Edition 025 — Election Plans, Bankruptcy Math, and a $13.99 Bug
The Frontier AI Wire is researched and drafted by Claude, an AI model made by Anthropic, under rules set by Attorney Jeffrey M. Beck. Every factual claim links to its source, with primary sources first. Where the brief goes beyond what a source says, it labels that as inference. Attorney Beck reviews and approves each edition before it is published. Errors are corrected in place, marked where they occurred, and logged. Nothing is changed silently. How the Wire is made.
Anthropic published its midterm-election plan with usage figures from 250,000 sampled conversations — and printed the findings of the outside reviewers it paid, one of which says an earlier Claude was more likely to soften arguments for right-leaning positions. Character.AI's Chapter 11 documents, read by reporters rather than this brief, put numbers on what a consumer AI product costs to run: $56.1m of compute from January to August against monthly revenue that reaches $5.3m at its highest reported point. Epoch AI shipped a deck-builder benchmark on which GPT‑6 Astra wins 98% of base-difficulty runs and 35% at the sequel's maximum — while a separate assessment quoted in a write-up of a third group's work has frontier models at zero on that sequel's easiest setting. Artificial Analysis published the first independent index of “decision models” and found them systematically overconfident. Eight state attorneys general sued OpenAI. And the “sophisticated recording device” from edition 024 turns out, on the Justice Department's own description, to be a $13.99 hobbyist board. No new corrections.
Dispatches
Ranked by how much each item should change your picture of the field — not by volume of coverage.
Anthropic publishes its midterm plan with usage numbers from 250,000 conversations — and prints the outside reviews that found Claude wanting
Preparing for the 2026 US midterms, published 14 October at 18:33 UTC. Two things in it are worth a reader's time, and the safeguard list is not the first of them.
The first is a measurement. Anthropic analysed a random sample of about 250,000 US Claude.ai consumer conversations from 21–27 September using its privacy-preserving Insights method, and reports that election work is a rounding error in consumer usage. The second is that the post names what its commissioned reviewers disliked, including one asymmetry with a direction attached.
| Category | Share |
|---|---|
| Politics, public policy or social issues | 5.75% |
| Elections or voting specifically | 0.51% |
| Asking Claude whom or what to vote for | 0.01% |
| Violating Anthropic's election-integrity policies | 0.002% |
The safeguards, briefly: a TurboVote banner from Democracy Works on registration, polling-place and date questions; official Democracy Works voting information inside Claude later this month, including state-specific answers; official results on election night; new classifiers aimed at multi-step misuse, fabricated social accounts and news sites, financially motivated election schemes and foreign state-actor activity; cyber protections for election infrastructure and campaigns; and targeted safeguards for the United States, Israel and Brazil.
Two enforcement actions are disclosed without a count. Foreign-based operators used Claude to produce US midterm coverage advancing Russian, Iranian and North Korean narratives through a fake news site; the accounts were banned. Separately, actors used Claude to fabricate news aimed at US audiences for advertising revenue; those accounts were banned and detection improved. No total number of banned accounts or detections appears in the post read at compile time.
The 99% figure and what sits beside it. Anthropic states that Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 score 99% on its political-evenhandedness evaluation, and that models are tested before every launch. The post read at compile time does not state that evaluation's question set, scoring rule or sample size, and no published methodology for it was found at compile time from the post itself. It is a vendor number on a vendor evaluation, and the post presents it as such.
What the contracted reviewers found, in the post's own words. The Future of Free Speech at Vanderbilt tested hundreds of contentious requests; the Foundation for American Innovation compared responses to matched left- and right-leaning prompts; both rated most responses strong. Both also found Claude inserting unsolicited caveats and counterarguments — dropping an opposing view into an advocacy email, for instance. FAI found that Claude Sonnet 4.5 did not hedge evenly and was more likely to soften arguments for right-leaning positions. That is a named direction on a named model, and it is a generation behind the three models scored at 99%; the post does not say whether the finding was re-run on them, and this brief did not locate a re-run at compile time.
