Claude Fable 5’s July 17 Usage-Credit Error: What the Record Proves
Anthropic’s status page classifies the July 17 Fable 5 usage-credit requirement as erroneous. What the official record settles, and what it leaves to the subscriber.
Takashi Fujino
AI Token Cost by Language: What Five Vendors Actually Publish
Vendors converge on a single rule of thumb for English. Beyond it, the published record fragments, and for Arabic there is nothing to check against. Nine pages from five vendors, read on one date and sorted by what each figure actually measures.
Metaso AI Search for Buyers Outside China: Access, Payment, Output Rights, and Contract Gaps
Metaso is publicly distributed outside China. Registration, payment, commercial output rights, and its video model terms are not established.
Meta AI Has Three Privacy Routes. They Do Not Do the Same Job
Meta prices data use by model ID on the API, markets Incognito Chat
separately, and offers record management in ordinary chat. FSR audited one Japan web account to map what each route actually governs.
Rust’s LLM Policy: What AI-Assisted Contributors Must Change Before the Next Pull Request
On August 5, 2026, five Rust teams ratified an LLM usage policy for the rust-lang/rust monorepo. It does not ban AI-assisted code. It adds reviewer, disclosure, authorship, and testing conditions that code quality alone does not satisfy. A document-first briefing on what changes before your next upstream pull request.
What Can Qwen3.8-Max Do? Capabilities, Limits, and What the Documents Leave Open
Qwen3.8-Max takes image, text, and video input and publishes four separate ceilings, at $2 and $6 per million tokens. This Tier C, document-first briefing reads Alibaba’s own pages as of August 3, 2026, and names what they leave unresolved. FSR ran no hands-on test.
EU AI Act Article 50: The Law, the Code of Practice, and What a Buyer Can Verify
Article 50 of the EU AI Act applies from 2 August 2026, with a narrower marking transition running to 2 December 2026. This Tier C briefing reads the Regulation, the Digital Omnibus and the Code of Practice, and sets out what a buyer can verify in a vendor’s own documents before signing.
Kimi K3 for Technical Buyers: Which Route You Take Changes What You Bought
Kimi K3 ships under a bespoke license, not Modified MIT. API, Business, third-party host and self-hosting each change the terms, the data default and the cost.
Fujitsu PHOTON’s 475× Claim: What the Metric Actually Measures
Fujitsu’s 475× is a throughput-per-memory ratio from one 1.2B table cell, not a reported wall-clock speedup. What the paper measures, what it cost in benchmark quality, and which serving conditions it never discloses.
Claude Opus 5 vs Opus 4.8: Six Runs, No Detected Difference, and a Migration That Still Needs Work
Six runs, three per model, on one frozen task. Claude Opus 5 and Opus 4.8 matched on every scored field. What that tells buyers, and what it does not.