Four things worth your attention today, and the one everybody covered wrong is the Google one. I run three WordPress properties on this stack, so I read the Preferred Sources launch the way a publisher reads it, not the way a press release reads it. Here is the operator’s version.
Google Gave Publishers a Button, Not Their Traffic Back
On Thursday Google released an embeddable “Preferred Sources” button, and TechCrunch’s writeup framed it as a new way for publishers to fight AI-driven traffic losses. Every outlet that picked it up stopped at “Google shipped a button.” Nobody read the docs.
So I did. The Search Central documentation is where the actual story is. The standard install is genuinely two lines: a script tag pointing at news.google.com/swg/js/v1/publisher.js and an empty where you want it to render. There is an ES Module path if you want to control theme and language yourself, and a plain deeplink to google.com/preferences/source?q= if you would rather not load Google’s JavaScript at all. That last option is the one most small publishers should take, and none of the coverage mentioned it exists.
Now the parts that matter more than the install. Eligibility is domain and subdomain level only, so you cannot nominate /blog or a single section. Your site has to already appear in Google’s source preferences tool before the button does anything, which means the button is a conversion device for demand that already exists, not a discovery mechanism. And Google’s own documentation frames all of this as “examples of how you can build your audience,” with no ranking promise attached anywhere on the page.
Google says people are twice as likely to click through to a preferred source when one is available, and that readers have now selected more than 600,000 unique sources. Both numbers are Google’s, measured by Google, about a Google product. Treat them accordingly. The mechanism underneath is also worth naming plainly: a preferred source shows up more often in Top Stories and gets a badge in AI Mode and AI Overviews. A badge inside an AI Overview is still an AI Overview. The reader still got their answer without leaving Google.
Verdict: marketing. This does not fix the thing it is being sold as fixing. Zero-click is a structural change to how Search works, and a button that asks your existing loyal readers to manually opt into seeing you more is not a countermeasure to it, it is a loyalty program. That said, install it anyway. It costs two lines, it is free, and converting your existing audience into a persistent preference is a real if small win. Just do not put it in a deck as your AI-search strategy.
OpenAI Cut Sol to $4 In and $20 Out, and the Output Side Is the Real Story
OpenAI dropped developer pricing on GPT-5.6 Sol on Friday. Reuters reported the move as a cut of more than 20%, taking input from $5 to $4 per million tokens and output from $30 to $20. It applies to the API and to credits on ChatGPT Work and Codex. Pro, Plus and Business subscriptions are untouched. It runs three months.
The headline number undersells it for the people who will feel it most. Input fell 20%. Output fell 33%. If you are running chat, that gap barely registers. If you are running agents, output is where your bill actually lives, because an agent loop generates far more than it consumes: plans, tool calls, retries, self-corrections, the whole recursive mess. A third off output is a materially bigger deal than “more than 20%” suggests, and it is aimed squarely at the workload OpenAI most wants to take from Anthropic.
Which it now does on price. Reuters put Claude Opus 5 at $5 in and $25 out, and Claude Fable 5 at $10 and $50. At $4 and $20, Sol undercuts Opus 5 on both sides of the meter. That is the first time OpenAI’s most expensive frontier model has been the cheaper option against Anthropic’s comparable tier, and it is not subtle positioning.
Two caveats before you rewrite your config. First, this is a promotion, not a price. It expires in three months, and a promotional rate that lapses in November is a terrible reason to re-architect anything in August. Benchmark against it, do not budget against it. Second, understand why OpenAI can afford the move: CNBC reported that CFO Sarah Friar told investors on August 14 that enterprise revenue had overtaken the ChatGPT consumer business for the first time, at roughly $40 billion in annualized recurring revenue. A company whose growth now comes from business contracts can rationally sell frontier inference near cost to lock in the agent and coding workloads those contracts run on. You are not being given a discount. You are being bid for.
Verdict: matters. Real money, real margin, and worth a fresh cost model. Just price the November cliff into whatever you decide.
