The Scientist

One thread runs through today's items: the meter moves before the price list does. Anthropic's temporary 50% boost to Claude Code weekly limits expires September 13, and a permanent 25% increase replaces it on September 14, per TLDR IT. That is an effective capacity drop of about 17% — roughly 20% more seats to hold weekly throughput. Instrument per-workflow token burn before September 13, and re-plan every sweep dated after September 14 against base × 1.25. Separately, OpenAI is resetting paid Codex and ChatGPT Work quotas after fixing token-burn bugs that silently consumed 10–50% of customer quota; nightly eval sweeps and judge-scoring runs are the exposed workloads.

Tool-call traces caught all 50 planted cheats and flagged none of 207 honest runs. So any grounded-QA or RAG score reported without those logs is an upper bound, not a measurement. The same Techpresso coverage finds frontier models surface both sides of a genuinely disputed fact only about 52% of the time.

ChinAI's Jeffrey Ding re-queried the SecurityScorecard data behind the widely repeated claim that OpenClaw use in China is "almost double" the U.S. On August 29, 2026 he found slightly more instances in the United States — from a scanner that observes only publicly reachable hosts, which is not the same thing as adoption.

kimi-k3-in-c ran all 2.78 trillion parameters of Kimi K3 on a single CPU, and its throughput curve is a step function. Going from 8GB to 64GB of RAM bought only 1.34x; 64GB to 128GB bought 3.54x once the working set became fully resident, per Unwind AI. Nvidia is making the same argument commercially, shipping Vera Rubin with dedicated storage and networking racks to stop flash stalling before the GPU. Profile wait-on-I/O per step before the next capacity conversation.

Sony Music Publishing, Warner Chappell and others sued Anthropic in the Northern District of California over allegedly torrented works used to train Claude, seeking up to $150,000 per infringed work plus $25,000 per stripped copyright notice, as Pivot 5 reports. It follows the $1.5B Bartz settlement, where the court held that training was legal but piracy-based acquisition was not. Liability attaches to how a corpus was obtained — a property your ETL controls and can log.