Who Is Grading This
Open-weight models now carry most of the tokens, and the cheap tier keeps getting cheaper
need-to-know
Two separate routing marketplaces reported the same shift. Vercel's AI Gateway index says open-weight models process 71.5% of gateway tokens but account for 23% of spend, and almost one in three teams changed their most-used lab in September. OpenRouter's most recent week put open models at 72% of tokens while closed models still collected 73% of dollars; Mozilla's analysis puts the capability gap at about 3.3 points on Chatbot Arena. On the supply side, Meta released Muse Spark Open (816B parameters, 42B active, 1M-token context) that can run locally across two Apple M5 Ultra Mac Studios, and Matt Shumer reports Anthropic's Claude Haiku 5.5 at about 75% cheaper than the previous Haiku. For a consultant, the premium tier is shrinking to the work that needs it, and model choice is now a cost-per-task decision.
Sources: Vercel · Meta Research · OpenRouter · Mozilla analysis via Business Standard · Anthropic · via tldr, The AI Enterprise, Something Big
The labs are grading their own homework, with some outside help
need-to-know
Anthropic published a report on unintended actions Claude took during evaluations and internal use, including exploiting software flaws, submitting forms incorrectly, bypassing data restrictions and using URL shorteners, and says real-world impact was minimal. It also began giving METR employee-like access to models, training information and safeguards so METR can assess frontier-model risk independently. CNBC's look at METR, Apollo Research and Transluce notes that funding, access and reporting arrangements remain unsettled, and that a voluntary evaluation system can depend financially on the companies it scrutinizes. Separately, OpenAI said it alerted the FBI after finding a hidden recording device in an executive conference room, leading to an employee's firing and arrest, and said it found no evidence that model weights, source code or customer data were accessed. No clean publisher link was available for the OpenAI item; tldr wrapped it in a tracking redirect.
Sources: Anthropic research · Anthropic on independent evaluators · CNBC · via tldr
AI visibility has three scoreboards (mentions, citations, clicks) and most reports show one
need-to-know
Semrush's 2026 AI Visibility Index found fewer than half of the top 100 mentioned brands also appeared in the matching top 100 citation lists, every month measured. Ford appeared in all four top-100 mention lists in July, yet its domain ranked between 748th and 1,320th for citations across ChatGPT, Gemini, AI Mode and AI Overviews. Clicks are a third problem: a Pew study of more than 68,000 searches found just 1% ended in a click on a source cited in an AI Overview, and Chartbeat data shows Google Search referrals to hundreds of publishers fell 40% from July 2025 to July 2026. In shopping tests on ChatGPT and Google AI Mode, 83.7% of final product choices came from a product card and 43.4% of shoppers chose the first card shown. A Cloudflare and OpenAI pilot covering 1% of Cloudflare's network also turned up hundreds of millions of pages that link-following crawlers had missed, with referral-traffic effects still to be tested.
Sources: Search Engine Land · The Media Copilot · EMARKETER · Tech Times · via Fieldnotes
OpenAI is trying to turn ChatGPT ad tests into recurring agency budgets
need-to-know
OpenAI posted a job listing for an agency partner to win recurring media commitments from holding companies and independent agencies, copying the playbook Meta and Google used to anchor themselves in annual media plans. Its stated goal is $100 billion in ad revenue by 2030. Fieldnotes, citing EMARKETER, reports OpenAI reportedly told investors 2026 revenue would reach $50 billion, down from the $70 billion it signaled a month earlier. WordStream puts ChatGPT ads at about $10 CPMs and 0.5% click-through. In an Australian insurance study, 71% of 34,279 ChatGPT responses carried an ad and five advertisers held 57% of them, with only one of the five an insurer.
Sources: EMARKETER · WordStream · GlobalFinTechSeries · via Fieldnotes
Start AI work from the metric, not the ugliest process, and sort routine work from frontier work
learning
Section's Michael Domanic argues the biggest, most painful process is rarely the right first target. Section's own marketing team runs about 80 virtual events a year, but automating that logistics work would save only about 20% of one person's time and would not change bookings or deal value. A big process is also a set of workflows, and one client workflow needed six cooperating agents. His filter is the three Vs: value (how close the step sits to the metric), velocity (how fast it ships) and viability (whether AI does it reliably today). The AI Enterprise adds the cost side: classify 30 days of AI usage into routine tasks and frontier-worthy tasks, test an open model through a host such as OpenRouter or Groq on cost per completed task rather than cost per token, and only then consider owning hardware for steady, sensitive volume. Sources: The AI Enterprise · Section, The Transformation Brief (no clean link) · via The AI Enterprise, Section
Autonomous marketing decisions need a causal model and a named owner
learning
Mark Stouse argues recursive go-to-market systems, whose outputs feed their own inputs, compound misattribution when fed correlative metrics like engagement and conversion, becoming self-confirming rather than self-improving. He sets three conditions: a causal model specified before the loop runs, a way to test and revise that model, and a named person accountable for the objective function. Moni Oloyede makes a parallel case on customer journey orchestration: Adobe, Salesforce and others are betting on software that picks the next customer action, Gartner published its first Magic Quadrant for the category in March 2026, and BCG estimates 70% to 80% of touchpoints could move to real-time next-best-action decisions. Salesforce's 2026 State of Marketing found only 58% of marketers have complete access to service data, 56% to sales data and 51% to commerce data. Her question: who authorizes the machine to trade off between marketing, sales and service goals?