Anthropic Turned Three Sampling Parameters Into Hard Errors
This one is not an announcement. It is a line in a docs table, which is exactly why it will bite people.
I was reading Anthropic’s model deprecation page this morning to check something unrelated, and the parameter section has teeth now:
temperature,top_pandtop_kare deprecated on Claude Opus 4.7 and later, and on Mythos Preview.- Setting any of them to a non-default value returns a 400 error. Not a warning, not a silent ignore. A failed request.
- The Python SDK from v1.0 onward removes them from the request types entirely, so passing one raises a
TypeErrorbefore you ever reach the network. - Anthropic’s guidance is to drop them and steer behavior through prompting instead.
If you have a wrapper written against Claude 3 or 4-era models with temperature=0.2 baked in as a default, that wrapper does not degrade gracefully when you point it at a current model. It stops. And because the SDK failure mode and the API failure mode are different, you get a TypeError in one environment and a 400 in another, which is a genuinely annoying afternoon.
Same page, second item: claude-mythos-preview is now deprecated with a migration path to claude-mythos-5. Worth flagging because Mythos Preview is one of the two models in the protein-design results we covered on Thursday. The model that just produced a lab-validated 40% hit rate against RBX1 is already on the deprecation list. That is the tempo now, and if your pinned model id is a constant in a config file rather than something you review on a schedule, the tempo will eventually run you over. While you are in there: claude-sonnet-4-5-20250929 carries a tentative retirement of no sooner than September 29 this year, and claude-opus-4-5-20251101 no sooner than November 24.
Verdict: breaks your stack. Go grep your repo for temperature today. It is a ten minute job that turns into a production incident if you skip it.
Pax Silica Turns Your Model Choice Into a Policy Bet
Slower fuse, bigger blast radius. Reuters obtained a draft US State Department letter to the 35 countries that signed Washington’s AI Opportunity Statement, warning that membership in the US-led Pax Silica framework cannot be combined with what the draft calls “duplicative initiatives whose expectations conflict with our own.” CNBC’s summary of the Reuters reporting is the cleanest short version.
“To be part of everything is to be part of nothing.” The draft letter urges partner countries to “choose deliberately” on AI, and never names China once.
Pax Silica launched in December 2025 to secure supply chains for semiconductors and critical minerals. Roughly two dozen countries have joined, including Japan, Australia and South Korea, and also Kazakhstan, which has joined China’s competing framework as well. Reuters could not establish when the letter goes out, or whether the text survives review intact. So this is developing, and I am not going to pretend it is settled.
Here is why a builder should care anyway. This roundup has spent weeks recommending Chinese open-weight models on price and capability. GLM, Qwen and DeepSeek are on real production stacks right now, at real US companies, for the obvious reason that they are cheaper. Separately, US officials have been weighing Entity List designations, federal procurement bans and liability rules aimed at exactly those labs. Put the two together and your model selection stops being a purely technical decision. It becomes a bet on a policy environment you do not control and cannot forecast.
I am not telling you to rip out a Chinese open-weight model. Weights already downloaded are weights you already have, and the enforcement question there is genuinely murky. I am telling you that “we standardized on GLM because it benchmarked well and cost a third as much” is now a sentence with a political dependency inside it, and the cost of an abstraction layer that lets you swap providers has never looked cheaper than it does this week. If you were ever going to build that layer, build it before someone else’s letter forces the timeline. Related context on how fast that particular ground moves: Z.ai held GLM-5.3’s open weights back over its own security findings barely a week ago.
Verdict: breaks your stack, eventually. Nothing to do this morning. Something to design for this quarter.
What I’d Do With Today
Install the Google button because it is free, and then forget about it, because it is not a strategy. Rebuild your inference cost model against Sol at $4 and $20 while the promo is live, but keep the November expiry in the spreadsheet where you can see it. And grep for temperature before Monday. Of everything above, that is the only one that will break something you already shipped.