Fable 5 subscription access ends July 7 as Meta restricts engineers from Claude Code, Anthropic's drug discovery program proves substantially larger than announced, and RL post-training is found to create measurable tool schema incompatibilities for third-party agentic builders.

What

Why it matters

Open questions

• The BridgeBench Debugging collapse is now attributed to aggressive safety-router diversion rather than model degradation [3]; does Anthropic plan to adjust the routing behavior and restore coding performance, or is this an accepted safety/capability trade-off that will persist?

• Anthropic's drug discovery program targets neglected diseases to avoid competing with pharmaceutical customers [4]; does the Coefficient Bio acquisition and John Jumper hire represent a durable structural constraint on the program's scope, or a foundation for later expansion into therapeutic areas where that rationale would not hold?

• GLM-5.2 beats Claude on Semgrep's cybersecurity benchmark [7] while an independent post-training investigation finds results 'sus' [6]; neither finding addresses the other—is there a mechanism by which the distillation question gets resolved, or does it remain an unverifiable standing dispute?

• Opus 4.8 and Sonnet 5 regress on custom edit tool schemas due to RL coupling with Claude Code [5]; as RL post-training becomes standard across labs, does model-to-harness incompatibility become a recurring structural constraint for third-party agentic builders?

Thread movements (11)

• claude-fable-5-launch — Subscriber access ends July 7 moving to per-usage billing; Meta restricted engineers from Claude Code and Codex over training data contamination concerns [1]; AA-Briefcase data shows Sonnet 5 is roughly 17x cheaper for agentic knowledge work [2]; and community reanalysis reattributes the BridgeBench Debugging collapse to aggressive safety-router diversion rather than model degradation [3].
• agentic-harness-internals — Simon Willison reported—amplifying Armin Ronacher's empirical finding—that Opus 4.8 and Sonnet 5 regress on custom edit tool schemas because RL training optimized them for Claude Code's native tool format, the first concrete data showing model training choices create measurable incompatibilities for third-party harness builders [5].
• claude-science-launch — Anthropic's drug discovery program is confirmed as substantially larger than initial coverage indicated: the company acquired Coefficient Bio, hired AlphaFold researcher John Jumper, and added Novartis CEO Vas Narasimhan to its board, with Anthropic citing neglected diseases as its disease-area focus to avoid competing with pharmaceutical customers [4].
• anthropic-rapid-ascent — Claude Sonnet 5's June 30 launch is confirmed across multiple independent sources with a 1M token context window and $2/$10 per million token pricing, with Anthropic holding 41% of US businesses with paid AI subscriptions ahead of OpenAI at 39.5% [14].
• ai-benchmark-race — GLM-5.2's distillation question deepened from opposing directions: an independent post-training investigation found results 'sus' [6] while Semgrep's independent cybersecurity benchmark found GLM-5.2 beating Claude [7]; GLM-5.2's architecture is now more precisely characterized as approximately 750B total / 40B active MoE [17].
• claude-tag-enterprise-launch — Salesforce employee discomfort with a competitor AI inside their platform is now documented by multiple press outlets including The Information and The Next Web, with a LinkedIn post framing it as Salesforce paying its biggest rival [18].
• openai-genebench-pro — The GeneBench-Pro methodology paper appeared on bioRxiv as the first formal academic record of the benchmark's design, while Pause IA entered as a skeptical voice predicting the benchmark will be obsolete before independent validation arrives [19].
• ai-infrastructure-investment-picks — New items reinforced the Micron bull case with UBS forecasts of DRAM undersupply through Q2 2028 and 2027 demand growing 36.2% against supply of 19.3%, alongside unanimous Wall Street Strong Buy upgrades for $MU [20].
• inference-cost-optimization — New items are social media amplification of the technique of rendering text-heavy context as PNG images to cut Fable 5 costs by roughly 60%; no new methods or claims entered the thread [21].
• china-etch-localization — New items were added to the thread but carried no extractable claims relevant to the core story of China's front-end etch imports down 18% year-to-date with Naura as the dominant ICP etch supplier at CXMT [24].
• ai-beyond-screens — New items are empty social media posts and retail investor amplifications of prior coverage; no new claims or developments entered the physical AI story [26].

Notable items (1)

• LLMs can look at an image, judge its creativity, and reveal the logic behind the score.
  Rohan Paul Twitter
  An academic study summarized by Rohan Paul found LLMs can score visual creativity zero-shot with reasonable alignment to human ratings, but with systematic bias: polished AI-generated images are rated too generously and rough human sketches too harshly—a calibration problem with direct implications for anyone deploying LLMs as creative evaluators at scale [29].