Home
Aug 24, 2026
View All

What $20K on Devin Looks Like, and Claude's Adoption Problem

The Short Version#

A solo founder's $20K/month Devin experiment gives us the most concrete data yet on what AI agents cost at operational scale, while Simon Willison flags a meaningful market signal: Anthropic's best model is struggling to attract users even as the company grows. Meanwhile, Vercel Sandbox going globally available quietly closes a gap that was limiting what AI coding tools could do in production.

Lenny's Podcast — Ryan Carson Spent $20K on Devin in a Month#

Source: https://www.lennysnewsletter.com/p/i-spent-20000-on-devin-in-a-month Credibility: High (first-person account from a named solo founder on a widely-read PM publication)

What happened: Ryan Carson, a solo founder, ran 15 concurrent Devin agents simultaneously to handle engineering, customer success, and investor updates. Over one month, he spent $20,000 on Devin alone. He reportedly uses a handwritten list to manage the parallel agent workload. This is among the most detailed public accounts of what running AI agents at operational scale actually costs for a small operation.

Key operational patterns:

  • 15 concurrent agents running simultaneously across multiple job functions (engineering, CS, investor comms)
  • $20,000 in a single month from one person at one company
  • Manual coordination overhead: a handwritten tracking list to manage agent state
  • Scope spans beyond code: customer-facing and investor-facing communications are also delegated

Why it matters for PMs: This is the data point product teams building agent products need to see. $20K/month from a solo founder means the willingness-to-pay ceiling is much higher than most AI products are priced at — but it also means the ROI bar is extremely high. If you're building an agent product, the question isn't just "will people pay for agents?" — it's "can we demonstrate enough value to justify this spend before the trial period ends?" The manual coordination overhead (the handwritten list) is the real signal: orchestration tooling for multi-agent workflows is still a gap that no product has cleanly filled.

Critical questions:

  • What's the actual revenue or time savings Carson can attribute to the $20K? Without that, this is a cost, not an investment.
  • Which of the 15 agent types drove the most value? If it's engineering, the CS and investor comms uses are experiments. If it's CS, that's a different story.
  • How does $20K/month scale if the founder hires even one person? Does the agent stack replace headcount or supplement it?
  • What does "handwritten list" tell us about the maturity of multi-agent coordination? It suggests no existing tool is doing this well enough.

Action you could take today: If you're building or evaluating an agent product, frame your pricing and value prop around this case. What would your product need to deliver for a solo founder to justify $1,500/month? That's 1/13th of Carson's spend. Start there.

Simon Willison — Claude Opus 5 Is Struggling to Attract Users#

Source: https://simonwillison.net/2026/Aug/23/anthropics-best-ai-model-struggles-to-attract-users-as-cheaper-t/ Credibility: High (Simon Willison is one of the most rigorous independent AI tool commentators; he's synthesizing a credible external report)

What happened: Simon flagged a report indicating that Claude Opus 5, Anthropic's best and most capable model, is struggling to attract users — even as cheaper AI tools (presumably Claude Haiku, GPT-4o mini, and others) continue to grow. This is notable because Anthropic released Opus 5 in late July 2026 as a step-change improvement for long-running agents, coding, and professional work.

Key signals:

  • Opus 5 is the flagship, positioned for agentic and professional workloads
  • Cheaper alternatives are winning adoption despite lower capability
  • This mirrors patterns seen in other markets: users often don't choose the "best" option, they choose the "good enough and cheap" option
  • Anthropic's revenue model depends on enterprise and API customers paying for premium tiers

Why it matters for PMs: This is a direct signal on willingness to pay for capability — and it's a cautionary tale. If your product's AI tier strategy assumes users will naturally upgrade to the best model once they experience it, this data challenges that assumption. The pull toward cheaper models is strong even when the quality gap is real. For any PM deciding which model tier to default users to, or deciding how to tier AI features by price, this is the evidence that "best" doesn't automatically win.

Critical questions:

  • Is this a pricing problem (Opus 5 costs too much relative to perceived value) or a distribution problem (users don't know what Opus 5 unlocks)?
  • Are the users of cheaper tools getting 80% of the value at 20% of the cost? If so, Anthropic has a real positioning problem, not just a pricing one.
  • Does the "agentic workloads" use case require Opus 5 performance, or is Flash good enough for most agents? Carson's $20K Devin experiment might tell us.
  • What does this mean for Anthropic's API revenue if enterprise customers start substituting down?

