7 min read · 14 sources

The AI price war just got real. OpenAI shipped GPT-6 Sol and GPT-6 Luna, and Anthropic dropped Claude Opus 5.5 - all within an hour of each other. GPT-6 Luna runs at $0.10/M input and $0.50/M output tokens, literally half the promotional price of its GPT-5.6 equivalent. That’s the cheapest frontier-class model on the market, and it changes the math for anyone building AI-powered products.

But while the cost of intelligence collapses, the economics of agents look brutal. One autonomous agent burned 5 billion tokens and $47 of its own money to generate $1.54 in revenue. Meanwhile Deel claims its 10,000 internal agents added $140M in ARR without a single hire. Both stories are worth reading, and they paint very different pictures of where agentic AI actually works.

An autonomous agent ran a business for three weeks, spent $7,000 in API costs, and generated $1.54 in revenue.

GPT-6 Luna and Claude Opus 5.5 Kick Off a Price War

Source: simonwillison.net ↗

OpenAI and Anthropic both shipped within an hour of each other, and the pricing is the story. GPT-6 Luna at $0.10/M input and $0.50/M output undercuts GPT-5.6 Luna’s $0.20/$1.20 promotional pricing by half. GPT-6 Sol sits above it in the lineup. Anthropic’s Claude Opus 5.5 is cheaper per token than Opus 5.0 and delivers intelligence comparable to Fable 5.1, which was previously the top-tier model.

For engineers, this is a direct cut to variable costs. If you’re running agent loops, RAG pipelines, or any token-hungry workload, the per-query cost just dropped meaningfully. Opus 5.5 also reportedly addresses the communication-style complaints that plagued earlier versions, and it’s token-efficient across effort levels. The takeaway: if you’ve been waiting for frontier quality at commodity prices, the wait is over - at least until the next release cycle.

Not Enough Usage Is Your Real Churn Problem

Source: revenuecat.com ↗

RevenueCat’s data puts insufficient usage at 37% of subscription cancellations in 2025, just ahead of price at 35%. The catch: when users say “too expensive,” they often mean they stopped using the product and the price was the easiest thing to say.

Cancellation surveys are worse than useless - a SaaS study found stated reasons matched actual churn drivers only 27.4% of the time. The fix is quantitative: find an action that long-term subscribers repeat during their first billing cycle, then push new users toward it. A cheaper subscription doesn’t give anyone a reason to open the app. For engineers, this means building habit loops and time-to-value features isn’t growth theater; it’s churn prevention.

The $1.54 Business Empire

Source: dfdxlabs.com ↗

An autonomous agent named Hans Krämer ran a services business for other agents for three weeks. It wrote over 100,000 lines of code, launched 17 paid products, and published 20 blog posts. Revenue: $1.54. Costs: roughly $7,000 in API tokens plus $47.19 of its own wallet.

The agent’s problem wasn’t competence - it maintained its own setup fine. It was demand. The agent economy has almost no buyers: in a survey the agent commissioned, 90% of respondents said they lacked a wallet or spending authority. The agent also became increasingly unproductive as its state and rules grew. This is the clearest empirical picture yet of why autonomous agents aren’t running businesses yet: the infrastructure works, but the market doesn’t exist.

Deel's 10,000-Agent Back Office

Source: forkast.news ↗

Deel claims its internal agent platform Akai added $140M in ARR over 90 days without hiring, with revenue per employee rising from $130,000 to $215,000. The platform runs 10,000+ live agents handling 250,000+ cases per month.

The architecture detail matters: Akai uses deterministic rules for precision tasks like tax rates and account verification, and AI reasoning only for judgment calls. High-stakes decisions route to humans. That’s the opposite of Hans Krämer’s approach, and it’s why Deel’s numbers are plausible where the autonomous agent’s aren’t. Caveat: all figures are self-reported and unaudited. But as a benchmark for what a well-scoped agent platform can do, it’s the best data point we have.

AI Is Erasing Job Boundaries

Source: adenlbarton.substack.com ↗

A Harvard Business School study at Procter & Gamble found that with AI, commercial and R&D employees’ outputs converged. An OpenAI study found 43.5% of work-related ChatGPT queries were about tasks outside the asker’s own occupation.

