Daily Brief Oct 27, 2026
How we checked every item today except "Also happened"
For readers who want to dig in. What happened, in brief, is on the Today tab.
This issue draws on last night's collection. The three main-line events are dated October 26; the three preprints in Research & models are dated October 23 to 25; the Named commentary post is dated October 26; and From the archive revisits an October 18 topic. Last night's collection took in 1,677 pieces, and this issue uses 22 outside sources with checkable links.
1. [Today] (filing, October 26) Anthropic's filing shows its Amazon, Google and Microsoft compute deals must be paid whether used or not, while 80% of revenue is usage-based (deep dive)
Not proven yet： the first S-1 doesn't list how much is due in each year
How we checked
This is the first version of the S-1, and the contract section hasn't been audited. The Amazon true-up is "subject to certain adjustments and exceptions if available." The first S-1 says neither what those exceptions are nor how much is due each year. So which year's bills would exceed cash plus revenue can't be calculated today. Anthropic writes only that its cash plus revenue will cover at least the next 12 months.
2. [Today] (launch, October 26) OpenAI launches GPT-6.1 Terra as a mid-priced tier for computer use; every capability figure is OpenAI's own
Not proven yet： cost per task has been measured only by OpenAI
How we checked
OpenAI's launch page is a vendor describing its own product. The prices are official, and the capability scores are self-reported by OpenAI with no third-party measurement. The task count and scoring method of Agents' Last Exam aren't on the launch page. Sanders's comment reaches readers through Willison, and Sanders works at OpenAI, the same party as the launch page, so "midrange" is also OpenAI's description.
3. [Today] (post, October 26) AVERI reviewed 800 prompts from Amazon's internal AI safety test set and found some refusals may hinge on surface wording; Amazon is changing four things
Not proven yet： Amazon supplied the sample, and AVERI didn't verify how it was drawn
How we checked
AVERI's account appears in Brundage's quote post (Miles Brundage, 2026-10-26) and on AVERI's report page (AVERI report page). According to the report page, whether the sample was random rests only on Amazon's word, and AVERI didn't verify it. AVERI also didn't check the four changes Amazon describes, writing only that it has no reason to doubt they were made. The review didn't examine model outputs, so "refusals triggered by surface wording" is a possibility inferred from the prompts, not a measurement. On conflicts of interest, the report page discloses that Amazon has donated API credits to AVERI. Amazon didn't pay for this project but had review and approval rights over what was published. The whole account comes from AVERI and Amazon alone, with no scores and no pass rates.
[Today] (announced October 26) Unconventional AI, a startup founded by former Databricks executive Naveen Rao to build AI chips on an unconventional architecture, has agreed to acquire NeuReality, an Israeli AI inference chip company; the price wasn't disclosed.
How we checked
The announcement gives no performance figures for it.
(Unconventional AI, 2026-10-26, announcement) (Naveen Rao, 2026-10-26)
[Evidence update] (filing, October 26) Amazon said in October 2025 that Project Rainier, an AWS compute cluster, gives Anthropic nearly 500,000 of Amazon's own Trainium2 chips for training and inference, but Anthropic's IPO filing gives no chip count.
How we checked
It gives no chip counts, and the main text doesn't mention Rainier (S-1).
[This week] (post, October 26) Jamin Ball, who writes the software-stock newsletter Clouded Judgment, worked out from Anthropic's IPO filing that the gross margin on serving customers rose from 12% in the first half of 2025 to 85% in the first half of 2026.
How we checked
[Trend] (preprint, October 23) A preprint by Shao Zhang, Hangfan Zhang and co-authors tested 4 general-purpose coding-agent harnesses with 4 models on 6 benchmarks, 52,168 runs in all, and found that harness rankings change with the model and benchmark underneath.
How we checked
(arXiv 2610.32483, 2026-10-23) (arXiv 2610.32131, 2026-10-23)
[Trend] (preprint, October 24) A Nanjing University preprint by Changlian Ma and Limin Wang had an LLM agent spend 20 days building a simulator for Hearthstone Battlegrounds and training a specialist agent, whose average placement over its last 20 games beat the server's top-10 players. The authors conclude that LLM agents do far better as builders than as players.