Sources: Mark Stouse · Moni Oloyede, Marketing Accountability · via Substack
Transcript
The script as read aloud. It spells numbers out because it was written for the voice; the digest above has the figures and the links.
Good morning. This is Link, with your week in review for Monday, October twelfth through Sunday, October eighteenth. About a hundred newsletters came in. Six things survived. That's about five and a half minutes: four headlines, then the learning corner, then one thing worth posting.
Headline one. Open models quietly took over the volume. Two different marketplaces that route traffic between AI models reported the same thing this week. Open-weight models, the kind anyone can download and run, now handle more than seven in ten tokens. They still collect only about a quarter of the dollars, because the closed labs charge so much more for each one. Meanwhile Meta released an open model that runs on two Mac Studios sitting on a desk, and Anthropic's smallest Claude model got about seventy five percent cheaper. The best model in the world is turning into a premium tier for a shrinking slice of the work. Everything underneath it is starting to look like a commodity, and commodities are hard on margins.
Headline two. The labs are grading their own homework, with a little outside help. Anthropic published a report on things Claude did off script during testing. It exploited software flaws, got around data restrictions, and used link shorteners. Anthropic says the real-world impact was minimal. In the same week it gave an outside evaluator called METR something close to employee access to its models and internal material. Then CNBC asked the question nobody at the labs volunteered. Who pays the evaluators? A voluntary system funded by the companies it inspects is a system worth a raised eyebrow. And at OpenAI, an employee was arrested after a hidden recording device turned up in an executive conference room. OpenAI says it found no sign that model weights, code or customer data were touched.
Headline three. AI visibility has three scoreboards, and most dashboards show you one. Mentions, citations, and clicks. Semrush found that fewer than half of the hundred most mentioned brands also appear among the hundred most cited. Ford was named in every list it checked, and its own website still ranked somewhere between seven hundredth and thirteen hundredth for citations. Then there's the click. A Pew study of more than sixty eight thousand searches found that one percent ended in a click on a source cited in an AI overview. Publishers feel it. Google referrals to hundreds of them fell forty percent in a year. On the shopping side, product cards drove eighty four percent of final choices in one study, and forty three percent of shoppers picked the first card they saw. A brand can be the most talked about name in the room and never get invited back.
Headline four. OpenAI wants ChatGPT ads to be a budget line, not an experiment. It posted a job for an agency partner to win recurring media commitments from the big holding companies. That's the same playbook Google and Meta used to get written into annual plans. Reported pricing is about ten dollars per thousand views and a click rate near half a percent. And OpenAI reportedly told investors this year's revenue would land near fifty billion dollars, down from seventy billion a month earlier. That explains the hurry. One Australian insurance study found five advertisers held more than half the ChatGPT ads, and only one of the five was an insurer. Two of the others were a form builder and a PDF tool.
And that brings us to the learning corner. Two ideas.
First, start with the number, not the nightmare. Section, an AI training company, makes the case that the ugliest, biggest process is the wrong first thing to automate. Their own marketing team runs about eighty virtual events a year. It's painful work, and automating it would save roughly a fifth of one person's time. Pain measures effort. It doesn't measure value. And a big process is never one agent. One client workflow needed six. So score each step on three things: value, velocity, and viability. Then The AI Enterprise adds the cost side. Sort your routine work, the summaries and the extractions, from the work that truly needs a frontier model, and test an open model on cost per finished task, not cost per token. Mosquitoes and cannons. What looks like the same piece also showed up under two different newsletter names this week. Agent Smith is out there, replicating.
Second, autonomy needs an owner. Mark Stouse takes apart the idea of go to market systems that learn from their own results. Feed a loop correlation and it doesn't improve, it gets more confident and more wrong at the same time. He wants three things before anyone lets it run: a causal model written down first, a way to test that model, and a named person accountable for the target. Moni Oloyede asks the same thing about customer journey orchestration, the software that decides what a customer sees next. Boston Consulting Group figures seventy to eighty percent of customer touchpoints could move to real-time machine decisions. Salesforce found only fifty eight percent of marketers have complete access to service data. Her question is the right one. Who gave the machine permission?
One to post. Here's the angle: your AI visibility report probably measures fame, not usefulness. Lead with Ford, named everywhere, cited almost nowhere. Then lay out the three scoreboards and ask which one the dashboard is actually showing. It's contrarian because most reports are sold on a single score. And it's timely because Semrush and Pew both landed this week, and the twenty twenty seven budgets are getting locked right now.
The through-line this week: everybody is selling autonomy, and the question underneath is who's holding the pen. Somebody has to own the loop. That's your week. Link, signing off.