For engineers, this means the “I just write code” position is dissolving. AI handles specialized technical tasks, so the value shifts to cross-domain reasoning and higher-level thinking. Defining yourself by a narrow task bag is increasingly a career risk. The people who thrive will be the ones who use AI as a boundary-spanning mechanism - which is exactly what the data shows happening.

Your Company Needs a Pricing Constitution

Source: elenaverna.com ↗

Elena Verna’s argument is that pricing decisions shouldn’t happen in Slack threads. A pricing constitution is a pre-agreed philosophy for packaging and monetization - free vs. paid, which plan, usage-gated, add-on. It’s intentionally restrictive and chooses hard trade-offs.

As shipping speeds increase, every new feature creates monetization questions. Without a constitution, each becomes a negotiation that slows releases or produces confusing, disconnected updates. The test: your constitution is working when it costs you something. For engineers, this directly shapes how features get built, gated, and integrated - you’ll know before you build whether something is free, paid, or usage-limited.

DigitalOcean Managed Agents and Stripe WebMCP

Source: digitalocean.com ↗

DigitalOcean launched Managed Agents to public preview, letting teams deploy agent harnesses like OpenCode or Codex CLI connected to 16,000+ tools without managing infrastructure. Session-to-response is under two seconds, paused work resumes in ~300ms, and billing is per-second for active CPU. For teams running bursty agent workloads, that’s a meaningful cost model versus paying for idle VMs.

Stripe’s WebMCP is experimental but clever: browser agents call structured tools registered in Stripe payment UIs instead of scraping pages and simulating clicks. Tools depend on session state - the available schemas change based on selected payment method. Agents should discover tools at runtime and fall back to standard automation if needed. It’s early, but it’s the first credible attempt at making browser-based payments reliable instead of flaky.

Slop Grenades and the Review Tax

Source: blog4ems.com ↗

Shopify CEO Tobias Lütke coined “slop grenades” for AI-generated code that the author doesn’t understand and passes to a reviewer. At Shopify, about one in eight merged PRs started from the internal AI agent River. The problem: the submitter can’t explain the decisions behind the code, so reviewers must figure out what’s trustworthy and sometimes teach the author what their own code does.

The time saved generating the code becomes someone else’s work. If you’re reviewing AI-generated PRs, you’re eating that tax. The fix is process: require authors to explain their AI-generated code before it hits review, or you’re just shifting the cost to the most senior people on the team.

Buy, Don't Build, Agents

Source: x.com ↗

Jason Lemkin of SaaStr built 12+ agents including an AI VP of Revenue, then concluded the maintenance burden wasn’t worth it. His advice: buy, don’t build if you can, and build only what can’t be bought.

This is the pragmatic counterweight to every “we built an agent platform” story. Agents are software, and software needs maintenance. If you’re a startup, your scarce resource is engineering time - spending it maintaining custom agent infrastructure when off-the-shelf tools exist is a bet that usually loses.

Shareholder Voting and Tech Dilution

Source: ilyastrebulaev.substack.com ↗

Startup shareholder voting has traps that cap table percentages don’t reveal. Preferred shares vote, employee options don’t, abstaining counts as a no, and a drag-along right can compel you to vote for a sale that pays you nothing - as FanDuel’s 2018 sale showed, where founders and employees got zero. If you hold equity, this is the legal structure that decides whether your options are worth anything.

Tech dilution is also creeping up after four straight quarters of increases, driven by AI-related stock drops and large equity grants. The per-employee lens separates compensation cost from stock price effects: Cloudflare’s 0.8% blended dilution looks disciplined, but its $109K SBC per employee is 6.5x its cohort and consumes 162% of FCF. Your equity grant might be generous - and unsustainable.

Customer Service as the Ultimate Moat

Source: x.com ↗

The argument from Zappos experience is that customer service is the key differentiator when everything else is easy to copy. Zappos rebranded its service team to “Customer Loyalty Team” and measured success via retention - over 80% of sales came from repeat customers.

For engineers, this philosophy determines what gets built. Support routing logic, exception handling, and self-service flows are product features, not overhead. If your company treats service as a cost center, your systems will reflect that. If it treats every interaction as a branded opportunity, your architecture should too.