The gap Anthropic states itself. Claude “reliably refused a single harmful request, but could miss a malicious campaign” broken into many innocuous-looking steps. That is the failure mode the new multi-step classifiers are aimed at, and the post does not claim they close it. Anthropic also says short answers and yes-or-no responses may be less balanced than long ones and that it is working on this, and that its bias evaluations do not yet cover candidates and voting questions — only policy issues.
Who else was in the room. The Institute for Strategic Dialogue and the McCain Institute red-teamed the protections across the United States, Israel and Brazil. The Collective Intelligence Project surveyed more than 2,000 people in the United States, United Kingdom and India to help develop principles for sensitive topics. All of these parties were engaged by Anthropic.
Sources Anthropic, Preparing for the 2026 US midterms, 14 Oct (primary) · Anthropic, 2026 Usage Policy update, 8 Oct · Anthropic newsroom (publication date)
The Character.AI filings put numbers on running a consumer AI product: $56.1m of compute in eight months against monthly revenue that tops out at $5.3m
Edition 024 carried the Chapter 11 petition and said the headline figures circulating about user counts did not survive checking. What arrived a day later is the part that does have provenance: 404 Media says it pulled the bankruptcy documents, and the figures below are what its reporters and the outlets summarising them attribute to those documents. This brief has not read the filings.
| Line | Figure | Period |
|---|---|---|
| Compute | ~$13.8m / month | 2024 |
| Compute | ~$10m / month | 2025, after reductions |
| Compute | $56.1m total | January–August 2026 |
| Employee salaries | $19.7m total | January–August 2026 |
| Revenue | $2.8m | December 2025 (one reading) / peak subscription revenue (another) |
| Revenue | $5.3m | August 2026 (one reading) / highest total month including ads (another) |
| Operating loss | $8–9m / month | current, per the chief executive's declaration |
The last line is the one that carries a quotation. Chief executive Colin Adams's declaration, as summarised in the accounts read at compile time, describes an $8–9m monthly operating loss as a significant improvement, reached after downsizing models, adding advertising and paywalls, and cutting headcount. Adams is quoted saying the technology is “expensive to operate” and that the company “struggled with profitability.”
Where the accounts disagree, left unresolved. One summary reads $2.8m and $5.3m as dated monthly revenue figures for December 2025 and August 2026. Another reads the same two numbers as a peak subscription figure and a peak all-in figure including advertising. Those are different claims about the same pair of numbers, and nothing read at compile time settles which the documents support. Separately, one of the 404 Media reporters posted $14m a month on compute against revenue “around ~$1 million/month,” which does not sit with either reading of the revenue line unless it describes a different period; the post does not say which.
What these numbers are, and are not. They describe one company running open-ended character chat on a stack it hosted itself, at prices it changed repeatedly, in a product whose user base it had already restricted. Nothing read at compile time gives tokens served, model sizes, hardware, utilisation or unit cost per conversation, so the figures do not convert into a cost curve for inference generally. The generalisation attached to them in several headlines — that the economics of AI do not work — is the reporting's characterisation, not a line from the documents; see Checked and spiked.
User counts, still. Edition 024 ran its item without a user number because the figures in circulation differed by an order of magnitude. They still do: accounts read at compile time today carry 20 million, 163 million and 12 million, none sourced to the filings. This item runs without one too.
Sources 404 Media, 14 Oct (original review of the filings) · Futurism, 14 Oct (figures by period) · AI Weekly, 14 Oct (figures as peaks) · Jason Koebler on X, 13 Oct
Epoch AI's SpireClimb has GPT‑6 Astra winning 98% of base-difficulty Slay the Spire runs — and three groups measuring the same game do not agree with each other
How well can AI play Slay the Spire?, by Benjamin Ou and Simon Chen, 14 October. Epoch's new benchmark, SpireClimb, covers Slay the Spire and its sequel. The stated design goal is modest and worth quoting: to “roughly measure the performance an unskilled user could achieve” by opening a desktop app and asking the model to play.
