Issue #10

THE CLAIM. Claude users are canceling en masse over the invisible watermark Anthropic switched on August 2 to comply with the EU AI Act's transparency code.

THE CHECK. count the evidence. Business Insider spoke with four Claude Max subscribers who canceled, and reports 'dozens' of X posts claiming the same; Gizmodo's own phrasing hedges to 'many supposed paying Claude users'. Against that, Anthropic says it has not seen a cancellation uptick. Nobody in this fight has published a real number: the coverage has anecdotes without a denominator, the company has a denial without data, and a cancellation screenshot costs nothing to post. What the wave-framing buries is the legitimate grievance underneath: the mark attaches when Claude merely touches text, and by Anthropic's own description 'a detected mark means Claude touched the text, not that Claude invented it'. Your essay, edited by Claude for commas, carries the same invisible stamp as a fully generated one, and the mark travels with copied text and survives some editing, into a world of employers and schools that will read any detected mark as an authorship verdict. Four people quitting is not a wave. The attribution problem is not a tweet.

THE CLAIM. Dario Amodei publicly supports mandatory pre-deployment testing for frontier AI models. He compared it to airplane certification: models should be blocked if they fail.

THE CHECK. Anthropic spent $3.53 million on federal lobbying in H1 2026, nearly tripling its full-year 2025 spend. The lobbying focused on export controls, cybersecurity, and AI safety standards. A Harvard Kennedy School study found mandatory compliance transforms startup margins from +13% to -7%, with actual costs hitting $344K per deployment against $10K budgets.

THE TWIST. the company that designs the safety framework also has the resources to comply with it. That is not necessarily corrupt. It is a structure where the regulated help write the rules, and the rules happen to be affordable only to the regulated. The name for that structure is regulatory capture.

THE CLAIM. DeepSeek is raising API prices by up to 1,100% with a new peak/off-peak billing structure, effective August 16.

THE CHECK. confirmed. V4-Pro cached input tokens go from $0.00362 to $0.044 at peak hours (1,113%). Output tokens from $0.87 to $3.96 (355%). V4-Flash sees 57-400% increases. Peak hours: 01:00-04:00 and 06:00-10:00 UTC, which covers prime US working hours.

THE TWIST. seven months ago this company panicked Wall Street by training a frontier model for under $6 million and pricing it at a fraction of American competitors. The cheap-AI disruption story just hit its sell-by date.

THE CLAIM. the week's coding-agent launches stack neatly on DeepSWE v1.1: GPT-5.6 Terra 69.6, Grok 4.6 High 65.9, Gemini 3.7 Flash 65.3, Muse Spark 1.2 59.3. Aggregators and social posts quote these as one ranking.

THE CHECK. no shared harness produced those numbers. Google's launch chart, Meta's launch chart, and xAI's launch table each ran a same-named suite independently, and the same models land in both of the two charts that overlap. That overlap is the tell. GPT-5.6 Terra: 69.6 on Google's chart, 64.8 on Meta's. Muse Spark 1.2: 59.3 on Meta's chart, 54.9 on Google's. Roughly five points of daylight per model, and in each case the chart owner's rival scores lower on the owner's chart. As the one careful comparison in circulation puts it, cross-chart readings are 'suggestive but not proof, since it's two different labs running the same-named suite independently', or shorter: 'a shared benchmark name is not a shared benchmark'. The individual numbers are ordinary first-party launch stats. The leaderboard assembled from them is fiction with a spreadsheet aesthetic.

THE CLAIM. Z.ai launched GLM-5.3 on August 14 claiming CyberGym 84.5%, beating Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%), plus 2,436 confirmed vulnerabilities across 269 open-source projects.

THE CHECK. searched for independent CyberGym replication of GLM-5.3 on August 16, 2026: zero results. Every score is Z.ai-reported, Z.ai-tested, on Z.ai's own harness configuration. Of the 2,436 "confirmed" vulnerabilities, only 53 have public CVE assignments. The remaining 2,383 are under embargo. Confirmed is doing work it has not earned.

THE TWIST. the weights are withheld for two weeks "for safety testing." That doubles as a perfect scarcity launch window. No one can replicate the benchmark until after the press cycle ends.

