Home
Jun 3, 2026
View All

Uber Caps Claude Code, Cursor Goes Enterprise, and Figma Make Gets Smarter

·1 underrepresented voice

The Short Version#

Uber putting a hard cap on Claude Code usage is the most honest signal we've seen yet about what AI coding tools actually cost at scale — and Teresa Torres's package-hack post is a useful reminder that even PM-builders who love their new superpower still have to think about supply chain security.

Simon Willison / Uber — Uber Caps Usage of AI Coding Tools to Manage Costs#

Source: https://simonwillison.net/2026/Jun/3/uber-caps-usage/#atom-everything Credibility: High (Simon Willison reporting on an industry signal; Uber is a credible, large-scale operator)

What happened: Uber has placed caps on employee usage of AI coding tools — specifically Claude Code — to manage escalating costs. Simon Willison flagged this as a notable signal today. This isn't a performance complaint or a safety concern. It's a budget problem. At Uber's scale, if engineers are running long Claude Code sessions freely, costs compound fast.

Key patterns:

  • This is one of the first credible, large-scale reports of an enterprise hitting cost ceilings on AI coding tools and responding with usage caps rather than model switching
  • The cap appears to be on per-seat or per-session usage, not a removal of the tool entirely — so Uber still wants the value, just bounded
  • This is distinct from "we tried it and stopped" — the tool is still in use, the problem is economics, not utility

Why it matters for PMs: If you're thinking about internal AI tooling rollouts or recommending AI coding tools to your eng team, this is the scenario you need to stress-test before you get there. "We'll see how much it costs when we scale" is not a plan. Uber's situation suggests that token-heavy agentic sessions — the ones engineers love most — are exactly what drives the cost spike. The tools that feel most powerful are the most expensive to run at scale. That creates a real tension between adoption and budget, and it's going to hit every company that goes broad with these tools. The build vs. buy question here is also interesting: at some threshold of usage, the economics of self-hosted models start to look very different.

Critical questions:

  • Is this an Anthropic pricing problem, a Claude Code usage-pattern problem, or just Uber not having budgeted correctly? Those have different implications for the rest of the industry.
  • Are other large enterprises hitting the same ceiling quietly? This could be the visible tip of a pattern.
  • Does capping usage actually preserve the benefit, or does it just frustrate the engineers who were getting the most value?
  • How do you set caps without creating perverse incentives — engineers gaming usage windows or switching to cheaper/worse tools?

Action you could take today: If you have AI coding tools in your team's stack, ask engineering leadership for a monthly cost report. If they don't have one, that's your answer — nobody is watching the meter.

Cursor — Organizations for Cursor Enterprise#

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

What happened: Cursor shipped "Organizations for Cursor Enterprise" on June 3. Enterprise customers can now manage multiple Cursor teams from a single place, with separate security, governance, budget, and feature controls per team. This is generally available to all Enterprise customers now — not a beta.

Key capabilities:

  • Multi-team management from one admin interface
  • Per-team controls for security policy, feature access, and budget
  • Separate governance settings per organizational unit — so a security team can have different rules than a product team
  • Generally available (not staged rollout)

Why it matters for PMs: This is the move that turns Cursor from a developer power tool into an enterprise procurement line item. The prior model — one Cursor team per company — was fine for a single eng team adopting it bottom-up. But it breaks down the moment an enterprise wants to deploy across multiple divisions with different compliance requirements or cost centers. By adding multi-team org management with per-team budget controls, Cursor is directly addressing the objection that enterprise IT and finance teams raise when asked to approve broad rollouts. This also connects to the Uber story above: if you can cap budgets per team at the tool level, you have a real answer to the "how do we control costs" question.

Critical questions:

  • What does "budget controls" actually mean in practice — per-seat caps, token caps, dollar limits? The changelog doesn't specify.
  • Does this change the pricing model for Enterprise, or is it an additive feature on existing contracts?
  • How does multi-team org management compare to GitHub Copilot Enterprise's admin controls, which enterprise buyers already know?
  • What's the minimum team/seat count where Organizations becomes relevant?

Action you could take today: If you're evaluating AI coding tools for a team larger than one squad, pull up Cursor's Enterprise page and compare the org management controls to whatever you're currently using. The gap between "individual tool" and "managed enterprise deployment" is exactly where deals get made or lost.

Teresa Torres — What I Learned from the Recent Wave of Package Hacks (And Is Cowork Immune?)#

Source: https://www.producttalk.org/package-hacks-cowork/ (2026-06-03) Credibility: High (Teresa Torres is a recognized PM thought leader; this is first-person product builder reflection)

What happened: Teresa Torres published a post today about the recent wave of open-source package supply chain hacks and what she's learned as someone actively building Cowork — her own AI-assisted product. The excerpt shows she's writing from the perspective of a builder who loves having "a new superpower" and is thinking carefully about the security surface area that comes with it. The post engages directly with whether Cowork is vulnerable to the same class of attacks.