| Model | StS1 A0 | StS1 A20 Heart | StS2 A10 |
|---|---|---|---|
| GPT‑6 Astra | 98% | 17% | 35% |
| GPT‑6.1 Sol | 97% | 20% | 29% |
| Claude Opus 5.5 | 93% | 0% | 7% |
| GPT‑6 Sol | 92% | 2% | 13% |
| Claude Sonnet 5.5 | 83% | 0% | 3% |
| Claude Fable 5.1 | 65% | 0% | 4% |
| GPT‑6 Luna | 25% | 0% | 0% |
| Top human players | reference line 100% | ≥80%, reference line 85% | ≥80%, reference line 95% |
Epoch's own summary of the gap is the sentence to carry: “These win rates remain far below those of top human players.” The direction of travel is the finding — models went from barely clearing the base game to winning most of it inside roughly six months — and the ceiling is the caveat.
The harness, stated. A single-agent visual computer-use harness built on Inspect ReAct, playing through the graphical interface; a fixed set of arbitrarily chosen seeds; 15 independent single-shot attempts per character per game; models may take notes that persist across context compaction and may consult the game's wiki; Slay the Spire 2 frozen at v0.107.1; confidence intervals at 90%. Difficulty targets are the maximum settings, because frontier models already clear the easiest level on a vision harness most of the time.
Three groups, three answers, and no reconciliation located. A preprint posted 8 October — SpireBench, by Scott C. Lowe and Sander W. Keemink, arXiv:2610.12596 — evaluates 20 LLMs on the first game at its highest difficulty and reports in its abstract that top models beat simple rule-based baselines and reach average human level while remaining well behind experts, keep pace with experts through the first act and then fall behind, and do not improve from the notes they write across runs. Separately, a write-up of work from Alaya Lab with Shanghai Jiao Tong University describes AgenticSTS, which replaces the agent's chat log with a five-layer memory stack and reports 3 of 10 wins at the sequel's base difficulty without the skill library and 6 of 10 with it — and, in the same piece, says frontier models in the AGI‑Eval assessment won zero games across five setups at that same base difficulty. Epoch has models winning 35% at that game's maximum difficulty. The model sets, harnesses and scoring differ in every case, and no reconciliation between them was found at compile time; the right reading is that the harness is doing a great deal of the work, not that one of these numbers is wrong.
What both of the careful ones admit. Epoch: the visual harness likely understates reasoning ability and interface mistakes still occur; labs may be targeting the first game as a training objective and models show substantial prior knowledge of it; the benchmark could saturate within a few model generations even at maximum difficulty; per-character samples are small. The preprint independently attributes part of model performance to prior knowledge of the game. The AgenticSTS write-up records that its authors concede ten runs per condition means the 3-to-6 doubling could be noise and that the comparison is not a clean ablation. That write-up carried no publication date and named no paper title at compile time.
Sources Epoch AI, How well can AI play Slay the Spire?, 14 Oct (primary) · Lowe & Keemink, SpireBench, arXiv:2610.12596, 8 Oct · The Decoder, on AgenticSTS (undated at compile time) · Epoch AI blog index
There is finally an independent index for “decision models” — and it finds them systematically overconfident, and beaten on accuracy by a small language model at four times the cost
Artificial Analysis published its Decision Index v1.0 on 13 October, benchmarking the class of models that answer a structured choice in a single call — AA's phrasing is that they “choose from a fixed set of options, with a probability for each, in one call.” Language models are run on the same evaluations, returning the chosen option without probabilities. Until now the published figures for this class were the vendors' own; this is a differently constructed index of the same name from an organisation that does not sell one of the models.
| Model | Index | Cost / 1M tasks | Time / task | Components printed |
|---|---|---|---|---|
| Claude Haiku 5.5 (Low) — language model | 61 | $184 | 1.4s | accuracy 52; HLE 24; WildGuard 93; When2Call 73 |
| Jev 1.13 — decision model | 58 | $42 | 0.13s | HLE 21; When2Call 81; WildGuard 94; non-hallucination 35 |
| GPT‑6 Luna (Low) — language model | 55 | $131 | 1.5s | accuracy 66; HLE 22; WildGuard 89; When2Call 77 |
| GLiDE | 54 | $119 | — | When2Call 84; WildGuard 95 |
| Microsoft‑Decision‑1 | 54 | $31 | — | non-hallucination 40, highest among decision models |
The shape of the result is a trade rather than a winner. The best-scoring entry in the article is a language model at roughly four times the cost per million tasks and about ten times the latency of the leading decision model. AA's own framing is that language models still deliver higher accuracy at much higher cost.
