The Intake
The Intake — Sunday, August 16, 2026
On the substrate
Fable 5's biology classifier was retrained on August 7 — biology-related fallbacks down 85% overall
If your applications query Claude on biology, chemistry, or molecular design and you've been fielding elevated fallback rates, Anthropic retrained Fable 5's biology classifier on August 7. Anthropic says overall biology-related fallbacks dropped approximately 85%.
The reductions were not uniform across surfaces. Claude.ai saw a 67% drop. Cowork saw 55%. Claude Code saw 17%. Claude Platform saw 7%. Anthropic rewrote the classifier's training rules to more specifically carve out benign use cases. The team developed updated training data and retrained the classifier. Dual-use blocks in virology, toxicology, and molecular design remain in place.
If your work involves biology, chemistry, or molecular design, the reduction varies by platform. Claude.ai and Cowork saw the largest drops — 67% and 55% respectively.
Meta launched Muse Code in beta — a terminal coding agent powered by Muse Spark 1.2, priced at $1.25/M input tokens
If you've been building primarily on Claude Code or Codex as your terminal coding agent, Meta entered the field in beta this week. Muse Code runs in the terminal on macOS and Linux. It is powered by Meta's Muse Spark 1.2 model.
The agent plans changes and writes code. It validates results. It launches parallel sub-agents in isolated worktrees to avoid collisions on large codebases. Pricing is $1.25 per million input tokens and $4.25 per million output tokens. Meta AI chief Alexandr Wang cited cost competitiveness against Claude Code and Codex as a primary selling point.
Meta released Muse Glimmer on August 10. The model is 30B, distilled from Muse Spark. The license is Apache 2.0. Wang indicated open weights for Muse Spark 1.2 would follow. If you're evaluating terminal coding agents, Muse Code is a new option in the field. If open weights matter to your deployment, Muse Glimmer is already available.
Google released Gemini 3.7 Flash on August 13 with agentic video processing enabled by default — 65.3% on DeepSWE v1.1
If you're evaluating models for agentic workflows that include video, Google released Gemini 3.7 Flash on August 13. It is the first Gemini model to enable agentic video processing by default.
The model is generally available via Google AI Studio, Android Studio, and the Gemini Enterprise Agent Platform. Access through the Gemini Spark tier requires an AI Pro or Ultra subscription. Per Google's release, the model reaches 65.3% on DeepSWE v1.1, up from 49.0%. FrontierCode 1.1 Main scores 43.6%, up from 34.4%. AutomationBench scores 30.4%, up from 17.0%.
Pricing is $0.75 per million input tokens and $3.75 per million output tokens. This rate holds through end of 2026. If you're benchmarking models for agent workflows that include video processing, this is the baseline for that capability in the Gemini family.
---
For operators
Anthropic's text watermarking is live globally — built on SynthID-Text, detection API forthcoming
If you generate text output through Claude and your deployment falls under EU content-transparency requirements, Anthropic's text watermarking went live globally on August 14. Anthropic published implementation details the same day.
Per Anthropic's published details, the implementation uses a version of Google DeepMind's SynthID-Text method. It modifies the source of randomness across semantically equivalent word choices. The watermark adds no extra tokens and carries no user or organization identifying information. It is imperceptible to readers. A detection API is forthcoming. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. The global rollout applies because Anthropic does not yet have a durable region-scoping mechanism.
Two limitations are on record. Complete rewrites can remove the watermark. Code receives minimal watermarking due to syntax constraints. If your output is primarily code, the watermark coverage is limited. If your output is prose text, the watermark is live now.
August Patch Tuesday closes an elevation-of-privilege path in GitHub Copilot and VS Code
Microsoft MSRC BleepingComputer SecurityWeek
If you're running GitHub Copilot in VS Code, Microsoft's August 12 Patch Tuesday patched an elevation-of-privilege flaw. CVE-2026-70335 is classified Important severity. No active exploitation was reported at disclosure. The fix is in Microsoft's standard update channel.
---