Home
Jul 28, 2026
View All

Cursor Goes Local, Perplexity Goes Hardware, and the PM Who Built Claude

·1 underrepresented voice

The Short Version#

Three signals today that all point at the same thing: AI tools are moving closer to where users actually are — local pricing, local hardware, and the organizational design decisions that made one of the most widely used models possible.

Cursor — Local Pricing for India with Cursor Start#

Source: https://cursor.com/changelog/cursor-start Credibility: High (first-party changelog)

What happened: Cursor launched Cursor Start, a new plan priced at ₹649/month for developers in India. It's designed specifically for local payment infrastructure, including UPI support. Existing free-tier users in India can upgrade from the dashboard.

Key details:

  • ₹649/month is roughly $7.75 USD at current rates — well below the standard Pro plan
  • UPI support removes a real friction point: international card penetration in India is low, and many developer tools have historically struggled with India market payment conversion
  • Positioned as a path to "daily agentic development" — not a stripped-down tier, but a regionally priced entry point

Why it matters for PMs: This is a textbook market-entry move, and it's worth watching carefully. The question with developer tools in emerging markets has always been: do you price for willingness to pay, or do you hold the global price and lose the market entirely? Cursor just picked a side. The UPI integration is the tell — this isn't just a marketing gesture, they actually rebuilt the payment flow to match local infrastructure. For any PM thinking about international expansion, this is the pattern: local pricing plus local payment rails equals real adoption. The global pricing plus "just use a card" approach leaves real markets on the table.

Critical questions:

  • What feature set comes with Cursor Start vs. standard Pro? The changelog says "daily agentic development" but doesn't detail limits or caps.
  • Is this India-first, or does it signal a broader emerging-market pricing strategy (similar to how Spotify has approached Southeast Asia)?
  • How does Cursor measure success here — signups, retention, or actual agentic usage? Those could diverge significantly.

Action you could take today: If your product has meaningful web traffic from India or other emerging markets but low conversion, pull your payment failure data by country. Chances are the friction is payment method, not intent.

Dianne Penn on Lenny's Podcast — How Anthropic Built Claude from the Inside#

Source: https://www.lennysnewsletter.com/p/anthropics-first-technical-pm-on Credibility: High (Lenny's Podcast, direct interview with Anthropic's first PM)

What happened: Dianne Penn, Anthropic's first technical PM, sat down with Lenny to talk through the product bets that shaped Claude: the pivot toward coding, the eval-driven development loop, and how she thinks about "living in the future" as a PM at a frontier AI lab. This is one of the more substantive inside accounts of what it actually looks like to do product at a foundation model company.

Key patterns from the interview:

  • The coding pivot was a deliberate bet, not an obvious one. Penn describes how Anthropic made a strategic call to go deep on coding use cases when it wasn't yet the consensus play. That focus shaped model training priorities and product positioning.
  • Eval-driven development is the loop. Rather than traditional user research cycles, the product process at Anthropic runs heavily on evals — structured tests that measure whether Claude got better or worse at specific tasks. PMs are deeply involved in designing what gets measured.
  • "Token maxing" as a product concept. Penn talks about users who deliberately push Claude to its context limits — a behavior pattern that reveals something about how power users interact with LLMs that traditional PM instincts wouldn't surface.
  • The jagged edge is real and operational. Claude is dramatically better than expected in some areas, and worse than expected in others — and PMs have to build product strategy around a capability map that keeps moving.

Why it matters for PMs: Most accounts of AI product work are abstract. This one is operational. If you're building on top of foundation models, Penn's description of eval-driven development is directly applicable: you need a way to know if your product got better or worse when the underlying model changes, and most teams don't have that. Her framing of "living in the future" is also useful — AI PMs need to prototype with capabilities that don't exist at scale yet, which requires a different relationship with uncertainty than traditional product work.

Critical questions:

  • How does Anthropic decide what goes into an eval versus what stays as qualitative judgment? That boundary seems load-bearing.
  • What happens to product strategy when the "jagged edge" shifts — i.e., Claude gets much better at something users were working around? Does the product roadmap adapt in real time?
  • How does a PM stay grounded in user problems when the product changes faster than user research can track?

Action you could take today: If you're shipping AI features and don't have evals set up, start with one: pick the one user task your feature is supposed to do well at, write five test cases, and run them manually before and after any model change. That's the minimum viable eval loop.

