Latest / UpNext AI / Anthropic’s AI Spending Reports, Microsoft’s OpenAI Revenue, and Google’s Search Ads
Transcript
- Welcome to the UpNext AI podcast. It's Thursday, October 29th, 2026, and here's what matters in AI today. Anthropic has introduced reporting tools for enterprise customers who want to understand what they get from spending on its models. TechCrunch reports that the tools combine internal analytics with employee surveys. Managers can examine a team’s work over a selected period, see what it produced and what it cost, and identify where workflows ran into trouble. The reports also offer recommendations for improving results. Anthropic is adding pooled group budgets so teams can manage token spending together. That gives buyers more to work with than a usage total. A team could see whether a costly workflow is producing useful output, or whether a recurring obstacle needs attention before it spends more. It also gives Anthropic a way to help customers refine their workflows through the product itself. The timing matters. TechCrunch cites a Jellyfish report on agentic coding tools used by 276,000 customers. In that dataset, spending on tools built on Anthropic models had grown tenfold since the beginning of the year, and their cost was about two percent of total engineering cost. Those figures describe spending in Jellyfish’s customer data; they do not establish a return on investment. Anthropic’s reports are designed to help customers make that assessment within their own teams. The company says the report data will not be used for model training by default. That question of where AI money goes leads to a remarkable number from Microsoft. In a regulatory filing, Microsoft said it generated $10 billion in revenue from commercial arrangements with OpenAI during the quarter ended September 30th. TechCrunch reports that the arrangements include revenue-sharing payments. The amount was nearly eleven percent of Microsoft’s total quarterly revenue of $92.3 billion. Microsoft also recorded $24.1 billion in revenue from OpenAI-related arrangements over its full 2026 fiscal year. The quarterly figure is revenue to Microsoft, not a measure of OpenAI’s total sales. It makes clear how financially important this partnership remains even as the companies compete for some of the same customers. There is an important distinction inside that headline number: Microsoft has not specified how much came from revenue sharing and how much came from Azure cloud services. It separately reported $7.2 billion in outstanding receivables from OpenAI at the end of September. Those are amounts owed for previous transactions, not additional revenue. For anyone tracking the economics of frontier AI, the filing shows the scale of the relationship without breaking out which commercial arrangement contributed what. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. Google, meanwhile, is extending ads in its AI-powered Search experiences to international markets. TechCrunch reports that the announcement came with Alphabet’s third-quarter earnings. Google said ads in AI Mode are expanding internationally, while multi-turn ads in AI Mode are being introduced in the United States. That gives advertisers another route to reach people as search becomes more conversational. The expansion comes as Google Search advertising growth slows. Search revenue grew twelve percent year over year in the third quarter, compared with seventeen percent in the second quarter. Alphabet’s overall quarterly revenue reached $127.6 billion, up twenty-five percent from a year earlier. Google chief business officer Philipp Schindler said click-through rates for sponsored product listings in AI Mode are already comparable to organic listings there. That is the company’s early assessment of a particular ad format, rather than a measure of revenue from AI ads. CEO Sundar Pichai also said overall Search query growth was at its fastest year-over-year rate in five years and credited AI features with changing how people search. Google is building ad formats around that usage; the earnings report makes clear why monetizing it matters. For the research note, consider an agent planning several steps ahead. A world model gives it an internal way to predict what may happen after an action. A paper from earlier this week asks whether those internal predictions actually follow the environment’s dynamics, rather than merely helping an agent complete a task. The researchers propose a measure called latent rollout consistency to check action-conditioned predictions in a model’s internal representation. They applied that test to the advanced latent world models they studied. Their central finding is stark: none faithfully modeled the environment’s dynamics by their criterion, even when assessed as systems built for prediction and planning. For builders, task performance alone may leave a blind spot. A model can be useful on an evaluation while its internal picture of how actions change the world fails this more specific test. The finding concerns the models studied and the paper’s definition of faithfulness; it does not establish that every world-model approach fails. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... A brief security update: The Information reports that Atomicwork CEO Vijay Rayapati told roughly thirty software engineers to prioritize cybersecurity after several incidents demanded an urgent check of the company’s exposure. Atomicwork found it was not directly affected. Rayapati now plans to hire one cybersecurity engineer for every three technical hires. And GitHub’s Octoverse report shows a sharp rise in activity. As highlighted by Simon Willison, merged pull requests tripled in less than a year to 220 million per month, while commits more than tripled to 7.38 billion per month. The counts show a much busier platform; they do not, on their own, isolate the contribution of coding agents. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!