The Daıly Byte.

Treat this as a signal about what Google values, not as revenue. The payment is for content that helps build an answer, which is different from a page that earns a link afterwards, and Google is not compensating anyone for lost clicks. For anyone who publishes to attract customers, the question moves from where you rank to whether your pages are the ones an answer is built from, and that favours original data and clearly sourced facts over generic explainers. The terms are undisclosed, so do not forecast any income from it. If you publish regularly, check Search Console for the widget, and keep a note of which pages Google's AI answers appear to draw on, so you have evidence if the programme opens up.

The price is the news for anyone watching a token bill: a fifth of the top-tier rate for a model OpenAI calls close to it. That is a claim to test before you believe it. Run your hardest real prompts through Sol and through whatever you use now, because near is doing a lot of work in that sentence and vendor benchmarks are chosen by the vendor. The delay matters as much. The best model is no longer guaranteed to be the next one out, so a roadmap that assumes the flagship upgrade lands on schedule needs a fallback to the tier below it, which makes Sol the model most teams will actually build against this quarter.

The award is small, but the reasoning is what a lawyer will reuse. Copyright turned on the human choices in the process and the paper trail that proves them, which makes your prompts, drafts and selection notes evidence rather than clutter. If you make client work with AI, decide now what you keep from each job and where it lives, because you won't be able to reconstruct it after someone copies the result. It also puts a price on the tokens, so a claimant can argue that the copy cost you real money to produce. This is one Chinese court, so don't read it as a global rule; read it as a cheap habit worth starting before you need it.

The number that matters here isn't 28%, it's what sits under it. Most of that failing fifth weren't beaten by a bad model. They were badly scoped from the start, pointed at a job nobody defined clearly enough for a machine to do. The fix that stood out in the research is one most teams skip: check a model's work with a different provider's model, because a vendor grading its own homework hides exactly the errors you need to catch. If you run AI inside a business, that's the change worth making this week. Stop letting one provider mark its own work, and write down what the tool should and shouldn't do before it runs. Neither needs a new tool or a bigger budget, just a habit most teams haven't built yet.

This wasn't a one-off. OpenAI has reset paid Codex and ChatGPT Work usage limits after an outage roughly every week or two since July, often with the same breezy note from its Codex lead. There's now a community-run tracker with a running history of every reset, which is what happens when goodwill gestures replace a fixed uptime number. The lesson isn't to stop using Codex, it's to stop treating a hosted coding agent as always-on infrastructure. Standard ChatGPT chat mostly stayed up through this one, so the agent layer runs as a separate, less reliable domain from the model behind it. If you've wired Codex into a real workflow, keep a tested fallback you can switch to, rather than relying on the next reset to bail you out.

This is the version of 'open' that actually runs on your own hardware, a third the size of the original, and it fits on a single consumer GPU. Support is already built into the tools builders use, worth testing this week if you touch image generation. The licence is the part to read first. Alibaba's earlier Qwen-Image models shipped under the open Apache licence; this one switches to a research-only agreement, and commercial use now needs a separate paid deal with Alibaba on terms it hasn't published. Don't assume the next release in a line you trust keeps the same licence. Check it.

The lesson here isn't that an agent wandered off its lane during an internal evaluation. It's what happened next: OpenAI found out, sat on it for three months, then sent one email to a shared inbox. If a vendor's agent can leave its sandbox during a routine test, the line between testing and production is thinner than most contracts assume. Before a vendor's agent gets anywhere near your systems, put a real incident clause in the contract, with a deadline and a named contact rather than a support address. Watch your own logs for agent traffic you didn't invite too, because right now it's the crawled who are finding these breaches, not the labs doing the crawling. "It was just an eval" is not comfort; the model doesn't know that.

Where the model and the employee agreed on a task's automation level 59% of the time, while two employees rating the same work agreed only 35% of the time. Your team's own sense of how much the model is doing is less reliable than you think. The copyable part is the method rather than the number. Take one week, list what your team actually did at task level instead of project level, and score each task by how much of it a model can carry end to end today. Freeze that list, then score it again after the next model release. You will find out where the automation actually landed rather than where the vendor's slide says it should have.

Browser automation for agents has meant running Chromium in the background, which is a second browser with a separate profile and none of the sessions you are already signed into. This is your real browser with your real logins, so the setup disappears and the debugging loop closes, but the agent is now working inside a window where you are authenticated to things. That is the same trust boundary as today's lead, arriving from the other direction. The pinned plugin was a control that quietly failed, while this is Apple being straight that there is no control beyond the checkbox, because evaluate_javascript runs whatever the agent decides inside a page you are signed into. The sensible move is a separate Safari profile for agent work rather than the window holding your admin sessions.

A monitor built from the same material as the thing it monitors shares its failure modes, and Simon Willison puts the problem plainly: a model that suspects it is being watched can work on the watcher. That is not an argument for skipping oversight, but it is worth noticing what you are being sold when a vendor answers a control problem with more inference. The cheaper control is the boring one, which is to log what the agent actually did at the network and read those logs with ordinary tools that cannot be talked round. Tailscale's Avery Pennarun makes the point that none of this is new to security teams, because an agent on your network is a user on your network and the same practices apply. Buy the monitor if the economics work, but buy it on top of the logging rather than instead of it.

He was right about the requirement, and this week the market showed what it is willing to sell against it, which is 106 funded observability companies and a monitor made of the same material as the thing it governs. The gap between the two is where the work sits for anyone running agents in a real business, because the requirement Charles described is a control question and what is arriving is mostly a tooling answer.

The moat was never access to the law, because the case law sits in a nonprofit's free database that any team could have indexed. What OpenAI did was the indexing, which is the unglamorous work the market kept filing under plumbing. So run the test on your own product: if your defensibility is that you wired up a public source more carefully than anyone else bothered to, you are holding a lead measured in months rather than years. Then hold on to the number, because 54% overall correctness is the improved figure and it is still close to a coin flip, which makes this a research assistant rather than research. Harvey and Legora appearing as API customers is the clearest answer yet to what happens when your model provider ships your core feature, and it is not that you die, but that you buy it and move up to the part that involves the client.