Several things happened in AI over the last two weeks that had nothing to do with each other and still rhyme. Anthropic, which makes the Claude models, started paying an outside firm to sit inside its own safety work and check its claims before a model ships, rather than after. Alibaba's Qwen team gave away the weights to a new image model, then wrote a license that stops most companies from actually using it. Two of the free tools people use to run AI models on their own computers, instead of sending the work to someone else's cloud, both shipped upgrades that make that option a little more practical than it was a week ago. Anthropic and OpenAI each cut prices within hours of each other. And a new kind of AI model, built to answer in a fixed choice rather than a paragraph, got tested by outsiders almost as fast as it shipped. None of it changes anything we build today. All of it is worth reading, because each piece says something about which way the ground is moving.
Why we build it this way
We read this kind of news every week for one reason: it tells us whether the industry is drifting toward more control for the buyer, or less. This week, on both fronts that matter to us, it drifted toward more.
A safety evaluator sitting inside Anthropic's own model development, with the access to see what actually happens rather than read a summary of it afterward, is Anthropic saying its own word on safety is not enough by itself — that the check has to come from somewhere the company does not control. That is close to a belief we already build around: a person has to be able to look at the reasoning behind a recommendation, not just trust the label on it.
A maker checking its own work is not the same as someone else checking it, however good the maker's intentions are.
The image model tells a quieter version of the same story. Giving away the weights to a model is not the same as giving away the right to use them, and Qwen's license makes that plain: a business that wants to build on it has to ask permission first. Open weights are a starting point, not a finish line. What actually matters to a buyer is what they are allowed to do with what they have been given, written down.
And the local tools getting better is the unglamorous version of the same trend. It keeps getting a little more practical to run a capable model on a machine a company owns, instead of sending its data out to get an answer back. That is the ground we already build on, and this week it got a little firmer under us.
The same lesson showed up a third way, from an unlikely place. A new kind of model shipped that answers instantly with a number attached saying how sure it is. Within days, outside developers had run their own tests on that number rather than taking the company's word for it, and found it was not as trustworthy as it sounded until someone checked it from outside. That is why every system we build has one point where a person decides, not a number.
What changed this week
- Anthropic is paying for an outside check on its own safety claims. Anthropic and the consulting firm Accenture announced a partnership in which a team from Accenture's safety practice works inside Anthropic, watching how models get built and tested rather than reviewing a report afterward. Both companies say they will each put in at least a billion dollars over five years. Confirmed by both companies, jointly, on 18 September 2026.
- An open, but restricted, image model shipped. Alibaba's Qwen team released the weights for a new model that generates and edits images, including cutting a subject cleanly out of a photograph, and put them up for anyone to download. The license only allows research use — a business that wants to use it commercially has to apply separately. Confirmed by Qwen's own release, 20 September 2026.
- The tools for running models on a company's own hardware got better. Ollama and vLLM, two of the widely used free tools for running AI models on a company's own computers rather than a vendor's cloud, both shipped upgrades this week. One added support for models that can read images as well as text; the other made it faster to answer several people at once from a single machine. Confirmed by both projects' own release notes, 19–23 September 2026.
Nothing here changes a roadmap, and we are not adopting any of it. What we did was read it, check where it came from, and note where it agrees with what we already believe.
New models, and what they now cost
Anthropic and OpenAI each shipped a new model on the same day, 22 September, and both cut prices while they were at it. Anthropic's newest, Claude Opus 5.5, costs meaningfully less to run than the model it replaces. OpenAI answered with two new models of its own, priced to undercut it. Neither move is really about capability — it is about cost: the same kind of work now costs less than it did a week ago, from either company. Two smaller pieces of model news, from companies that build for different markets: DeepSeek released a smaller, cheaper version of its main model that can also read images, and Sarvam — the Indian AI company most likely to sit next to Varahi on a customer's shortlist — released a new model built to read documents in English and twenty-two Indian languages, including handwriting.
What people are actually trying this week
The most argued-about launch of the last two weeks is Jev, from a new company called TypeSafe AI: a model that answers in a fixed choice — yes or no, or a score — with a number attached saying how sure it is, instead of writing a paragraph. Outside developers and researchers immediately built their own tests for it, for free, and published exactly what they did. What they found: the "how sure am I" number is not as trustworthy as it sounds straight out of the box. It runs too confident in some situations and too cautious in others, and the gap widens the further a question sits from anything ordinary. Within days, other outside developers had rebuilt the same trick using ordinary, freely available models running in nothing more than a web browser, no company server involved — proof the idea itself is not locked to one vendor. And on Hugging Face, where people share and download models, this week's most-downloaded item was not a new model at all: it was someone else's aggressive shrinking of an existing open model, small enough now to run on an ordinary laptop. Nobody has independently checked yet how much quality that shrinking costs; the download count only proves people are trying it, not that it works as well as the original.
Explained simply: Jev
A traditional chatbot answers a question by writing a sentence, one likely word at a time. Jev does something different: it looks at a fixed list of allowed answers — yes or no, or a score out of ten — and picks one instantly, the way you would fill in a bubble on a form rather than write an essay. Alongside the answer, it gives a second number: how sure it is. TypeSafe AI, the company that built it, says this makes it far faster and cheaper than a chatbot at jobs that are really just decisions dressed up as questions — approve this, flag that, score this ticket.
The idea is useful exactly where the company's own confidence number holds up, and less useful where it does not. Outside testers who ran their own experiments, for free, and published exactly how they did it, found the same pattern more than once: Jev is genuinely good at telling you which of two answers is more likely right. It is markedly less good, out of the box, at telling you exactly how likely — confident when it should not be for some kinds of questions, too cautious for others, and worse the further a question sits from anything ordinary. A short correction step, using a small number of known-right answers, fixes it. Skip that step, and the confidence number is closer to a guess dressed up as a statistic.
Nothing here changes what we build. But it is a clean example of something we already believe: a system's own stated confidence is not evidence. It has to be checked by someone who did not build it.
None of this moves what we ship. What it tells us is that the two things we bet the company on — a person checking the machine's work, and a customer owning the machine the work runs on — are not eccentric positions this week either. We will keep reading the same sources next week, and if the honest answer is a quiet week, we will say so instead of padding it.
Read about AI Transformation →