Action you could take today: Review your AI model tier strategy. If you're defaulting users to a premium model, check your usage data: are users generating enough value to justify the cost? If not, you might be burning money on capability users can't perceive.

Vercel — Sandbox Is Now Globally Available#

Source: https://vercel.com/changelog/vercel-sandbox-is-now-globally-available Credibility: High (first-party changelog announcement)

What happened: Vercel Sandbox moved from limited availability to globally available as of August 24. Sandbox enables isolated, ephemeral execution environments — the infrastructure that lets AI coding agents run code safely without touching production systems. This is the kind of backend capability that doesn't make headlines but meaningfully changes what's buildable.

Key capabilities:

  • Globally available (no waitlist, no region restrictions)
  • Ephemeral isolated execution environments
  • Designed for AI agent workloads where code needs to run and be tested safely
  • Pairs with Vercel's existing AI Gateway and v0 toolchain

Why it matters for PMs: Sandboxed execution is a prerequisite for any serious agentic coding workflow. Without it, AI agents can't safely run generated code before deploying it. The fact that this is now globally available means teams building on Vercel's platform can ship agent features that include code execution without building their own isolation infrastructure. It also signals Vercel is serious about becoming the deployment platform for the agent-era, not just the Next.js era.

Critical questions:

  • What are the pricing implications? Sandbox execution at scale could get expensive — is this included in existing plans or metered separately?
  • How does Vercel Sandbox compare to alternatives like E2B or Modal for sandboxed AI execution?
  • Does global availability include the same latency SLAs across all regions, or are some regions best-effort?

Action you could take today: If your team is building any AI feature that involves code execution or agent-driven automation on Vercel, check the Sandbox docs today. This removes a blocker that may have been on your backlog.

Quick Hits#

  • Lenny's Newsletter: Jen Abel breaks down every step of closing $100K+ enterprise deals — relevant if you're an AI PM working with an enterprise sales team on a land-and-expand motion. (2026-08-23): https://www.lennysnewsletter.com/p/how-to-close-100k-1m-deals-step-by

  • Karri Saarinen (Linear CEO): Posted about how Perplexity starts AI projects by exploring LLM capabilities with simple prototypes before designing the experience. That's a specific inversion of the normal product design order and worth sitting with. (2026-08-24): https://x.com/karrisaarinen/status/1889734737042002393

  • Aravind Srinivas (Perplexity): Announced a Perplexity-Intel collaboration bringing local models and hybrid inference to Intel Ultra Series 3 laptops. Local + cloud hybrid inference on consumer hardware is the architecture PMs building for privacy-sensitive use cases should be watching. (2026-08-24): https://x.com/AravSrinivas

  • Notion: Acquired ZeroEntropy, which builds efficient task-specific models for knowledge work. Notion continues to move toward building its own model layer rather than just calling OpenAI/Anthropic APIs. (Recent): https://www.notion.com/blog/zeroentropy-is-joining-notion

  • ElevenLabs: Deprecated their local MCP server in favor of a hosted MCP server. The local server and MCP player are archived and will no longer receive updates. If you're integrated with ElevenLabs via MCP, this is a migration you need to plan. (2026-08-22): https://elevenlabs.io/docs/changelog/2026/8/22

The Thread#

The "good enough and cheap" model is winning. Claude Opus 5 struggling against cheaper alternatives, Carson spending $20K/month on Devin while managing coordination manually, and Karri Saarinen's point about prototyping capability before designing experience — these all point to the same thing. Users and builders alike are making fast, pragmatic choices: cheap models that work adequately beat expensive ones that work brilliantly. The orchestration layer and the UX matter more than the model ceiling. If you're building on premium model capabilities, that's a bet you need to validate explicitly.

Sit With This#

Ryan Carson runs 15 concurrent Devin agents and still manages them with a handwritten list. That's the coordination overhead that no AI product has solved yet — and it's where the workflow breaks down at scale.

For your agent product or AI workflow: If a power user at maximum engagement still needs manual coordination to manage your tool, what does that tell you about the product's ceiling? And is "better orchestration" a feature you should build, or a reason your target user isn't who you think they are?