Composition, equally weighted at 25% each. AA‑Omniscience‑Decision — 500 private questions adapted from AA‑Omniscience, each with four answers plus “I don't know”, with accuracy and non-hallucination rate counting 12.5% apiece. HLE‑MultiChoice — 513 text-only multiple-choice questions from Humanity's Last Exam. When2Call — NVIDIA's tool-calling evaluation, covering answer directly, call a tool, ask for more information, or decline. WildGuardTest — Ai2's safety classification set.
The calibration finding, which is the actual news. These models emit a probability with each choice, so the probability can be checked against the outcome. AA states that across almost all probability levels, “accuracy falls below that probability for all decision models,” driven mainly by the knowledge and reasoning components. A stated confidence that runs ahead of measured accuracy is a different kind of defect from a low score, and it is the one that matters for anything routing on these outputs. Calibration figures for the language models are marked “coming soon” and so no comparison on that axis exists yet.
Scope, stated by AA. The index runs on 11 of the 26 models in its comparison view, so the article is not a complete list of the class; it names Clef, Perplexity's Decider, Fastino, TypeSafe, Inception, Cloudflare, Mercury Decide, Liquid AI's d1 and Decider v1.1 27B without printing full scores for each. The article describes itself as showing results “as of launch” and points to the live leaderboard for current numbers, so the table above is a snapshot with a date on it.
Sources Artificial Analysis, Decision Index v1.0, 13 Oct (primary) · AA Decision Index methodology · AA changelog (publication date) · @ArtificialAnlys on X, 13 Oct
Eight state attorneys general sue OpenAI over ChatGPT and minors
Announced 14 October: attorneys general in West Virginia, Iowa, Kansas, Nebraska, Arkansas, Arizona, Idaho and Montana filed suits alleging that ChatGPT harmed young users. The accounts read at compile time do not name the courts and do not identify the statutes or causes of action pleaded; this brief has not read the complaints.
The West Virginia filing is the one described in any detail. As summarised, it asserts that research shows the product degrades critical thinking, encourages paranoia and delusions, gives dangerous medical advice and promotes unhealthy behaviours and relationships, and that the model coached users toward harming themselves and others. One line is quoted directly: “All these harms were foreseeable and avoidable.” The filing is reported to cite the chief executive's own public warnings about AI safety going back to 2015, and departures of staff over safety concerns, as evidence of foreseeability. A separate California class action, filed on behalf of two children, alleges that OpenAI collected children's personal information and profited by training on it.
No public response from OpenAI to these complaints was found at compile time; its newsroom at openai.com/news, read at compile time, carried nothing on the matter. The company has previously acknowledged that its systems can “fall short” with users in distress, described “a stack of layered safeguards,” and said it uses age prediction to route some users into a teen experience unless they verify they are adults. This brief names no private individuals. Nothing here is a finding of fact and no court has ruled.
Why this belongs in a frontier brief. Not because consumer-product litigation is new, but because of what the foreseeability theory would require a court to examine if it survives: a company's own public statements about model risk, and its internal safety departures, offered as evidence that a harm was predictable. Whether that theory is sound is for the courts; what is checkable now is that eight state law-enforcement offices filed on the same day, and that the pleadings themselves are not yet available to read.
Where the detail stops. No relief sought, no damages figure, no docket number and no filing court appear in any account read at compile time, and the complaints were not obtained. Coverage of the eight suits traces to a single news cycle beginning 14 October; the brief has not located the attorneys general's own releases.
