Need to know
- Anthropic ships Claude Fable 5.1 and Mythos 5.1 with 75% cheaper cached tokens and a restricted tier for cybersecurity and life-sciences orgs.
- OpenAI Astra hits the company's first-ever Critical cybersecurity threshold, able to find and exploit zero-days, triggering a staged rollout.
- CrowdStrike launches Falcon Guardian, runtime AI-agent detection and response deployed across hundreds of millions of endpoints.
- Meta switches internal comms from Google Chat to Slack, citing agent infrastructure as the deciding factor.
- Clay closes a new funding round at a $7 billion pre-money valuation, up from $5 billion in January.
New Releases
Anthropic released Claude Fable 5.1 and Mythos 5.1 today, cutting cached token prices 75% and splitting the same underlying model into a broadly available tier and a restricted tier for approved cybersecurity and life-sciences partners.
- Fable 5.1 is available now via API, cloud platforms, and the Claude desktop app, including Microsoft Foundry; Cursor reports it scores 73.4% on CursorBench 3.2, the highest of any model they have tested.
- Mythos 5.1 ships with fewer safety guardrails and is limited to registered partners; Anthropic also expanded zero data retention so customers can run models on their own infrastructure with no data outflows.
OpenAI published a pre-release safety brief today confirming Astra is the first model to meet its Critical cybersecurity risk tier, capable of independently finding and exploiting zero-day vulnerabilities, and will launch with staged access controls.
- Astra scored 100% on ExploitBench and identified two zero-day vulnerabilities in internal testing; advanced cyber capabilities will be restricted at launch to selected defense-focused partners.
- OpenAI's preparedness framework now requires additional monitoring and access gates before any broader rollout, a structural change that signals capability has outpaced its previous safety categories.
CrowdStrike announced Falcon Guardian at Fal.Con 2026, an AI Detection and Response product that monitors and enforces policy on AI agents at runtime across endpoints, cloud, SaaS, and browsers.
- Falcon Guardian uses the existing Falcon sensor footprint across hundreds of millions of devices to discover, profile, and control AI agents without requiring new infrastructure deployment.
- Google Cloud integration is included at launch, with Falcon connecting into Google Agent Gateway, Gemini Enterprise, and Google Cloud's Agent Registry for cross-platform agent visibility.
Google released Gemini 3.8 Flash, codenamed Skimaki, a coding-focused model whose internal testers reportedly preferred it over Anthropic's Opus model in head-to-head evaluations inside Google's Jetski coding platform.
- Skimaki addresses user complaints about verbose outputs and improves logic for agentic task chains, positioning it for high-volume developer and agent workloads.
- The release cadence signals a strategic pivot under new DeepMind leadership to compete directly with Anthropic and OpenAI in the developer market with faster, cheaper coding-optimized models.
Funding
Clay closed a Wellington Management-led round at a $7 billion pre-money valuation, up from $3.1 billion a year ago, signaling that AI-native GTM tooling is compressing what used to be a multi-vendor sales stack into a single platform with venture-scale returns.
Palo Alto Networks acquired Console, an AI-native IT service management platform whose agents connect identity systems, SaaS tools, and security tooling, adding agentic workflow automation to Cortex on the same day the company reported Q4 revenue of $3.41 billion, beating consensus by $60 million.
Case Studies
Eight CISOs from Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo spent months with Anthropic defining the controls required to run frontier models in regulated financial environments, resulting in the Enterprise Frontier Safeguards framework shipping today alongside Fable 5.1.
- The Analysis and Resilience Center for Systemic Risk drove the requirements, focusing on what bank examiners need rather than benchmark scores, including audit log residency inside the customer's own cloud infrastructure.
- Zero data retention is now extended to allow customers to run Anthropic models on their own infrastructure with no data outflows, a prerequisite for several of the participating institutions.
Trending on X
- Astra zero-day exploit capabilities Security researchers and enterprise architects are debating whether staged access is sufficient containment for a model that scored 100% on ExploitBench, with some arguing the disclosure itself changes attacker incentives.
- Fable 5.1 vs Astra safety posture comparison Practitioners are contrasting Anthropic's tiered-access approach for Mythos 5.1 with OpenAI's staged rollout for Astra, arguing the two labs have converged on structurally similar answers to the same capability-safety problem.
- Meta switching to Slack for agents The AI community is treating Alexandr Wang's internal memo less as a procurement decision and more as a validation signal that agent-native infrastructure is now a real enterprise buying criterion, not a roadmap feature.
- Grok 4.7 at 2.1 trillion parameters Musk's confirmation of a September 12 release and a parameter jump from 1.5T to 2.1T is drawing skepticism from practitioners who note that raw parameter count has been a poor predictor of benchmark performance in recent model generations.
- Bill Ackman on Anthropic profitability Ackman's public claim that Anthropic may be one of the fastest-growing businesses he has ever seen, paired with rumors of profitability, is generating debate about whether Claude revenue is genuinely outpacing OpenAI's compute burn.