Archive

Today's Key Insights

  • Anthropic’s Fable 5.1 Gets Better, Up to 45% Cheaper — Customers running long autonomous coding or research tasks can get more than 30 percent better agentic-coding performance while paying up to 45 percent less than with Fable 5.
  • OpenAI Prepares Astra, Its First AI Model With ‘Critical’ Cyber Abilities — Select OpenAI partners will get Astra early enough to shore up their defenses before the model’s release.
  • Gemini Agent-Based Video Analysis Cuts Token Usage by Up to 88% — Google is adding agent-based video analysis to three Gemini Flash models, giving users of Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite a stated maximum token-use reduction of 88%.
  • CrowdStrike SafeMind: An Agentic System for Automated Defense — For cybersecurity teams, CrowdStrike SafeMind turns Jensen Huang's warning—automated attacks require automated defense—into a named agentic system from CrowdStrike.

Top Story

Anthropic’s Fable 5.1 Gets Better, Up to 45% Cheaper #

Next Big Future AI • The Decoder AI • TechCrunch AI

Anthropic released Fable 5.1 and Mythos 5.1, claiming sharper performance for agentic coding and scientific research. Fable 5.1 doubles its predecessor’s Terminal-Bench-Science score and improves agentic coding by more than 30 percent.

Fable 5.1 is estimated to cost 25 percent less than Fable 5 for typical workloads billed by token. For long autonomous runs with many tool calls, costs can fall by up to 45 percent, with Anthropic attributing the reduction in part to lower cache-read pricing.

TechCrunch reports that Fable 5.1 also includes changes meant to reduce token costs and false-positive restrictions from the model’s safeguards.

Why it matters: Customers running long autonomous coding or research tasks can get more than 30 percent better agentic-coding performance while paying up to 45 percent less than with Fable 5.

Key Takeaways

  • Fable 5.1 doubles its predecessor’s Terminal-Bench-Science score.
  • Anthropic estimates 25 percent lower costs for typical workloads billed by token, citing reduced cache-read pricing.
  • TechCrunch reports changes to Fable 5.1’s safeguards meant to reduce false-positive restrictions.

Industry Updates

OpenAI Prepares Astra, Its First AI Model With ‘Critical’ Cyber Abilities #

TechCrunch AI • Wired AI

OpenAI is preparing Astra, its first AI model with “critical” cyber abilities, and has previewed precautions ahead of its release. TechCrunch describes Astra as very good at breaking into computer systems.

OpenAI will give select partners early access to Astra so they have time to shore up their defenses. The reports name no participating partners and provide no public release date.

Why it matters: Select OpenAI partners will get Astra early enough to shore up their defenses before the model’s release.

Gemini Agent-Based Video Analysis Cuts Token Usage by Up to 88% #

The Decoder AI • DeepMind Blog

Google is adding agent-based video analysis to Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. The company says the capability cuts token usage by up to 88%.

DeepMind calls the feature “agentic video understanding.” The addition brings the same named video-analysis capability to three Gemini Flash models, with Google highlighting reduced token usage as its key result.

Why it matters: Google is adding agent-based video analysis to three Gemini Flash models, giving users of Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite a stated maximum token-use reduction of 88%.

CrowdStrike SafeMind: An Agentic System for Automated Defense #

NVIDIA Blog

NVIDIA and CrowdStrike introduced CrowdStrike SafeMind, an agentic cybersecurity system, at CrowdStrike's Fal.Con 2026 in Las Vegas on Tuesday, according to NVIDIA's blog.

NVIDIA founder and CEO Jensen Huang joined CrowdStrike CEO and founder George Kurtz for the announcement. Huang told a sold-out crowd: "We're at an inflection point in cybersecurity. Attacks are now automated. Defense has to be, too."

Why it matters: For cybersecurity teams, CrowdStrike SafeMind turns Jensen Huang's warning—automated attacks require automated defense—into a named agentic system from CrowdStrike.