Key patterns:

  • Torres is writing as a PM-turned-builder who uses AI tools to ship real product, not as a theorist
  • She's applying a practical security lens to her own vibe-coding workflow — asking "am I exposed?" rather than "isn't this exciting?"
  • The framing of AI as a "new superpower" combined with genuine security scrutiny is a useful model for how PM-builders should hold both truths at once

Why it matters for PMs: Most of the "PMs building with AI" content right now is celebration mode. Torres's post is something rarer: a practitioner who loves the capability and is still willing to ask hard questions about the risks it introduces. For anyone using AI-assisted development tools to build or prototype — Cursor, Lovable, Replit, etc. — dependency management and package security are real concerns that often get skipped in vibe-coding workflows. The AI won't warn you that the npm package it just suggested has been compromised. That's still your job.

Critical questions:

  • What does "immune" actually mean in this context — is Cowork's architecture meaningfully different, or is Torres documenting that she checked and found no exposure?
  • How do AI coding tools handle package version pinning and security audits by default? This is a gap worth investigating.
  • Is this a one-time check or a process Torres now runs continuously?

Action you could take today: If you've built or prototyped anything with an AI coding tool in the last 90 days, run npm audit or the equivalent for your stack. If you don't know how, ask your AI coding tool — and then verify the output manually.

Figma — Plan Mode in Figma Make#

Source: https://help.figma.com/hc/en-us/articles/40830441709719-Use-plan-mode-in-Figma-Make (2026-06-03) Credibility: High (first-party changelog)

What happened: Figma shipped "Plan mode" in Figma Make today. It's an opt-in mode that lets you shape what Make will generate before generation actually starts. According to the changelog: "Make takes a look at your design or prompt and proposes a step-by-step plan for what it will build." You can review, edit, or redirect that plan before any code or components are generated.

Key capabilities:

  • Opt-in pre-generation planning step
  • Make proposes a structured plan based on your design or prompt
  • You can edit the plan before generation starts — changing scope, removing steps, redirecting direction
  • Available in Figma Make today (new release)

Why it matters for PMs: The pattern here is "human checkpoint before AI execution" — and it's showing up across multiple products right now. Cursor has auto-review mode. LangGraph has checkpointing. Now Figma Make has plan mode. The insight is that users don't just want AI to generate things — they want to understand and steer what's about to happen before it happens. This reduces the "undo everything and start over" failure mode that makes AI generation tools frustrating at scale. For PMs thinking about how to design AI features in their own products: a visible plan step before execution is becoming a trust-building pattern worth copying.

Critical questions:

  • Is plan mode on by default or does the user have to discover and enable it? Opt-in features often go unused.
  • How detailed is the proposed plan — high-level steps or specific component-level decisions?
  • Does this meaningfully reduce the iteration count needed to get a usable output, or is it mostly psychological reassurance?
  • How does this interact with Make's codebase-connected mode (shipped May 28)?

Action you could take today: If you have access to Figma Make, enable Plan mode on your next generation task and compare the output quality versus prompting directly. The planning step is also worth screenshotting — it's a concrete artifact for sharing AI-generated design direction with stakeholders before committing.

Quick Hits#

  • Jason Boehmig: Former Ironclad CEO announced his first day at OpenAI on June 1, leading product for the legal vertical. A founder who spent 12 years building contract intelligence moving to OpenAI to lead a vertical is a strong signal about how seriously OpenAI is treating domain-specific product strategy: https://www.linkedin.com/in/jasonboehmig

  • Lenny Rachitsky: Live-tested Google's Gemini Omni avatar feature today — scanned a QR code, cloned his face, and shipped a hype reel in under 15 minutes. Short-form but useful for PMs thinking about how fast avatar/digital twin workflows are getting: https://www.lennysnewsletter.com/p/gemini-omni-clone-yourself-with-ai

  • Microsoft Build 2026: Mustafa Suleyman announced 7 new MAI models (image, voice, transcription), Frontier Tuning capabilities for enterprise model customization, and a collaboration with Mayo Clinic on a frontier health model. The "humanist superintelligence" framing is notable positioning. Full Build coverage at: https://news.microsoft.com/build-2026-live-blog/

  • LangChain: Shipped a post on building custom agent harnesses using create_agent plus middleware, walking through how to customize the core agent loop for specific task types. Useful if you're thinking about how tightly your agent architecture should be coupled to your use case: https://www.langchain.com/blog/how-to-build-a-custom-agent-harness

  • Cursor Enterprise Organizations: Worth noting alongside the main item — the changelog also shows the Auto-review run mode (May 29) is still recent. If you're tracking Cursor's enterprise push, both features together (auto-review + org management) represent a coherent push toward "AI coding tool your IT department can actually approve": https://cursor.com/changelog/enterprise-organizations

The Thread#

The cost of AI tools at scale is becoming a real product constraint — and "governance" is the feature that unlocks enterprise adoption. Uber capping Claude Code, Cursor shipping multi-team budget controls, and Figma adding a human checkpoint before generation all point to the same pattern: the first wave of AI tooling was about capability, and the second wave is about control. Companies that give buyers visible dials — cost caps, plan review, per-team governance — are the ones that survive procurement.

Sit With This#

Uber capped Claude Code not because engineers stopped finding it useful, but because the most valuable sessions are also the most expensive. The tools that feel most powerful are the hardest to budget for.

For your team: If you recommended or approved an AI coding tool rollout in the last six months, do you have a cost monitoring plan that would catch a "Uber situation" before it becomes a finance conversation? And if not — what would that plan actually look like?