How we checked
The abstract doesn't say how many games or what period the top-10 figures cover. This is one game, with an environment and benchmark the authors built, and 20 games is a small sample.
[Trend] (preprint, October 25) In a preprint, Haoyu Zhao and co-authors from UCL, UNC and other institutions built a "continual learning exam" to measure whether frontier LLM agents improve from experience, and the best system passed 47.7% of the stages.
How we checked
[Today] (announced October 26) Apple expanded Siri AI from English to five more languages, French, Japanese, Korean, Portuguese and Spanish, and the announcement says server-side AI features have daily usage limits, with expanded use to be offered for a fee later.
How we checked
The announcement gives no limit values, start date or price, and this is Apple's announcement alone.
[Looking back] (SecondSource deep dive, October 18) How tight GPU compute is shows up in contract prices; public list-price indexes can't read it.
How we checked
The 25% most likely landed on newly signed and renewed contracts; that is the deep dive's reading from the quarterly report and the website, and the CFO didn't say which price it applied to.
Still missing today. The gap that would most affect judgment is three sets of numbers the first S-1 lacks: the year-by-year payment schedule for contract obligations, the full commitments table and third-quarter revenue, which await the amended filing. How Amazon, Google, Microsoft and Broadcom account for these contracts may be in the 287 unread filings, which this issue doesn't draw on. OpenAI's GPT-6.1 Terra system card and pricing page aren't part of this issue's evidence, and the task count and scoring method of Agents' Last Exam aren't on the launch page. AVERI's full report page and Amazon's blog post weren't checked word for word. For the three preprints, only the abstracts were used.
Sources & accounting
The past 24 hours. 1,677 new pieces came in from October 26 to 27: 829 academic papers, 288 company filings, 262 company and personal blog posts, 163 social-platform posts, 90 from a second academic-paper feed (counted separately from the academic papers above), 22 show transcripts, 19 subscription newsletters and 4 macroeconomic data releases. Of those 1,677, 344 were read in full, 193 were screened out as off-topic, 1,139 haven't been read yet, and 1 is waiting to be fetched again. Of the academic papers, 139 were read, 68 screened out and 622 not yet read. Of the blog posts, 140 were read, 67 screened out and 55 not yet read. Of the social-platform posts, 36 were read, 57 screened out, 69 not yet read and 1 is waiting to be fetched again. Of the show transcripts, 21 were read and 1 screened out. Of the newsletters, 7 were read and 12 not yet read. Of the 288 company filings, only one, the main text of Anthropic's S-1, was read; the other 287 haven't been read, and they include new filings from Amazon, Google, Microsoft and Broadcom. None of the 4 macroeconomic releases has been read. The paper counts include papers picked up early this morning. Last night's collection picked up 884 original X posts; those 884 are counted differently from the 163 above, so the two can't be divided. One account, @teortaxesTex, accounts for 101 of the 884, followed by @renehaas, @pstAsiatech, @elonmusk and @GaryMarcus. Shows included Latent Space, The Cognitive Revolution, ChinaTalk, No Priors, Training Data and Dwarkesh Podcast; newsletters included Gary Marcus, ChinAI, Epoch Gradient, Exponential View, Import AI, Newcomer, Stratechery and The Zvi.
One-time catch-up. No new one-time catch-up of older material today.
Tracked sources. The long-term roster currently 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. The same source may be tracked on several channels, so the channel figures add up to more than 529. ⚠️ Those are counts of tracked sources and can't be added to or compared with the 1,677 new pieces. Representative names include Lucas Beyer, Graham Neubig, Daniel Kokotajlo and Arvind Narayanan on social platforms, Zvi Mowshowitz, Miles Brundage and Dean Ball among newsletters, and Sam Altman, Dario Amodei and Demis Hassabis among show interviewees. This issue uses 22 outside sources in the body. Only outside links actually cited in the body are counted, and links to SecondSource's own site are left out; this is a different population from the 529 tracked sources above. On the Taiwanese side, 7 new sources were added in September, bringing the total to 136; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.