Claude Fable 5.2 Ships and Stripe Builds Its Agent on Deep Agents
AI PM Research
Monday, November 9, 2026
Claude Fable 5.2 Ships and Stripe Builds Its Agent on Deep Agents
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 everywhere at once (AWS, GovCloud, Vercel's AI Gateway, Bedrock) the same day Lenny and Ethan Mollick both reviewed it independently, and buried in the launch wave is a much better story: Stripe built its internal AI agent, Kai, on LangChain's Deep Agents and got 5,000 internal users in four weeks, which is the clearest "how do you actually ship an internal agent fast" case study we've seen all quarter.
|
Anthropic - Claude Fable 5.2 and Claude Mythos 5.2 launch across every major platform simultaneously 2 on November 9, positioning them as "our most capable models for daunting technical work" with claimed strength in coding, finance, and sciences. The release landed everywhere on day one: Amazon Bedrock, AWS GovCloud (US), Claude Platform, and Vercel's AI Gateway all turned on access simultaneously, which is a distribution pattern we haven't seen this tightly coordinated before from Anthropic. |
|
LangChain - Stripe built its internal knowledge agent, Kai, on Deep Agents in one week and hit 5,000 users in a month Stripe built Kai, a company-wide AI knowledge agent, using LangChain, LangGraph, and the Deep Agents framework. The build took roughly one week, and the agent reached 5,000 internal users within about four weeks of launch. |
|
Vercel - AI Gateway production index shows Anthropic winning new teams 2:1, decision models now 26% of requests Vercel published its November AI Gateway production index, reporting that "decision models" (routing/classification-style calls rather than generation) now account for 26% of all gateway requests, and that Anthropic is winning new team sign-ups over competitors at a 2:1 ratio. This is usage telemetry, not a survey, pulled from actual production traffic across Vercel's customer base. |
Also today
- Aravind Srinivas (Perplexity): Launched Teammates, personal work agents that take high-level goals instead of prompts, now rolling out to Max users; separately noted the median Perplexity employee gets 7x more agent-hours per week than in September (2026-11-05/06)
- Karri Saarinen (Cognition): Argues customer support may become a competitive moat once "software is solved," a provocative reframe of where defensibility goes when AI commoditizes build speed
- Amjad Masad (Replit): Disclosed that Replit's custom kernel sandbox design avoids 96% of Linux zero-day vulnerabilities, published the week after critical Linux kernel flaws were disclosed publicly, a concrete security architecture reveal rather than a marketing claim
- Sarah Guo: Argues AI pricing will move from seats to outcomes-based models because "all pricing models have to break through inertia"
- Figma: Shipped "Select area," an AI image editing tool for faster object selection in the desktop app, plus improvements to agent-generated vector illustrations and icons
AI PM Research · Daily updates on AI products and strategy