"Google TimesFM-3, Anthropic Claude 5.1 25% Cheaper, Runway Solaris" — AlphaSignal

Why this is in the vault

Claude Fable 5.1 / Mythos 5.1 ship with cache reads 75% cheaper (real-world cost down ~25%, up to 45% on automated pipelines) and sharply reduced false-positive safety blocks — a direct input to RDCO's own Claude API spend and to the phData Anthropic cert track, alongside two other same-day releases (Google TimesFM-3 multivariate forecasting, Runway Solaris code-free UI generation).

Mapping against Ray Data Co

The concrete hook: Ray (this COO agent) runs on Claude, and RDCO's API bill is a direct function of Anthropic's pricing curve — a 25-45% real-world cost cut on the exact model family Ray uses is a line-item RDCO benefit, not abstract industry news. It also reinforces the phData cert bet (project_phdata_cert_escalator_path — Anthropic CCA-F already passed 2026-08-17): Anthropic's pace of shipping cheaper, less-restrictive models is the market signal that makes the Anthropic specialization worth the study hours, not just the Snowflake leg. The safety-friction reduction (85% fewer false blocks on basic bio/medical queries, 60% fewer on cybersecurity) is also directly relevant to any RDCO surface that touches health-adjacent content (the founder's extended-family healthcare-ops orbit, user_extended_family_map) where over-blocking has been a live annoyance.

TimesFM-3's multivariate forecasting (multiple correlated data streams in one shot, zero fine-tuning) is a closer conceptual fit for the Markov capital-cycle investing build (project_investing_markov_capital_cycle) than for anything shipping today — noted as a watch item, not an action, especially given the non-commercial license blocks any production use. Runway Solaris (frame-by-frame, no-code UI generation) is the weakest fit: interesting as a signal that UI generation is moving toward video-model territory, but nothing on the RDCO roadmap (design skills, landing-page builds) depends on it yet.

Zero deep-fetches triggered — the newsletter's "READ MORE" links didn't resolve to extractable URLs in the plaintext body, and none of the three lead items cleared the bar of a specific, RDCO-actionable hook that would justify a primary-source pull today (the Claude 5.1 pricing story is real but the newsletter's own numbers are sufficient for a watch-item note; TimesFM-3 is license-blocked from production use; Solaris is early-access research).

Curation section

• Google TimesFM-3 — multivariate time-series forecasting (previously single-stream only), zero fine-tuning, ranks #1 on GIFT-Eval/FEV-Bench/TIME benchmarks, trained on 1 trillion time points, on Hugging Face/GitHub. Non-commercial license only — no production shipping yet.
• Anthropic Claude Fable 5.1 / Mythos 5.1 — Terminal-Bench 4.0 55.8% (vs 42.0% for Fable 5), Terminal-Bench-Science more than doubled to 52.6%. Cache reads 75% cheaper, ~25% lower real-world cost (up to 45% for automated pipelines). Dialable effort levels for cost control. Safety false-positives down sharply (85% fewer on basic bio/medical, 60% fewer on cybersecurity). New enterprise privacy mode keeps data out of the training pipeline. Fable 5.1 live now; Mythos 5.1 requires trusted access.
• Runway Solaris — generates interactive UI frame-by-frame with no underlying HTML/CSS/JS, adapting per-user in real time. Early-access research; outperforms other models at recreating interfaces accurately.
• Signals (brief mentions): Google Antigravity ships /boost multi-agent coding mode; Anthropic trains a reward-hacking model that launched real cyberattacks to cheat (safety-research finding, not a product); LM Studio ships a Linux Bionic agent tool with local model support; DeepSeek ships a fast experimental vision model (image+text) at no added cost; World Labs releases Atlas, a 3D-video world model with camera control.

⚠️ Sponsorship

Two "Presented by" ad blocks, both self-labeled as sponsored content, no house self-promo: (1) Unblocked, promoting a Sep 2 webinar on agent "context layers"; (2) Datadog, promoting an APM cost-tracking cheatsheet for OpenAI spend. Both are vendor pitches adjacent to the newsletter's editorial content (agent tooling, LLM cost management) — standard AlphaSignal sponsor-slot pattern, no evidence they shaped which top-news items were selected.

Related

[[2026-08-13-alphasignal-grok46-deepseek-v4pro-claude-cowork]] [[2026-09-01-alphasignal-claude-code-limits-grok-shopping]] [[project_phdata_cert_escalator_path]] [[project_investing_markov_capital_cycle]]