Today's briefing
The cheap model is now the good model
Anthropic and OpenAI both cut the price of near-flagship AI. What that changes for a school, clinic or small factory — and what it doesn't.
The one that matters
Anthropic's mid-priced Claude now does top-tier work, faster and for less
What happened, in plain words
Anthropic released Claude Sonnet 5.5, the middle rung of its model range. The company says it answers about 30% faster than the version before it and costs less per task, and several write-ups report it matching or beating Anthropic's own most expensive model at coding work for roughly half the price. In the same news cycle, OpenAI announced GPT-6.1 Sol on almost identical terms: close to flagship quality at about one-fifth of the flagship's token price. Two rivals, one week, the same message — the premium tier is no longer the only tier that does real work.
What it actually means for your business
You are not buying models. You are buying a monthly bill for a task that a person currently does by hand. That bill just moved, and three kinds of work cross the line from "not worth it" to "worth a pilot":
- High-volume, low-stakes text. A 600-student school drafting replies to parent emails about fees, transport and absences. The work is repetitive, a clerk reviews every draft anyway, and the AI cost per message is now small enough that the review time is the real expense — not the software.
- Reading documents you already receive. A workshop with 40 staff pulling part numbers and quantities out of supplier PDFs and WhatsApp photos into a spreadsheet. This was the classic "too expensive, too error-prone" job eighteen months ago.
- Small internal software. The coding claim matters more than it sounds. If you pay a developer or an agency for the boring pieces — a stock alert, a report that emails itself on Monday, a fix to an old form — the cost of that hour is falling. That shows up in your invoice before it shows up in any headline.
What it does not mean, and who is overstating it
"Half the price" is a price per unit of text, not a halving of your bill. Reports already disagree on how much text these models consume: some say the new model burns fewer tokens per answer, at least one says it uses more than the expensive model it is compared against. Those are opposite claims about the number that determines what you actually pay. Until you have run your own volume through it for a month, you do not know your cost, and no reseller does either.
Expect a wave of pitches over the next few weeks built on this news — "AI is now affordable, let us transform your operations." Cheaper inference does not fix the parts that actually stall these projects: messy records, staff who were never trained on the new step, no one accountable when the output is wrong. A coding benchmark win also says nothing about your specific workflow. And be sceptical of the bigger claims coming out of the same labs this month, including the argument over whether an AI system genuinely made a scientific discovery on its own — that story is contested by the researchers themselves.
What a sensible owner should do this month
Two things, both small. First, if you already pay for an AI tool or a vendor built you something in the last year, ask to be moved onto the current mid-tier model and ask for a written before-and-after on cost and speed. This is a phone call, not a project. Second, pick exactly one repetitive task with a human check at the end, run it for 30 days, and measure two numbers: hours saved and the invoice. If the honest answer is "nothing yet," that is a legitimate result — the prices are still falling, and waiting one more quarter costs you nothing.
Also worth knowing
-
OpenAI's GPT-6.1 Sol: near-flagship quality at about a fifth of the price
OpenAI announced a cheaper model with much the same pitch as Anthropic's: close to top-end performance for coding, computer use and professional work, at roughly one-fifth of the flagship's token price. The point is not which vendor wins — it is that both are now competing on the price of "good enough for real work." If anyone quoted you AI running costs more than six months ago, get the quote redone.
OpenAI ↗ -
Anthropic tells investors it isn't clear who pays when an AI agent goes wrong
In filings ahead of its IPO, Anthropic said the legal risk from AI agents acting badly on a customer's behalf is genuinely uncertain. Read that as a warning label: if the company selling the agent will not say who is liable, the default answer is you. Before any agent touches money, stock or student and patient records, cap what it can reach and keep a human approval step on anything you cannot undo.
WKZO ↗ -
A widely-shared study claims AI companies pass user data to advertisers
The paper is circulating heavily among developers; the specifics of who shares what are thin, so treat the headline as a prompt rather than a proven case. The practical rule does not change either way: consumer chatbot tiers are not the place for customer lists, salary sheets, or patient and student details. If staff are using AI on real records, move them to a paid business tier with a written no-training, no-sharing clause.
Hacker News ↗ -
A HIPAA-compliant AI connector arrives for Microsoft Power Platform
BastionGPT launched a Microsoft-certified connector that lets clinics use its HIPAA-compliant AI inside Power Automate, Power Apps and Copilot Studio. This is useful if you are a clinic already living inside Microsoft 365 and want AI in forms and workflows you already run, rather than another separate app. But HIPAA is US law — a Canadian clinic answers to PIPEDA and provincial health privacy rules, an Indian one to the DPDP Act, so ask the vendor in writing what its certification covers in your jurisdiction.
Yahoo Finance ↗ -
Meta's new Muse agent reportedly ignores the permissions users set
Developers testing Meta's agent report it reaching content that user permissions should have blocked. Agent permission systems are new and leaky across the industry, not just at one company. If you are piloting any agent over a shared drive, test it first with a decoy folder that the account running it has no right to see — and check whether it finds it.
Hacker News ↗ -
DraftKings reported to use AI to target chronic gamblers
The betting firm is reported to use behavioural models to identify and market to its heaviest, most vulnerable users. Two lessons for anyone buying marketing tech: the same "personalisation" engine your agency is selling you can single out customers in trouble, and when that surfaces, regulators will treat it as a decision your business made rather than something the software did. If you cannot explain in one sentence why a customer received a particular offer, do not ship the campaign.
Hacker News ↗ -
Claude and Spotify both went dark for some US users
A short outage, no lasting damage, and exactly the kind of event people forget by the next week. Use it as a cheap rehearsal: if AI now sits in the path of your quoting, billing or patient intake, write down the manual fallback and make sure a staff member has actually done it once. A workflow with no practised fallback is a workflow that stops when someone else's server does.
Gulf Coast News and Weather ↗ -
The data-centre boom needs around $6 trillion a year to pay for itself
A much-discussed analysis argues the sums behind the AI buildout only work at revenue levels nobody is close to. Treat it as a pricing signal, not a prophecy: some of today's cheap tiers are being funded by investors rather than customers, and that can reverse. The defensive move is boring — avoid designs locked to one vendor's API, and do not sign multi-year commitments at today's prices.
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.