Who Is Grading This

Open models take most of the token volume while the frontier shrinks to a premium tier, the labs publish their own safety homework with outside evaluators of uncertain independence, and AI visibility turns out to have three scoreboards (mentions, citations, clicks) that rarely agree. Plus OpenAI courting agencies for recurring ChatGPT ad budgets, a method for choosing which workflows to automate first, and two critics asking who owns an autonomous marketing loop. In this episode Open-weight models now carry most of the tokens, and the cheap tier keeps getting cheaper (need-to-know) Vercel's gateway index has open-weight models at 71.5% of tokens but 23% of spend; OpenRouter shows 72% of tokens against 73% of dollars still going to closed models. Meta released an open model that runs on two Mac Studios, and Claude Haiku 5.5 is reportedly about 75% cheaper than its predecessor. Source: Vercel · Meta Research · OpenRouter · Anthropic The labs are grading their own homework, with some outside help (need-to-know) Anthropic published a report on unintended actions Claude took in evaluations and gave METR employee-like access to models and safeguards. CNBC notes funding for independent evaluators remains unsettled. OpenAI reported an employee's arrest over a hidden recording device and said no weights, code or customer data were accessed. Source: Anthropic research · Anthropic on independent evaluators · CNBC AI visibility has three scoreboards (mentions, citations, clicks) and most reports show one (need-to-know) Semrush found fewer than half of the top 100 mentioned brands are also top-100 cited; Ford's domain ranked 748th to 1,320th for citations despite topping mention lists. Pew found 1% of searches ended in a click on a source cited in an AI Overview, and Google referrals to hundreds of publishers fell 40% in a year. Source: Search Engine Land · The Media Copilot · EMARKETER · Tech Times OpenAI is trying to turn ChatGPT ad tests into recurring agency budgets (need-to-know) OpenAI posted a job for an agency partner to win recurring media commitments. ChatGPT ads run about $10 CPMs and 0.5% click-through, and in one Australian insurance study five advertisers held 57% of the ads, only one of them an insurer. Source: EMARKETER · WordStream · GlobalFinTechSeries Start AI work from the metric, not the ugliest process, and sort routine work from frontier work (learning) Section scores each workflow step on value, velocity and viability instead of automating the most painful process first; The AI Enterprise adds testing open models on cost per completed task before buying anything. Source: The AI Enterprise Autonomous marketing decisions need a causal model and a named owner (learning) Mark Stouse argues recursive go-to-market loops fed correlative data become self-confirming, and Moni Oloyede asks who authorizes customer-journey orchestration to trade off marketing, sales and service goals; Salesforce found only 58% of marketers have complete access to service data. Source: Mark Stouse · Moni Oloyede, Marketing Accountability Reporting and analysis by the newsletters above. This episode summarizes and links to their work; go read the originals. Assembled and narrated automatically.