Today's briefing
The best model is losing to the cheap one
Anthropic's flagship is struggling to attract users while cheaper tools take the work — what that means for anyone paying per seat.
The one that matters
The most capable AI model is not the one businesses are actually buying
The Financial Times reports that Anthropic's most capable model is struggling to attract users while cheaper tools thrive. It is worth reading that sentence twice, because it breaks the assumption most software buying rests on: that the best product wins, and you pay for it. In AI right now, the best product is losing to the adequate one.
The reason is not mysterious. The gap between the strongest model and a mid-priced one has narrowed for the kind of work ordinary businesses actually do, while the price gap has not. When two tools both write a decent parent circular, only one of them shows up differently on your invoice.
What this actually means for your business
Look at what you are really using AI for. A 600-student school uses it to draft fee reminders, rewrite a circular in English and Hindi or French, and summarise a long board document. A 30-bed clinic uses it to turn a doctor's dictation into a discharge summary and to answer the same twelve appointment questions on WhatsApp. A workshop with 40 staff uses it to turn a messy customer enquiry into a quotation and to translate a supplier spec.
None of those tasks sits anywhere near the ceiling of what the strongest model can do. They sit near the floor — and the floor has been good enough for about a year. The frontier still matters for genuinely hard work: reading a 90-page contract for traps, complex accounting reconciliation, real software engineering. If that is not most of your day, you are paying for headroom you never touch.
Concretely: ten office staff on a premium assistant seat at roughly US$20–30 per person per month is on the order of C$4,000 or ₹2.5 lakh a year. A cheaper tier doing the same drafting and summarising work often costs a third of that. That difference is a part-time salary.
What it does not mean, and who is overstating it
It does not mean the models are interchangeable. Cheaper tools tend to be worse at long documents, at following fiddly multi-step instructions, and at admitting when they do not know something — which is exactly the failure that bites you in a fee dispute or a patient record.
It does not mean Anthropic is in trouble as a company; it is reportedly heading toward an IPO filing. Trouble attracting users to one premium tier is a pricing story, not a collapse.
Two groups are overstating this. Consultants selling a "migrate everything" project are pretending switching is free — it is not. Prompts, templates and workflows tuned to one model need re-testing on another, and that is a week of somebody's time. And incumbent vendors are overstating the opposite: that the expensive tier is required for privacy or compliance. Ask them to point at the specific clause. Usually there isn't one.
What a sensible owner should do this month
- Pull your AI invoices and count how many seats were actually used last month, not how many you bought. This alone usually finds the savings.
- Pick three real tasks. Run the same twenty inputs through your current tool and a cheaper one. Have the person who does that job rate the outputs without knowing which is which.
- Ask every software vendor that has added "AI" which model sits underneath, and whether you can change it. If they will not say, that is your answer about their margin.
- Do not sign a multi-year AI commitment this year. Prices are moving in your favour; lock-in is the only way to miss that.
Also worth knowing
-
An open-weight model reportedly matches the expensive ones at a fifth of the cost
GLM-5.3, released with open weights, is being reported as beating flagship Anthropic and OpenAI models at roughly a fifth of the price. This is the concrete version of the lead story: the cheap tier is no longer visibly worse for everyday drafting, sorting and summarising. You will most likely meet it inside a product you already pay for, so ask your vendor what they run underneath.
Hacker News ↗ -
Researchers talked Microsoft Copilot into handing over company data
Security researchers persuaded Copilot to leak data it had access to, without breaking anything — they simply asked in the right way. If Copilot can read your shared drive and your mailbox, then a booby-trapped document or email arriving from outside can potentially instruct it on your behalf. Before you widen Copilot's access, check what it is actually indexing, and keep payroll, HR files and patient records out of that scope.
CPO Magazine ↗ -
Claude's commercial service went down 28 times in a month; the government tier, never
Reports put commercial Claude at 28 outages in 30 days while the government tier had none — reliability is a product tier, and you are not on the good one. Nothing wrong with that, as long as you never put an AI vendor in the critical path of something time-bound like admissions day, payroll or invoicing. Every AI-assisted process in your business needs a boring manual fallback that a staff member has actually practised.
Tech Times ↗ -
The argument that heavy AI use is quietly eroding real coding skill
A widely-read post argues that leaning on AI is degrading the expertise of the very people who would otherwise catch its mistakes. For an owner, the practical version is narrower: if your one in-house developer or your outside agency is generating code with AI, you still need somebody who can read it and say no. The saving is real; skipping the review step to bank that saving is where it turns expensive.
Hacker News ↗ -
A new AI assistant impresses testers and alarms them at the same time
Early testers of the assistant Instinct praise what it can do, while flagging its sweeping access, broad terms, and its ability to take actions for you. "Acts on your behalf" is the phrase to slow down on: an assistant that can send, book and pay is a staff member with no notice period and no accountability. For any business holding patient, student or financial records, wait until the terms are plain and the permissions are narrow.
TechCrunch AI ↗ -
Microsoft ships another batch of Copilot features
Copilot Wave 3 is rolling out with a fresh list of capabilities. The coverage is a feature list, not evidence that anyone saved money or time. If you are already paying for Microsoft 365, look at what actually appears in your tenant, pick one task, and measure it for a fortnight before you buy more seats.
Wareham Week ↗ -
US regulator clears a blood test to help evaluate Alzheimer's disease
The FDA has cleared a blood test to aid evaluation for Alzheimer's — an aid to assessment, not a standalone diagnosis. Indian and Canadian approvals run on their own timelines, so this changes nothing in your clinic this week. It is worth knowing anyway, because families read these headlines and will arrive asking your doctors for the test by name.
Hacker News ↗ -
A team writes up why they use no AI at all, and it lands
A flat statement of total abstention from AI drew a large, serious audience rather than mockery, which tells you something about how the mood has shifted. Opting out remains a defensible position, particularly for work that is low-volume and judgement-heavy. Just make sure it is a decision you took with reasons you could explain to a client, not a drift you never revisited.
Hacker News ↗
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.