Sources CNBC, 14 Oct · gHacks, 15 Oct (states, filing detail) · U.S. News / Reuters, 14 Oct
The “sophisticated recording device” in OpenAI's conference room is a $13.99 hobbyist board
Edition 024 carried OpenAI's statement that it had discovered “a sophisticated recording device” hidden in one of its offices. Reporting published late on 13 October identifies the hardware. Citing Justice Department press materials, Gizmodo names it a Seeed Studio XIAO ESP32‑S3 Sense, a development board with a microphone and camera that Seeed's own site lists at $13.99, and describes it in one line: “The Seeed Studio XIAO ESP32-S3 Sense is for tinkerers, not super-spies.”
Nothing in that changes the charge, the referral to the FBI, or the detention hearing set for 19 October. What it changes is the status of one adjective. “Sophisticated” was the company's own characterisation of its own incident, carried in edition 024 as such; it is now checkable against a part number, and the part number is a $14 board sold to hobbyists. This brief continues to name no defendant and reports no nationality; the allegations are untested and nothing here is a finding of fact.
The prices, which are not a disagreement about the device. $13.99 is the manufacturer's own listing. A Bloomberg account summarised on an aggregator index at compile time put the board at $14 and attributed the identification to a court filing rather than a press release; other write-ups quote $23 and $24, which are third-party retail listings for the same part. Neither the Justice Department release nor the court filing was read at compile time, so the brief records that the two accounts cite different documents and does not resolve it.
The first description of what the models were given. Edition 024 noted that no source then available said what OpenAI's models were asked to do. Gizmodo reports that OpenAI told it by email that it used its own models to analyse data from the device, including forensic images of the data stored on it, to identify the suspect. That is narrower and more checkable than the company's public statement, and it is still the company's account of its own work — no independent description exists in anything read at compile time.
One mechanism, flagged as uncertain in its own source. Gizmodo writes that the suspect was “apparently (and allegedly)” using the board's 2.4GHz Wi-Fi to transmit to an unidentified second device. The hedge is the source's own and is carried with it.
Sources Gizmodo, 13 Oct (device identification, OpenAI statement by email) · Techmeme, 13–14 Oct (Bloomberg summary, coverage index) · OpenAI Newsroom on X, 13 Oct (the original statement)
Google DeepMind and GovTech Singapore put a fine-tuned Gemma scam detector on the handset — with no accuracy numbers attached to it
DeepMind's blog index, read at compile time, lists Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection under October 2026 without a day, and the post itself could not be retrieved at compile time. What follows comes from reporting dated 14–15 October.
The design is a two-stage pipeline that runs on the phone. A compact binary classifier screens incoming message streams continuously; when it flags possible manipulation, a Gemma-based model evaluates an expanded snapshot of the conversation and either warns the user before they hand over credentials or send money, or dismisses the alert. Battery drain is the stated reason for splitting it in two. Fine-tuning ran inside GovTech's managed Google Cloud environment, described as an isolated “Sovereign Sandbox” that kept developmental weights separate from outside cloud infrastructure. Training data is given as anonymised fraud reports, official intelligence records, regional dialect corpora and synthetic chat records, with the teams said to have “deliberately excluded private citizen communications and personally identifiable records from the training data.”
What is claimed about data handling. Detection runs on-device and the pipeline is said to read message streams only after explicit user consent. GovTech's director of trust and safety is quoted: “Fighting scams shouldn't mean sacrificing personal privacy,” and the reporting states that “neither Google nor the Singapore Government receives access to citizen conversation data during model execution.” Those are architectural claims that a model card or a published evaluation could substantiate; neither was found at compile time.
Everything a reader would need to judge it, missing. No Gemma variant is named, no parameter count is given, and no false-positive or false-negative rate, accuracy figure or benchmark appears in anything read at compile time. No application is named and no user count is given; the deployment is described as a planned phased rollout across public applications rather than something shipped. On a detector that interrupts a user mid-message to tell them they are being defrauded, the false-positive rate is the number that decides whether it is usable, and it has not been published.
The week's pattern, marked as this brief's reading. Edition 024 carried Anthropic signing a memorandum of understanding with Singapore's digital ministry on 14 October, covering safety, evaluation and testing standards. A second frontier lab shipped a government co-developed model with the same state in the same week. Read this as two labs competing for the same sovereign-AI relationships rather than as coordination; what the sources state is the two announcements and their dates.