LangChain — How LangChain Built an Agent-First Data Stack#

Source: https://www.langchain.com/blog/agent-data-stack Credibility: High (first-party blog, production case study from the company itself)

What happened: LangChain published a detailed write-up of how they rebuilt their own internal data stack to be agent-first — using Hex, dbt, semantic models, and observability tooling to support self-service analysis at 40x the previous scale. This is notable because it's LangChain eating their own cooking: they built a trusted data agent for internal use and are publishing what worked.

Key technical details:

  • The stack: Hex for analysis, dbt for data transformation, semantic models for consistent definitions, observability layered throughout
  • The 40x figure is for self-service analysis scale — meaning analysts and non-engineers can now run queries without blocking on a data team
  • Semantic models are the key piece: they give the agent a consistent vocabulary for business concepts, which reduces hallucinated or inconsistent query results
  • Observability isn't optional — it's how they maintain trust in agent outputs over time

Why it matters for PMs: "Agent-first" data infrastructure is becoming a real category, and LangChain's write-up is one of the cleaner descriptions of what it actually requires. The semantic model layer is the piece most teams skip — and it's why their data agents produce inconsistent answers. If you're evaluating internal tooling for AI-assisted analytics, this is a useful benchmark for what a production-grade setup looks like versus a demo. The 40x self-service number is also worth pressure-testing: it likely reflects a very specific definition of "self-service," but even a 5x improvement in analyst leverage is material.

Critical questions:

  • What does "trusted" mean in practice here — how do they validate agent outputs before users rely on them?
  • The 40x figure needs a denominator: 40x from what baseline, and what's the error rate at that scale?
  • How does the semantic model stay current as business definitions evolve? That's often the failure mode for this kind of infrastructure.

Action you could take today: If your team uses a data tool with an AI query feature (Hex, Mode, Looker, etc.), try asking it a business question using the natural language interface and then verify the query it generated. That gap between the question you asked and the query it ran is your semantic model problem.

Quick Hits#

  • Aravind Srinivas (Perplexity): Thanked Intel for partnering on a Perplexity Personal Computer with local models and hybrid inference on Intel Ultra Series 3 laptops. This is Perplexity moving into hardware distribution — a significant strategic expansion if it ships. (2026-07-28): https://x.com/AravSrinivas

  • Simon Willison: Published "An opinionated guide to which AI to use to do stuff" — a practical breakdown of when to reach for which model for which task. High signal for PMs building multi-model workflows or advising teams on AI tool selection. (2026-07-27): https://simonwillison.net/2026/Jul/27/an-opinionated-guide-to-which-ai-to-use-to-do-stuff/#atom-everything

  • Dario Amodei: Published a clarification post on open-weight models and AI policy — says Anthropic does NOT support banning open-weight models, but calls for mandatory safety testing before release for both open and closed models. Relevant for any PM thinking about model selection, vendor lock-in risk, or open-source strategy. (2026-07-27): Referenced widely but direct URL not in collected data.

  • Vercel: Regional inference now available on AI Gateway — lets you route model inference requests to specific geographic regions, which matters for latency-sensitive features and data residency compliance. (2026-07-27): https://vercel.com/changelog/regional-inference-now-available-on-ai-gateway

  • Microsoft: Launched MAI-Cyber-1-Flash, their first cybersecurity-specific model built to find vulnerabilities in complex codebases. Combined with their MDASH tooling, they're claiming world-class security performance at 50% of the cost of leading models. (2026-07-27): https://blogs.microsoft.com/blog/2026/07/27/rethinking-security-for-the-age-of-ai/

The Thread#

Localization is the new frontier for AI tool adoption. Cursor's India pricing with UPI rails, Perplexity's Intel laptop partnership for local model inference, and Amodei's policy position on open weights all point at the same underlying tension: AI tools built for US audiences, US payment infrastructure, and US regulatory assumptions are running into the limits of that model. The next wave of adoption — whether it's developers in India, enterprises that need data residency, or users who want local inference — requires product decisions that go deeper than translation.

Sit With This#

Dianne Penn describes Anthropic's product process as "eval-driven development" — where PMs are responsible for designing structured tests that measure whether Claude improved or regressed on specific tasks, rather than relying primarily on user research cycles.

For your team: Do you have a way to know, right now, if the AI feature you shipped last month is performing better or worse than it was at launch? If the answer is "we'd check support tickets or NPS," what would it take to build one eval that gives you a faster, more reliable signal?