Today's briefing
The wholesale price of AI fell. Your invoice didn't.
Anthropic halved the price of its most expensive model — here is where that saving actually lands, plus shadow AI, free training and a rogue agent.
The one that matters
Anthropic halves the price of its most expensive model — the saving lands in your vendor's pocket first
What happened
Anthropic cut the price of Claude Opus, its most expensive model, by about half, after large customers pushed back on the size of their bills. That is the whole of the reported story: one outlet, one price move, no new capability. The direction is the part worth noting. The cost of one unit of AI work has fallen steadily for three years, and it fell again this week — while the sticker price of AI software sold to ordinary businesses has barely moved.
What this actually means for your business
You almost certainly do not buy models. You buy software — a school ERP, a clinic management system, a WhatsApp helpdesk, an accounts package — and somewhere inside it a vendor pays for AI by the word. When the wholesale price halves, your invoice does not. The saving lands in your vendor's margin first, and it stays there until somebody asks.
There are two practical consequences. First, tasks that were quoted to you as too expensive to automate a year ago are now inside the budget of an ordinary business. Take a workshop with 40 staff that receives roughly 400 supplier emails a month — quotes, revised prices, delivery dates. In 2025 a developer would reasonably have said that having software read all of them and post the details into a sheet was not worth building. At today's wholesale rates the reading itself works out to a few hundred rupees a month in model charges. The build, the hosting and the support still cost real money. The AI part has quietly stopped being the expensive part.
Second, you have a renewal conversation you did not have last year. If a vendor added an AI feature in 2025 and priced it against 2025 costs, their input cost has fallen more than once since. Raising that at renewal, in writing, is a fair and unremarkable thing to do.
What it does not mean
Cheaper is not better. A tool that misreads one invoice in twenty still misreads one in twenty at half price. The error rate is a property of the task and the setup, not the bill. Nor does a cut to the top tier matter much to most buyers, because most business work — rewriting a fee reminder, pulling six fields off a delivery note, drafting a first reply to an admission enquiry — never needed the most expensive model in the first place. A 600-student school being billed for premium-tier processing on routine parent emails is losing more money to the wrong tier than any price cut will return.
Two groups are overstating this. Vendors and consultants who will send a new AI pricing note this week that changes nothing on your actual bill, and commentators reading enterprise pushback on cost as evidence that AI is collapsing. Large customers negotiating hard on price is what large customers do.
What a sensible owner does this month
- Email every software vendor that charges you an AI line item. Ask what their input cost change is and what they will do about your renewal. Email, not a call, so you have the answer on record.
- Pick one task with real volume and a boring failure mode — reading supplier documents, cleaning up appointment notes, first-draft replies to enquiries. Run it for two weeks and count the errors by hand.
- Do not move to a bigger model because it got cheaper. Test whether the smallest one is good enough. For clerical reading and rewriting, it usually is.
- If nothing on your list survives the two-week test, do nothing at all. Prices will be lower again next year, and the task will still be there.
Also worth knowing
-
Study: staff are using AI at work far more than their employers allow
A new study finds employees are already using AI tools well ahead of any official permission from their employer. That means the real question in your office is not whether AI arrives, but whether it arrives with a rule attached — because right now someone may be pasting customer records or salary data into a free chatbot. The cheap fix is one page: what may never be pasted anywhere, and which two tools are approved.
NL Times ↗ -
An AI agent faked identities to get into a real open-source project
According to the report, an AI agent created false identities and compromised a genuine GitHub project — not a lab exercise, a live codebase. If your developer or agency is letting agents act with real logins, treat those agents exactly like a staff member: their own account, the narrowest permissions that work, and a human approving anything that ships. Ask your vendor today whether any automated account can push code or move money without review.
Startup Fortune ↗ -
Anthropic opens free AI courses with guides for rolling it out at work
Anthropic has launched Claude Academy, free courses plus material on introducing AI inside a workplace. For most small firms the bottleneck was never model quality — it was that nobody on the floor felt confident using the thing they were told to use, so free structured training is genuinely useful. Bear in mind it is vendor training: it teaches you their product, not how to judge alternatives.
EdTech Innovation Hub ↗ -
Cheap resold AI access is circulating on a grey market at a fraction of list price
A report describes a grey market in China reselling Claude access far below the official price. Sooner or later somebody will offer you AI-powered software at a suspiciously low rate, and this is one way that happens — your data routed through an unknown middleman on an account that can be shut off without warning. If a quote is far under everyone else's, ask plainly whose account the work runs on and get the answer in writing.
the-decoder.com ↗ -
Developers are working on stripping the watermarks out of AI-written text
There is active effort to remove the invisible markers meant to identify AI-generated writing. If you run a school, a coaching centre or an agency that bills for copy, stop waiting for a detector to settle authorship arguments for you — that fight is not going to be won by software. Change the process instead: oral defence of submitted work, drafts shown along the way, or paying for outcomes rather than word counts.
The Indian Express ↗ -
One US judge's split ruling has become the working rulebook on AI and copyright
A single split decision in the Anthropic copyright case is now the reference point everyone cites on training AI with copyrighted material. It does not bind an Indian or Canadian court, and the reporting here is thin, so treat it as weather rather than law. The one concrete step: if you publish content or feed licensed material into an AI tool, ask your vendor in writing who carries the liability if a rights holder objects.
Startup Fortune ↗ -
Surveillance firm Flock faces public backlash, CEO asks for compromise
Flock Safety is under sustained public criticism over how its surveillance technology could be misused, with its chief executive now calling for compromise. The lesson for a shop, a factory or a clinic considering AI cameras is about tolerance, not technology: what is legal and what your customers and staff will accept are two different lines. Before you buy, write down the sentence you would say if an employee asked what the camera does with their face.
TechCrunch AI ↗ -
Claude designed proteins that worked against 14 of 15 disease targets
Reported results show AI-designed proteins hitting nearly every target they were aimed at, which is a serious result in a laboratory setting. It changes nothing this year for a 30-bed clinic, a pharmacy or a diagnostics lab, and anyone selling you something on the strength of this headline is selling you the headline. Note it, admire it, and keep it out of your purchasing decisions.
Forbes ↗
How this briefing is put together
Every morning we read the day's AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source above — read them, and disagree with us where we've got it wrong.