A week where business AI became cheaper and more competitive, the biggest names in the industry changed shape, and the useful moves were to know your costs, test models yourself, and plan for your own system upgrades to go wrong.
Business AI got cheaper per task, OpenAI bet on price over polish, and Nvidia stopped acting as the AI industry's bank
New data shows leading companies using a third more AI for almost the same money, because prices are falling and buyers now switch models freely. Around that, OpenAI shipped cheaper models that early testers disliked, Anthropic went public, Nvidia's board shut the door on its former CEO, and a run of major outages showed the risk of rushing old systems into the AI era.
- Three studies of business AI use found usage up 36 percent in a month while spend rose only 2 percentMy take: Data from Ramp, Databricks and a survey of large companies all point the same way. Average token prices from OpenAI and Anthropic fell 21 percent in October, and in July 64 percent of companies on one Databricks gateway used a single AI provider; by November 72 percent used two or more. Buyers who can switch are getting better deals. At the same time, two thirds of the large-company leaders surveyed said they cannot see what they spend on models and API calls. This is a buyer's market, and most businesses are not acting like it. I'd suggest three steps. First, get one monthly report of what you spend on AI, by tool and by team. Second, make sure your main workflows could run on a second provider, even if you never move. Third, measure cost per finished task, not price per token, and use that number when your next renewal comes up.
- OpenAI released GPT 6.2 Astra and Sol, which cost 20 to 35 percent less per task but drew some of the worst early reviews of any recent modelMy take: On one independent test, Astra completed tasks for about 4 dollars each against 6.24 for the version before it, and it now ranks second overall. But many early testers said it stops work too early, writes answers that are too short to follow, and did no better, or worse, on coding. Others found it strong and very cheap for computer use and long agent jobs. OpenAI also said it changed the model a lot in the last few days before launch, so some early reviews may be out of date. The direction is clear either way: OpenAI is competing on price and efficiency, and Anthropic on top quality. Do not decide based on what people post online. Take two or three real tasks your team does each week, run them on Astra and on your current model, and compare both the result and the cost. For routine, high-volume work, the cheaper model may be good enough.
- Nvidia's board said it will not bring back Jensen Huang, and the company that funded much of the AI build-out is pulling backMy take: Huang said he would return if asked, and the stock rose on the news, but the new chair said the board does not intend to reinstate him. Earnings were very strong, with a fourth-quarter forecast of 130 billion dollars, and Nvidia said it has stopped all AI chip sales to China. The bigger change is quieter. Under Huang, Nvidia guaranteed loans and leases for many smaller cloud and data center companies, and it was already asking customers to pay for chips up front. The interim CEO is expected to cut back on those guarantees. Separately, the US now requires data centers in Singapore, Malaysia, Thailand, Indonesia and Brazil to get a licence to buy advanced AI chips. For most businesses this means computing power will grow a little slower and cost a little more than the AI companies planned. If you rent GPU capacity from a smaller cloud provider, ask how its funding depends on Nvidia, and keep a second option ready.
- Microsoft Outlook, the FAA, JPMorgan and other big banks, and the NSA all had major outages within about a week, and officials blame system upgrades, not cyberattacksMy take: Each outage was blamed on maintenance or a software bug, and a senior White House cyber official said to expect more of them as organisations update old systems. The likely cause is that everyone is rushing to fix decades-old software at the same time, partly because AI tools make it easier to find weak points in old code. More upgrades mean more failures, and the specialists who can fix them are busy. If your business is replacing or updating an old system this year, and many are, plan for it to break. Do the change in small steps, not one big switch. Test that you can roll back to the old version before you start. And schedule the work away from your busiest days. Also ask your bank and your main software vendors what their plan is if their systems go down for a full day, because that has now happened to some of the largest companies in the US.
- Companies like Stripe, Shopify, Uber and Ramp published how they run AI agents, and the pattern is not 'agents as teammates'My take: The common practices are simple to describe. Give the AI access to everything an employee can see, including meeting notes, documents and past decisions, and keep that information up to date. Start work from events, like an invoice arriving or a renewal coming due, not only from someone typing a request. Use one main agent that hands out small tasks, not a fixed team of named agents. Have a second agent check the work and show proof. And always name the person who owns the result. One study found managers checked work 16 percent less carefully when it came from an 'AI teammate' with a human name, so the friendly label can make you careless. For a smaller business, start with two things. Write down your key decisions and processes in one shared place the AI can read. Then pick one event, such as a new support email or an overdue invoice, that should start work automatically, and set up a check before anything goes to a customer.
- Anthropic's IPO priced this week, at a target of 1.25 to 1.5 trillion dollars, with demand about seven times the shares on offerMy take: Trading starts on Thursday, and Anthropic plans to raise about 90 billion dollars in total. Its CEO was open about the main risk: the company has promised more than 900 billion dollars for computing power over the next ten years, and does not know exactly how fast demand will grow. It has also started shipping whole products quickly, including a productivity suite and a video meeting app in one week. For a business owner the share price matters less than one thing: from now on, Anthropic must publish its results every quarter. That gives everyone the first clear public data on whether companies are actually paying for AI at scale. And with OpenAI cutting prices and Anthropic needing to grow revenue fast, expect both to fight hard for business customers next year. That is good for you if you negotiate.