THE CLAIM. Grok 4.7 'will exceed all current models', with a SpaceX engineering corpus so strong Musk 'would be shocked if any model is better at real-world engineering than 4.7'. Reported size: about 2.1 trillion parameters. Timeline: '3 to 4 weeks'.

THE CHECK. every part of that is a promise about an object nobody outside xAI can touch. No benchmark exists, vendor-run or otherwise, because the model is unreleased. The 2.1T figure traces to Musk's remarks and appears on no xAI specification sheet. The training-corpus advantage is unfalsifiable and comes with its own asterisk in Musk's earlier post: the SpaceX data excludes everything ITAR-restricted, which is precisely the rocket-engineering core the sales pitch evokes. And the one checkable element, the calendar, has already failed once mid-claim: on July 25 the model was 'about four weeks out', pointing at August 22; on August 12 that became '3 to 4 weeks', pointing at September 2 to 9. The desk notes the base rate: Grok 4.6 was framed for August 7 and shipped August 12. This is a pre-release superiority claim scored against a track record of sliding windows, which is why it files as UNSUPPORTED rather than wrong: there is simply nothing to check yet, and the claimant knows it.

THE CLAIM. Alibaba's Qwen accumulated 3 billion downloads in six months, eclipsing Meta and Google to become the world's top open-weight AI family. Bloomberg, Yahoo Finance, and Fortune ran it.

THE CHECK. the Hugging Face report covers January through August (seven months, not six). Qwen open-sourced 460+ models. Meta has three main Llama variants. Google has four Gemma sizes. Per model: Qwen 4.4 million downloads, Meta far more per variant, Google 104.5 million. Qwen's per-model adoption trails Meta by 17x and Google by 24x.

THE TWIST. downloads count every variant, quant, and finetune separately. Qwen's 300,000 community derivatives each generate their own download count feeding back into the parent total. The metric measures ecosystem breadth, not model quality.

THE CLAIM. Qwen has amassed more than 3 billion downloads, versus 418 million for Google and 227 million for Meta, so Alibaba now leads open AI, per coverage of Hugging Face's August 14 State of Open Models report.

THE CHECK. the report is real and unusually careful, and every careful sentence got shed on the way to the headline. The count is 'activity observed on the Hugging Face Hub during the first seven months of 2026': one platform, one window, and by design 'nothing is credited for merely having existed longer'. It excludes, in the report's own words, 'API usage, private deployments, or models distributed through other channels'. The 227 million next to Meta's name is that window on that hub, not Meta's tally: Meta's own LlamaCon figure was 1.2 billion cumulative downloads back in April 2025. And the 3 billion sums 460 plus Qwen models against rivals who ship a handful, on counters that meter pulls (CI jobs, mirrors, re-quantizations), not people. Qwen's momentum is genuine and the report documents it well. The crown in the headline is a window-scoped, hub-scoped, SKU-summed crown.

THE CLAIM. Anthropic signed a $9.1 billion compute deal with Riot Platforms for 191MW of capacity.

THE CHECK. Riot's 8-K filing with the SEC on August 10 describes the customer as a "leading frontier AI company." It does not name Anthropic. Bloomberg attached the name citing "people familiar with the matter." Neither company has officially confirmed the counterparty. The $9.1 billion is a 20-year lease running to June 2048, which is $455 million per year. First capacity: 96MW by December 2027. Full 191MW by June 2028.

THE TWIST. Riot stock surged 25% on the headline number. The headline number is a cumulative 20-year figure for a customer the company will not name in its own SEC filing.

THE CLAIM. Vals AI raised $40M led by a16z at a $400M valuation on the pitch that it is the independent evaluator of AI. Its benchmarks appear in model cards from OpenAI, Anthropic, Google, Meta, and xAI.

THE CHECK. Artificial Lawyer reported Vals AI disclosed a customer relationship with one or more of the participants it evaluates. The benchmarks that appear in model cards are produced by a company billing the labs whose products carry those scores.

THE TWIST. independence is the product. The customers are the subjects. The conflict is structural, not hidden. Vals disclosed it. But disclosure does not eliminate the conflict; it just puts it on the record.