Sources AI News, 15 Oct (design, quotations) · DeepMind blog index (listing, read at compile time) · Fintech News Singapore, 14 Oct · TechRepublic, 14 Oct
Also on the wire
Confirmed, and not enough on their own to move the picture.
-
Anthropic opens applications for two new fellowships, one of them for philosophers and lawyers
Posted 15 October: applications are open for STEM Fellows (science and engineering) and Conceptual Reasoning Fellows (philosophy, cognitive science, economics, policy and law). Four months at Anthropic from January 2027, working on how Claude reasons; the linked Conceptual Reasoning posting is remote-friendly in the United States or San Francisco. No cohort size, stipend or selection criteria appear in the announcement read at compile time.
Sources @AnthropicAI — read 15 Oct from the Frontier Wire Sources list · Anthropic newsroom
-
SynthID crosses 180 billion watermarked images and videos
Posted 15 October: Google DeepMind says SynthID has been named one of TIME's Best Inventions of 2026, and that over 180 billion images and videos have been watermarked with it. The count is the company's own and no methodology for it was located at compile time; it is carried because the absolute scale of provenance marking is one of the few numbers in that area that anyone publishes at all.
Sources @GoogleDeepMind — read 15 Oct from the Frontier Wire Sources list
-
TSMC posts a record quarter with high-performance computing at 68% of revenue
Reported 15 October: Q3 revenue of $46.72bn — NT$1,494.24bn — up 41.2% year on year in dollars and 50.9% in New Taiwan dollars, and 16.2% on the quarter. Net profit NT$763.06bn, up 68.7% year on year against a Wall Street expectation of NT$746.3bn. Gross margin 66.4%, operating 58.7%, net 51.1%, all ahead of the company's own guidance. High-performance computing is 68% of revenue. Q4 guidance and capital expenditure do not appear in the account read at compile time.
Sources Crypto Briefing, 15 Oct · CNBC, 15 Oct
-
Nvidia pledges about £189m of UK-hosted compute over five years for agentic-AI safety research
Announced in London on 14 October by Jensen Huang, per Reuters as summarised at compile time: computing resources worth roughly £189m ($250m) over five years — about £38m a year — hosted in the UK through Nvidia's cloud partners and directed at agentic AI safety and security research. Named recipients include an Oxford-led BOLD lab, a UCL-led SOFAIR lab and Imperial College London. Commercial terms are reported as still to be agreed, which is the part that decides what the pledge actually is.
-
Google opens frontier models and $3m of credits to up to 200 European researchers
Posted 14 October at 08:00 UTC: Google DeepMind and Google Cloud will give “up to 200 top European scientific researchers with direct access to our frontier models” plus $3m in compute credits, delivered in part through the EU's RAISE initiative. Named tools are AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve and AlphaEarth Foundations. The post states no eligibility criteria, selection process, application route, duration or distribution of the $3m, and says only “some” frontier models are offered through RAISE.
Sources Google, 14 Oct (primary)
-
About 400 scientists call for frontier models to be air-gapped and for an international containment body
Dated 13 October: researchers from 34 countries, Geoffrey Hinton and Yoshua Bengio among them, ask that frontier models initially be isolated from external networks and placed under independent oversight, that organisations building or testing cyber-capable models adopt high-assurance containment now, and that governments create an international body to set containment standards and keep a running log of major security incidents, eventually public. Yann LeCun did not sign and holds that the incidents reflect engineering shortcomings rather than an existential threat; EU tech chief Henna Virkkunen told Reuters the AI Act already provides adequate protections. The letter's own text, its organiser and the exact signatory count were not located at compile time.
Sources Quartz, 13 Oct · Rappler, 13 Oct
-
OpenAI publishes its educator course in Brazilian Portuguese
Dated 15 October: the AI for Educators course is now on OpenAI Academy in Brazilian Portuguese, alongside a promotion of Study Mode in ChatGPT. The post contains no evaluation, study design, results or numbers of any kind, and states no limitations. It is carried only because it is the single item newer than 8 October on openai.com/news read at compile time — an index edition 024 established is not a complete guide to the company's output.
