Today's briefing
The cheap end of AI got cheaper — and noisier
A small Claude model that reads a million tokens for ten cents changes one kind of work, and almost nothing else. Plus eight things worth knowing.
The one that matters
A small Claude model now reads a million tokens for ten cents — what that actually buys you
What happened
Anthropic released Claude Haiku 5.5, the small and cheap end of its line-up. Two numbers matter. It accepts up to a million tokens of input in one go — call it several thousand pages at once — and input is priced at ten cents per million tokens. For scale, a million tokens is roughly 700,000 words: about two years of a small company's email, or a tall filing cabinet.
It landed in a busy week of model releases, and a lot of the coverage framed it through Anthropic's planned stock listing. Ignore that framing. A listing does not change your invoice. The price per page does.
What it actually means for your business
Cheap input pricing changes the economics of exactly one thing: a machine reading large piles of text you already own. Not deciding, not writing to customers — reading, sorting, extracting, summarising. That is the task that used to be too expensive to bother with.
- A 600-student school holding three years of fee receipts, leave applications and parent emails — say 40,000 pages — can have the lot read and indexed for a few dollars of model cost. Not thousands. A few.
- A 30-bed clinic with 12,000 pages of discharge summaries and lab reports, roughly 8 million tokens, pays about 80 cents — under 70 rupees — to have every page read once.
- A workshop with 40 staff can push two years of supplier invoices and job cards through a model to find which jobs actually lost money, instead of putting a clerk on it for three weeks.
A year ago you would have thought twice about those jobs because of the per-page cost. Now the model fee is close to a rounding error. That is a genuine shift, and it is a narrow one.
What it does not mean, and who is overstating it
The price of the model is not the price of the project. On real deployments, the model is usually the cheapest line item; the money goes into getting documents out of WhatsApp, scans and Tally, defining what a correct answer looks like, and having a human check the output until you trust it. Budget on that basis, not on ten cents.
A million-token window is capacity, not reliability. You can put 4,000 pages in front of a model; that does not guarantee it notices the one fee waiver on page 2,310. Treat it as a fast, cheap, occasionally careless reader, not a search oracle. And note that only the input price is cheap here — what the model writes back is billed separately, so tasks that generate long reports cost more than tasks that answer in a line.
Two groups are overstating this. Resellers who will quote you a monthly retainer to put your whole business into AI — the underlying model now costs them almost nothing, so ask what you are actually paying for. And anyone citing the context window as evidence of accuracy. It is evidence of appetite.
What a sensible owner does this month
- Pick one high-volume, low-risk reading task you currently pay people to do. Invoice matching, admission-form data entry, insurance claim triage.
- Take 100 real documents — your messy ones, not clean samples — and run them twice. Count the errors yourself.
- If the error rate is under what your current clerk makes, you have a project. If not, you have a demo, and that is fine: wait.
- Ask any vendor for a cost per document processed, not a cost per seat per month. Make them show their working.
- Keep a human signature on anything that touches money, medicine or a legal deadline. Cheap reading does not change who is liable.
If you do not have a pile-of-documents problem, the honest answer is that this changes nothing for you this month.
Also worth knowing
-
Anthropic rewrites its usage rules — the part that matters is not the polite-to-Claude bit
The headlines are about a new ban on sustained, needless abuse of the model, which is what everyone is arguing about. The part that affects you is further down: tightened rules covering health and financial uses, surveillance, weapons software and election work. If you are building anything that gives medical or lending guidance, read the actual policy before you build on it, because your vendor's terms are now a constraint on your product.
The Verge AI ↗ -
OpenAI's revenue is running about $20B below what it had signalled
Heavily discussed, and it matters for a dull reason: the prices and free tiers you plan around are downstream of how much money these companies are actually making. Avoid multi-year commitments that assume today's pricing, and make sure your data and prompts can be moved to another provider without a rebuild.
Hacker News ↗ -
Microsoft's fall event: new Surface hardware and Copilot running locally
The notable thread is AI that runs on the laptop rather than in a data centre, which eventually helps anyone nervous about sending client files off-premises. But nothing here justifies pulling forward a hardware refresh — if your machines are fine, they are still fine. Revisit it when your normal replacement cycle comes due.
PCMag UK ↗ -
A legal software firm cut its AI coding bill 65% by matching cheap models to easy tasks
This is a vendor-published case study, so discount the enthusiasm, but the method is sound and transferable: use the cheapest model that passes your own test for each task, and reserve the expensive one for the hard cases. If you or your dev team are on a flat diet of the most capable model, you are almost certainly overpaying by half.
OpenAI ↗ -
OpenAI shut down two influence operations built on fake journalists and a fake think tank
The specific operations are geopolitical, not commercial, but the capability leaks downward fast. Expect better-dressed fake credentials in your inbox — a convincing reporter requesting comment, a plausible consultancy requesting data. Confirm any payment change or media request through a phone number you already had, not one in the email.
OpenAI ↗ -
Hexaware becomes an Anthropic preferred partner, so expect the pitches to start
Relevant mainly because large IT services firms will now actively sell Claude-based builds into Indian mid-market businesses. A partner badge is a commercial arrangement, not a quality guarantee. If one comes calling, ask for a fixed-price pilot on one real workflow with a measurable pass mark before you discuss anything bigger.
PR Newswire ↗ -
A new open speech recognition model, Falcon ASR, is out
Speech-to-text keeps getting cheaper and more openly available, which is the thread to watch if you want front-desk calls, site instructions or consultation notes transcribed. The announcement does not give accuracy figures or language coverage, so assume nothing about how it handles your staff's accents or code-switching until you test it on your own recordings.
Hugging Face ↗ -
Fired OpenAI safety researchers publicly dispute the misconduct claims against them
Internal politics at a supplier, not a product change, so nothing to act on today. It is a useful reminder that a vendor's stated safety culture is a marketing position, not a contract. Whatever limits matter to you — data retention, what the system may never say to a customer — belong in your own written agreement.
TechCrunch AI ↗
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.