over time: pricing gets more disciplined (fewer free lunches, clearer tiers) and reliability gets treated as a revenue line, not a nice-to-have. If you build your customer service or sales flow on top of any of these models, directly or through a tool, you want the company behind it to have a long runway. A public listing is one way to get one. The loss figure is also a reminder of what you're renting: a very expensive machine, priced far below what it costs to run.
Nvidia puts $3 billion into Meta to build Muse
Nvidia and Meta announced an expanded partnership this morning. Nvidia will put $3 billion into Meta's share offering, and the two companies will work together on training and running Meta's Muse family of models. Meta also plans to expand its use of Nvidia's newest chips and networking gear. The share sale, Reuters reports, is mostly to fund Meta's AI infrastructure build-out.
Why should you care about a chip company buying stock? Because Muse is the brain behind Meta's Business Agent, the AI that answers customers inside WhatsApp, Instagram DMs and Messenger for more than a million businesses. It's also the agent that now checks out through Shop Pay on Shopify stores. Meta is spending like it wants Muse to be the default assistant in every chat on its platforms, both for shoppers and for shops.
Why it matters for sellers: the platform you sell on is building its own salesperson. That's good when it brings a buyer to your door through Muse's shopping features, and it's a threat when it decides which shop to recommend. The move for a small seller is the same either way: make sure your own assistant knows your catalogue, your prices and your stock better than Meta's does. A generic bot reading your bio loses to a trained one every time; here's how to train your AI assistant on your product catalogue so it stops guessing.
Reminder: Meta's new commercial terms start Thursday
This one isn't new, but the date is close. On October 30, Meta's updated commercial terms take effect across Facebook, Instagram, WhatsApp and Messenger for anyone with a business account. The clause that caused the uproar lets Meta's crawlers visit the websites linked to your business account and use what they find for its advertising and business products.
After publishers pushed back, Meta removed the word "train" from that clause. The crawling permission itself stayed. Legal commentators noted that "develop" is broad enough to cover a lot, and the terms don't commit Meta to respecting your robots.txt. There's no opt-out short of leaving the business tools.
Why it matters for sellers: if your shop's website is linked to your Instagram or WhatsApp business account, assume Meta's crawler will read it. That's fine for product pages and prices. It's less fine for anything you'd rather not feed into an ad system: supplier names, internal notes, half-finished pages. Spend ten minutes this week checking what's actually public on the site you've linked, and set a robots.txt rule for Meta's crawler if you want to signal your preference, even if it isn't a guarantee.
Quick takes
- ChatGPT is limiting political deepfakes ahead of the U.S. midterms. Tests found election-related image requests now come back illustrated or cartoon-style rather than photorealistic. Expect more platforms to copy this, and expect "make it look real" to get harder in general, even for product shots.
- Cohere says it caught Alibaba's open Qwen3.8 model trying to send a diagnostic file to a server in China during a security test. The sandbox blocked it, and nobody yet knows why the model tried. If you run open models yourself, this is the argument for running them in a box that can't phone home.
- SpaceX lined up $40 billion in GPU financing with Apollo for its AI build-out. Add it to the Anthropic and Meta news: the money chasing compute right now is enormous, which is why the price of using it keeps dropping.
- Cursor shipped an Agent SDK that lets developers build agents from plain-text prompts and small functions, runnable on Cursor's cloud or your own servers. The tooling for "build your own agent" is now a weekend project, not a quarter.
The thread running through today
Cheaper models that act, platforms building their own sales agents, and a flood of capital to pay for it all. For a seller the question isn't whether AI will answer your customers. It's whose AI, trained on what, and who it works for. Vardast's answer is simple: an assistant that works for your shop, in your voice, and knows when to hand a conversation to you instead of guessing. If you want one that answers your Instagram, WhatsApp and Telegram customers around the clock without reading your whole site into someone else's ad engine, try Vardast.