Morning Digest, September 2, 2026

11 newsletters, 12 overlapping stories

Top Stories

Claude Fable 5.1 and Mythos 5.1 ship with cheaper cache reads

(2 newsletters)

Anthropic made Fable 5.1 generally available and released Mythos 5.1 through restricted-access programs for vetted cybersecurity and life-sciences organizations. Fable 5.1 tops Artificial Analysis’ Intelligence Index at a record 66, roughly doubling Fable 5 on scientific research tests, and the safety filter now intervenes 60% less on cybersecurity work and 85% less on basic medical and biology questions. Pricing is $10 per million input tokens and $50 per million output, with cached context cut 75%, though Artificial Analysis found a max-effort task actually costs 20% more because the model writes about 1.7x the text.

Runway’s Solaris generates interfaces instead of code

(3 newsletters)

Runway introduced Solaris, an “Interface World Model” that renders websites and apps as real-time interactive video, generating every frame and every response to user input with no intermediate representation. In early tests it beat frontier LLMs on structural similarity and information retention when creating new interfaces. Early access is limited and Runway is candid about what it cannot yet do, but the same real-time video wave is already producing audience-prompted livestreams and scroll-generated short video apps via MiniMax H3 Max.

Meta takes Muse Code out of beta with multi-agent workflows

(3 newsletters)

Meta’s agentic coding CLI now handles larger engineering tasks by splitting jobs across teams of subagents that share context and sync state as they work, plus inter-session messaging. Paid plans start at $5 per month alongside existing pay-as-you-go, and there is a TypeScript SDK preview for building custom agents. Muse Spark is also now available on the Meta Model API.

OpenAI’s Astra is the first model to cross its critical cyber threshold

(2 newsletters)

OpenAI says Astra can find unknown flaws in computer systems and exploit them with no human guidance, reportedly acing ExploitBench and surfacing two zero-days on a tougher in-house version. The lab restarted the training run it had frozen after the Hugging Face breach and rates Astra as its first “Critical” cyber risk, with access limited at launch and a release expected this week. That puts a real stress test on the industry’s recent summer of caution.

Instagram will demote AI personas that do not label themselves

(2 newsletters)

Accounts built around an AI-generated human must now toggle on an “AI-generated profile” label, replacing the older optional “AI creator” tag, or lose Reels and Explore reach among non-followers. Properly disclosed AI influencers are not penalized, so Meta is targeting deception rather than banning virtual personalities. The rule follows backlash over an Instagram tool that generated people’s likenesses from public photos without explicit opt-in, and Meta’s $18B settlement over social media harms to children.

Vercel publishes DESIGN.md so agents build on-brand pages

(2 newsletters)

Vercel released a public Markdown file that any coding agent can load to get consistent design parameters for color, typography, and spacing, so agents stop shipping generic cookie-cutter UI. The framework pairs the guidance file with a public stylesheet and an evaluation loop, adapting through real-world feedback. The team is upfront that it will not fix messy code or replace real components yet.

Google’s TimesFM-3 forecasts many signals at once, zero-shot

(2 newsletters)

The 330M-parameter time-series foundation model is pretrained on more than 1 trillion time points and forecasts multiple related series in a single pass. It combines historical data with known future covariates such as promotions, holidays, or weather without task-specific fine-tuning, and ranked first across three major public forecasting benchmarks. That makes it immediately useful in retail, finance, healthcare, and manufacturing, where many signals drive the outcome.

OpenClaw 2.0 lands as its largest release yet

(2 newsletters)

What started as an effort to simplify installation and rebuild the browser experience turned into a full overhaul. The release incorporates more than 16,000 pull requests spanning memory, skills, models, automations, apps, plugins, and security.

Also Worth Knowing

Quick Hits