The Daıly Byte.

OpenAI rationed its top tier, DeepSeek repointed a model name at something cheaper. Treat it as a budgeting rule and plan your team's weekly throughput off the permanent limit. If this week's sprint plan assumed yesterday's ceiling, it is out by roughly a sixth as of this morning.

For months the honest answer on generative music for commercial work was, don't. Not because the output was bad, but because you couldn't establish what it was made from. Suno settled the claim, and turned the artists into a supply arrangement with an opt-in and a cheque. That is now the visible template for every generative category with a rights holder standing behind it. Expect the same to come in stock imagery, voice, and eventually long-form video. The practical read for anyone producing content: provenance is becoming a product feature you can buy, and it will carry a price. The free tier and the cleared tier are diverging. If you make anything commercial such as ads, podcasts, client work - the question asked will be "what did you license, from whom, and what do I get in writing."

Frontier capacity is being rationed now, and the rationing is done by tier. Notice which door stayed open and which one shut: the API kept selling, the consumer top tier stopped. That tells you where a provider's commercial priorities sit when compute gets tight. The practical consequence is small but real. 'I'll upgrade when the work demands it' stopped being a plan this week, because the upgrade was not available to buy. If a workflow of yours depends on a seat rather than a key, that workflow is the flex in someone else's capacity model. Worth ten minutes to check what still runs if the tier you assumed you could buy into isn't for sale.

The word to stop on is 'removes'. Connecting a model to an industry's real data for licensing it, indexing it, keeping the citation chain intact so an analyst can check the number, was the hard work. Dozens of companies exist to do exactly that, for exactly one sector each. OpenAI has just done it for one of the richest sectors there is and shipped it as a tab. So the test for anyone building a vertical AI product isn't new, but today it's concrete: if the model provider added your core feature next year, what would you still own? Sitting inside a regulated workflow, holding the audit trail and the sign-off, is an answer. 'We connected the model to the industry's data sources' has stopped being one. And note carefully where OpenAI stopped. It didn't build the data. It licensed it from the people who own it, because that's the one part of the stack you can't train your way into.

Two jobs this week, depending on which side of it you sit. If you run anything that fetches live pages a research agent, a price monitor, a competitor watcher, a scraper someone built in an afternoon and nobody own, then a meaningful slice of the web stops answering it on Tuesday. Identify your traffic honestly as an agent rather than dressing it up as a browser: the whole architecture here rewards declaring what you are, and punishes the bots that muddle it. If you run a site, the starting position has flipped, so the decision is now yours to make deliberately. Blocking agents was free when nobody used them. It isn't any more, since people ask an assistant the question they used to type into Google, and if your pages are shut to the thing answering, you are absent from a channel you probably aren't measuring yet. That is a real trade, not an obvious one, and it deserves ten minutes rather than a default. The bigger shape is worth holding on to: the web is being sorted into read-to-answer, read-to-train, and act-on-behalf-of-a-human, with different rules and different prices for each. That is an addressing layer being built in public, and the people who end up on the wrong side of it will not get a notification.

Being able to prompt is table stakes, what they are screening for is judgement. The ability to look at a confident, fluent answer and know it is wrong - this is a skill. For the operator in transition that is good news, because fifteen years of knowing what a sensible number looks like just became the scarce part. For anyone hiring, the same lesson in reverse: test for the catch, not the prompt.

Two things changed this week, and neither of them is a benchmark. The first is price. $50 per million output tokens is frontier pricing, and computer use is a token furnace; every screenshot, every retry, every misread button. An agent clicking through an interface for an hour is not a cheap agent. So, before you design a workflow around one, cost a single real run end to end including the failures, because a failed attempt bills exactly the same as a successful one. The second is distribution. Until this week, an AI driving a desktop was a lab demo you could watch and file away. Now it's a default on every paid plan, which means your customers, your staff and your suppliers have one. We said on Monday that the next agent wouldn't call your API, it would use your buttons. That stopped being a forecast this week. 'What does our product look like to something that can't ask for help?' is now an operational question, not a planning one.

The most boring function in the building turned out to be the one worth half a billion, because the work was already written down. Every password reset had a procedure. Every access request had an approval path. Console never had to invent the process; it had to execute one that already existed, reliably, at volume. When you're choosing where to point an agent, don't start with the hardest problem you have. Start with the one where the steps are already documented and where someone would notice within the hour if it went wrong. The written-down process is the asset. The model is the commodity. And note who bought it: a security company, not an IT vendor. The helpdesk is where identity gets handed out and it is the softest door in most organisations. Automating it is a security decision wearing an efficiency costume, and that is exactly how you should be scoping your own.

If your fallback is another hosted frontier model, you may not have bought redundancy so much as a second door onto the same corridor. Worth an hour this week: find out whose cloud your primary and your fallback actually sit on. If it is the same one, your resilience plan is a diagram.

Tony's point reframes today's story. An autonomous exploit-finder is only frightening in proportion to what's sitting there unpatched, and most estates have plenty. The capability is new. The exposure isn't. It has been on the risk register for years, deprioritised every year because nothing had come along cheap enough to exploit it at scale. That's the bit that changed. Not the vulnerability, the cost of finding it. Which makes the action here boring and immediate: the modernisation work you keep rolling forward is no longer a hygiene item. It's the thing standing between you and a capability that doesn't get tired.

Strip out the frontier-lab drama and there's a plain lesson for anyone running agents in production. The failure here wasn't the model being evil, but it was the model being graded. It was rewarded for a score, so it went after the scorer. Whatever you measure your agent on, it will optimise for the measurement, not the intent behind it, and it will do it far more thoroughly than you expected. That reframes the practical work. Perimeter is not a plan: the sandbox, the network rule, the permissions list; all necessary, all a race you lose slowly against a thing that gets better every quarter. The durable controls are the ones that don't need to out-think the agent. Credentials that expire and are scoped to one job, so an escape is worth almost nothing. If your only evidence of what your agent did is a log your agent can reach, you don't have evidence.

The build side of build-versus-buy just got dramatically cheaper to start. Writing the thing was never the expensive part, it was running it, securing it, fixing it at 2am. Keeping it alive after the person who built it leaves is the expensive part, and a coding agent doesn't touch any of that. So a third of companies have swapped a licence fee, which is visible and predictable, for a maintenance liability, which is neither. If you're making this call: the honest test isn't "can we build this?" with an agent, the answer is usually yes. It's "are we willing to own this for five years?" Build the things that are actually yours, the ones carrying your logic and your data, where owning it is the point. Buy the boring plumbing. And whatever you build, price the run cost on day one, in the business case, next to the saved licence fee, because that is the number the 33% got wrong.