Good morning, AI enthusiasts!
AI is getting more useful in everyday work, from longer video clips to a clearer view of your spending. Today’s briefing also covers new security tools, a fresh model release from Baidu and research on how AI is changing the pace of AI development itself.
Your AI Edge briefing:
THE QUICK SCOOP
Summary: Google launched Gemini Omni 1.2 Flash with native 15-second video generation, stronger reference following and conversational editing. The update lowers the price of 720p generation to $0.07 per second.
The Quick Scoop:
Create clips from 3 to 15 seconds, up from the previous 10-second maximum.
Refine a clip with follow-up instructions instead of starting every attempt from scratch.
The model is generally available through the Gemini API and Google Cloud, with access across Google’s business video tools.
Your AI Edge: Try a short product explainer using existing brand photos. Draft at a lower resolution, check product details and motion, then upscale the version you want to keep.
ANTHROPIC
Summary: Anthropic introduced Personal Finance in Claude, a beta that connects financial accounts so users can ask questions about balances, spending and recurring charges in one place.
The Quick Scoop:
The rollout starts with Pro, Max and Ultra subscribers in the US.
Accounts connect through Plaid. Claude cannot move, manage or invest your money.
Conversations inside the Personal Finance hub are excluded from model training; your usual privacy settings apply outside it.
Your AI Edge: Start with one account and one question, such as which subscriptions increased. Ask Claude to show the underlying transactions and verify important figures with your bank.
SPONSORED BY MAMMOTHCLUB
Stuck on a lesson or unsure what to study next? Bring your question to Pete, MammothClub’s AI Tutor, and turn it into a focused practice session.
Ask for an explanation, then test your understanding with a practice quiz.
Build a weekly study plan around the areas you want to improve.
Explore career paths or rehearse interview questions one at a time.
OPENAI
Summary: OpenAI launched GPT-6 Cyber, a specialized model for authorized security research. Access runs through Daybreak Red, with verification and organizational security requirements.
The Quick Scoop:
OpenAI reports an ExploitGym score of 48.5%, compared with 32.1% for GPT-5.6 Cyber.
The company classifies its cybersecurity capability as Critical under its own Preparedness Framework.
Access includes controls such as identity verification, multi-factor authentication and API IP allowlisting.
Your AI Edge: For security teams, stronger models make clear scope and human review more valuable. Evaluate findings on systems you are authorized to test and track confirmed fixes, not just the number of alerts.
METR
Summary: Independent evaluator METR estimates that Anthropic’s progress on well-specified model tasks is now roughly twice its historical pace. The assessment draws on eleven weeks of access to internal evidence.
The Quick Scoop:
METR believes AI use likely explains most of the current acceleration, but describes that judgment as uncertain.
Its internal-benchmark analysis found about a 1.9-times increase in the rate of progress.
Anthropic redacted much of the evidence METR considered most informative, limiting what readers can independently assess.
Your AI Edge: Treat this as an uncertain research finding, not proof of autonomous self-improvement. For your own team, compare verified results and time saved across repeated tasks rather than relying on productivity impressions.
BAIDU
Summary: Baidu released ERNIE 5.5 Flash Base, giving developers a pretrained foundation for further tuning and research. It is a base checkpoint, so it still needs additional training for many assistant-style uses.
The Quick Scoop:
The model has 280 billion total parameters, with 14 billion active per token.
Baidu says fine-tuning and inference support a context window of up to one million tokens.
Weights are available in FP8 and BF16 under a custom model license; a post-trained version is planned.
Your AI Edge: If your team develops specialized models, assess this as a starting point for adaptation. Review license terms, hardware requirements and training costs before putting it on a deployment shortlist.
SPONSORED BY MAMMOTHCLUB
Turn a blank page into a practical starting point. MammothClub’s ChatGPT for Business Owners course explores ways to support the decisions and conversations that fill your workday.
Develop ideas and refine them through follow-up questions.
Explore employee questions and onboarding tasks.
Practice clearer communication and collaboration with your team.
BASETEN
Summary: Baseten detailed its new Sandboxes service, which gives agents isolated computing environments that retain their work between tasks. The launch focuses on long-running jobs and reducing idle compute costs.
The Quick Scoop:
Each sandbox runs inside a microVM to isolate agent code.
Pausing preserves files, memory and running processes; Baseten reports resuming from standby in about 25 milliseconds.
Sandboxes are free to run until November 16. Additional storage and networking products remain on the roadmap.
Your AI Edge: Test a workflow that pauses for human input and resumes later. Check whether the files and processes survive correctly, then measure the cost of the whole task rather than individual model calls.
ANTHROPIC
Summary: Anthropic added security and administration features to Claude Enterprise, giving teams more visibility into settings and more ways to apply existing policies to AI activity.
The Quick Scoop:
A beta security center compares your organization’s settings with Anthropic’s recommended configuration.
Generally available inference hooks check prompts and tool calls through connected security tools.
New management and access controls include a shared admin console, self-service IP allowlists and member sign-out, with several features in beta.
Your AI Edge: Review which controls apply to the Claude products your team actually uses. Start with a configuration review and test policy checks on a small group before expanding them.
THE MAMMOTH LOWDOWN
⚙️ TOOLS FOR AI EDGE
Five tools to explore for checking the quality and reliability of your AI workflows.
Pydantic Evals - Build repeatable tests for AI outputs, from individual calls to agent workflows.
LangWatch - Trace AI applications and evaluate outputs or entire conversations.
Patronus AI - Evaluate model responses and investigate hallucinations and other quality failures.
Deepchecks - Compare AI versions, test applications and monitor production quality.
garak - Scan language models and AI systems for security weaknesses.
📢 What Else Is Buzzing In AI
Kimi launches hosted agents - Moonshot AI released Kimi Hosted Agents, Yicai reports. The service supplies a managed cloud runtime for long tasks and collaboration between agents, reducing the infrastructure developers need to maintain.
Arcjet introduces a dedicated prompt-injection model - Himalia 1 now powers Arcjet’s prompt-injection detection. Separate models for coding agents and autonomous security agents are entering early access.
Australian universities expand practical AI learning - OpenAI outlined university collaborations and staged ChatGPT Edu rollouts. Its newly published, commissioned survey found that 79% of students wanted more opportunities to use advanced AI on real projects; the survey was conducted in August.