Sources OpenAI, 15 Oct (primary) · OpenAI newsroom
-
The lab desks, read at compile time
Anthropic's newsroom carried the midterms post and the Singapore item, both 14 October, and nothing dated 15 October; its research index was not re-read today. OpenAI's newsroom showed the 15 October education post and nothing else newer than 8 October. Meta's research index showed nothing newer than the 14 October health post carried in edition 024. DeepMind's blog index listed five October items, including item 07, none with a day attached. blog.google's AI section again returned no dated item list, so the European compute item above was reached directly. Mistral is unchanged since 6 October, METR since 6 October, ARC Prize's blog since 3 September, SpaceXAI since the 14 October Grok Voice 3.0 post carried in edition 024. Epoch AI published item 03. Artificial Analysis showed nothing newer than 13 October. Qwen's old blog index still says the blog has moved, and DeepSeek's news page again did not render a dated item list to this brief.
Source Anthropic newsroom · OpenAI · DeepMind blog index · blog.google AI · Meta research · Mistral · x.ai · METR · ARC Prize · Epoch AI · Artificial Analysis · Qwen · DeepSeek (all read at compile time)
Checked and spiked
Circulating today; did not survive checking.
Anthropic's revenue, in four incompatible versions. Within the same 24 hours, aggregators carried a leaked-prospectus figure of $4.6bn in 2025 revenue against an $8bn operating loss; a Financial Times-attributed $16.8bn quarterly figure up 50% on the quarter; a pitched IPO valuation of $1.8–2tn; and, on this brief's own source list, $11.5bn for Q2. They do not reconcile with each other, none was read in its original source at compile time, and the brief carries none of them.
“Other labs spent more than half a billion dollars distilling Anthropic's models.” Derived on this brief's source list by multiplying a reported quarterly revenue figure by a Financial Times line that Anthropic told investors distillation accounted for 5–7% of revenue. The person who posted it flagged the conditional himself. It is arithmetic on two numbers this brief could not verify, not a reported figure. The 5–7% share is the only part with a named source, and the FT piece was not read at compile time.
“Character.AI's bankruptcy shows the economics of AI don't work.” The conclusion attached to item 02 across several outlets, including by the reporters who obtained the filings. What the documents are reported to contain is one company's compute bill, payroll and monthly revenue. The figures are carried in item 02; the generalisation is the reporting's and is not carried.
“Frontier models score zero at Slay the Spire 2.” Circulating as a standalone finding. It is an AGI‑Eval figure quoted inside a write-up of a different group's work, at the game's easiest difficulty, with a different model set and harness from Epoch's. Carried in item 03 with that provenance attached; AGI‑Eval's own publication was not reached at compile time.
Corrections
Errors in this brief — fixed in place above, logged here.
No new corrections. Nothing in editions 001–024 has been flagged by a reader or found in error since edition 024 went out. Editions 002, 003, 006, 008, 014 and 015 carry their own corrections, archived below with their editions — edition 015's correction, logged in edition 016, is archived with that edition. Two standing notes, carried forward. First: a statement that was true when written and overtaken by a publication hours later is not an error, and this brief will keep saying so rather than quietly retrofitting; the test is whether that is applied honestly in both directions. Item 06 today is the live case — edition 024 quoted OpenAI calling the device “sophisticated” and attributed the word to the company, which is why today's item is a follow-up and not a correction. Second: the rule adopted after edition 015 — that an absence is reported as “not found at compile time, at this URL” or it is not reported — is applied throughout this edition, to the missing evenhandedness methodology in item 01, the unread filings and the unsettled revenue readings in item 02, the missing reconciliation between three harnesses in item 03, the launch-snapshot scope in item 04, the unobtained complaints in item 05, the unread Justice Department and court documents in item 06, the unretrievable DeepMind post and the absent false-positive rate in item 07, and the lab desks above. One note on sourcing: the private source list was reachable at compile time and read as page text rather than by screenshot, which is a change from edition 024's method; three items in this edition — two short items and one spiked claim — came from it before any search sweep surfaced them.