Daily Brief Oct 12, 2026
How we checked the items below
For readers who want to dig in. What happened, in brief, is on the Today tab.
This issue draws on the research brief we write ourselves, produced early on October 12, and last night's sweep. The main-line event date is October 11; the background figures come from an OpenRouter post of September 22. This week's column checks one claim made on October 1 and settles a call we wrote down in August. Material in the later sections dates from October 9 to 11, except From the archive. We swept 118 pieces overnight, and this issue uses 25 outside sources with links you can check.
1. [Today] (event date October 11) Coinbase's CEO says Qwen3.5-9B, a 9-billion-parameter open model fine-tuned on Coinbase's own data, beats frontier models at fraud prevention
Not proven yet： the post names neither the rival model nor the metric; Coinbase hasn't published results
How we checked
we read both posts and the OpenRouter article in the original; for Laskin we read the show transcript. Coinbase has only the CEO's one post: it doesn't say which model it compared against or how it measured, and there's no public evaluation. The speed half of "lower cost and speed" is unclear. OpenRouter speaks only for its own platform; what enterprises buy directly from OpenAI or Anthropic isn't in its numbers. ⚠️ Our analysis is produced with help from Anthropic's models, and Anthropic is one of the closed-model vendors.
[Today] (post October 11) Investor Gavin Baker, citing Silicon Data's "token index," says B300 prices are up 38% and B200 prices up 10% since May 31; we haven't confirmed whether the index measures rental prices or what its base is.
How we checked
Silicon Data is a data company that publishes GPU rental-price indexes; his post doesn't say how this index ties to tokens. All we could find was the latest reading on Silicon Data's B300 index page, US$6.95 per GPU-hour; we found no late-May base, and we haven't confirmed that this page uses the same measure as the "token index" he cites.
[Today] (courtroom notes, October 10) On day four of Qualcomm's antitrust trial against Arm in Delaware, according to notes from someone watching in court, Arm's economist conceded under cross-examination that only five architecture licenses for the latest v10 architecture exist worldwide.
How we checked
[Today] (post October 11) Andrej Karpathy, former head of AI at Tesla and a founding member of OpenAI, roughly estimates that about 20 million people worldwide, about 0.2% of the population, use frontier models for math and programming work; he says the two sides arguing about how capable AI is are on different tiers.
How we checked
[Today] (post October 11) Wharton professor Ethan Mollick says one source of friction for large companies adopting AI is the labs themselves: products and models often change without warning.
How we checked
This is his observation from talking with a few leaders; he doesn't say how large the sample is.
[Today] (post October 11) Sakana AI introduced MASS, built with UC Berkeley: one open model with 27 billion parameters acts as a team with divided roles, then fine-tunes itself on its successful collaborations. The company's post says that after two rounds, its evaluation score per output token is 1.2 to 1.6 times that of the original model.
How we checked
The ratio is evaluation score divided by output length, not a 1.2- to 1.6-fold gain in capability; a higher ratio could simply mean shorter output, and the post doesn't give the score and length changes separately. ⚠️ The numbers come only from the company's posts; we haven't read the paper.
[This week] (show, October 9) Misha Laskin, CEO of Reflection AI, says pretraining its open model Beam used 6,000 GB300s for several weeks and reinforcement learning used more than 10,000 for four weeks, so more compute went to reinforcement learning than to pretraining.
How we checked
The usual assumption is that pretraining eats the most compute; his account says the opposite, but he doesn't say how much more reinforcement learning used. This is the company's own account with no third-party check; another part of Laskin's remarks in the same episode also appears in main-line item 1.
[Today] (October 11) River AI, founded by xAI co-founder Igor Babuschkin, launched an inference service that runs two Chinese open models, DeepSeek and GLM, at prices 20% below the two original vendors' own rates.
How we checked
[Today] (post October 11) Analyst Patrick Moorhead says that after Claude's chat and Cowork were merged, one of his two custom research workflows went from 20 minutes to 2 hours 57 minutes and the other from 6 minutes to 3 hours, with no better output.
How we checked
⚠️ Disclosure: we use Anthropic's models to help with analysis, and Claude is an Anthropic product.
[Looking back] (deep dive, October 7, 2026) The "subsidy" on AI coding subscriptions looks mainly like a problem of tying up compute, not an accounting loss.
How we checked
By SemiAnalysis's estimate, subscriptions bring in about a tenth of Anthropic's revenue but may use more than 40% of its inference compute. We didn't go back and check this round; no result yet. ⚠️ We also use Anthropic's models to help with analysis, and Anthropic is one of the subscription vendors discussed here.
What you are not getting today. What would most affect judgment is Coinbase's evaluation results for that task: Coinbase hasn't published them, so we haven't read them. Beyond that, we didn't read Ramp's original figures, the late-May base of the Silicon Data index, the official transcript of the Arm–Qualcomm trial, the full MASS paper, Culpan's paywalled story or the full South China Morning Post article.
Sources & accounting
The past 24 hours. 118 new pieces came in from October 11 to 12: 87 social-platform posts, 23 arXiv academic papers, 6 show transcripts, 1 subscription newsletter and 1 company or personal blog post. Of those 118, we read 44 and screened out the other 74 as off-topic, so none are left unread. Of the 87 social posts, we read 29 and screened out 58; of the 23 papers, we read 8 and screened out 15, 14 of them judged off-topic before reading; we read all 6 show transcripts. The company-filing, paid-analysis and macroeconomic-data channels had nothing new in the window; the channel that fetches official show transcripts failed last night and brought in 0; and 2 pieces from last night couldn't be loaded because of formatting problems. So "within what we swept" excludes those 2 pieces. The OpenRouter data post behind the main line and the Silicon Data index page are originals we read directly today.
One-time catch-up. No new one-time catch-up of older material today.
A note on source concentration. ⚠️ The main line rests on a single post by Coinbase's CEO, and the background usage and spending figures come from one platform, OpenRouter. ⚠️ The same No Priors interview with Laskin is a source for both main-line item 1 and the second Model watch item. ⚠️ Gavin Baker is both the speaker this week's column checks and the speaker in the first Chips & semiconductors item; they are two different claims. ⚠️ Our analysis is produced with help from Anthropic's models: Anthropic signed the accord in this week's column, the Claude in the second Product moves item is an Anthropic product, From the archive discusses subscription plans that include Anthropic's, and the closed-model vendors in main-line item 1 include Anthropic.
The sources we track. Our long-term roster has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters, with the rest spread across earnings reports, keynotes and other channels. ⚠️ Those are counts of tracked sources, a different population from the 118 new pieces above; they can't be added or compared. Representative names: on social platforms, Teortaxes, Ben Horowitz, Dean Ball, Zvi Mowshowitz and Ethan Mollick; among newsletters, Exponential View; among shows, 20VC, All-In, No Priors and Machine Learning Street Talk. This issue uses 25 outside sources in the body. We count only body links not on our own domain. We added 7 Taiwanese sources in September, bringing the Taiwanese total to 126; Taiwanese media, Taiwanese companies' official channels and local Taiwanese reporters count; foreign outlets relaying Taiwanese media do not.