A small pricing experiment in late April pulled me out of my AI honeymoon — and made me rethink how much of my workflow is sitting on someone else’s balance sheet.

A two-day glimpse of how dependent we’ve become

On April 21, 2026, Anthropic quietly removed Claude Code from its $17 Pro subscription on a slice of public-facing pages. By the next day, the change had been reverted. Anthropic’s head of growth, Amol Avasare, clarified on social media that the company was running a small test on around 2% of new prosumer signups, and that existing Pro and Max subscribers were not affected. The reasoning he gave was honest: when their existing plans were designed, agentic workflows running for hours simply weren’t a thing yet. Usage had outgrown the original economics, and they were exploring options.

Two percent of new signups. Two days. Reverted after developer backlash.

On paper, it was a footnote. For me, it was a wake-up call.

How quickly a tool becomes infrastructure

Over the past few months, I’ve moved from trying AI to building on AI. Like a lot of people, I’ve stitched Claude into the fabric of how I work — at my day job, on my indie projects, even in the way I plan my weekends. The pace is intoxicating: ideas to prototypes in hours, not weeks.

But the Anthropic experiment surfaced an uncomfortable question I had been quietly avoiding. If a single pricing change at a single company can ripple through my workflow in two days, how much of what I’m building is actually mine?

This isn’t a gripe about Anthropic. The opposite, really. They were transparent about the test, listened to feedback, and rolled it back. What rattled me wasn’t their behaviour — it was the realisation of how quickly a productivity tool becomes infrastructure. And infrastructure is something you depend on without thinking.

The economics nobody is hiding anymore

It’s worth being clear-eyed about why this kind of repricing is on the table for every major lab, not just Anthropic. The numbers behind the scenes are staggering.

Sam Altman has openly stated that OpenAI loses money on its $200/month ChatGPT Pro plan. According to public reporting, OpenAI booked roughly $13.5 billion in net losses in just the first half of 2025, with internal projections pointing to around $14 billion in losses in 2026 against approximately $13 billion in revenue. Anthropic, on its side, spent close to $9.7 billion in 2025 — $6.8 billion of that on compute alone. Epoch AI estimates that frontier labs are currently spending two to three times more than they generate.

So the $17 a month I happily pay isn’t a sustainable price. It’s a land-grab price — designed to attract users, train models on real-world signal, and build habit before the unit economics get rebalanced. That rebalancing is coming. It might be gradual: tighter token caps, more peak-hour throttling, new tiers introduced quietly. It might be sharper: features moving up the price ladder, like the Claude Code experiment hinted at. Either way, the era of “abundant and almost free” is a chapter, not the whole book.

The question worth sitting with

So the question I started asking myself is simple: if any one of these companies took a longer-than-expected coffee break — through pricing, policy, outage, or anything else — what would actually happen to the systems I’ve built?

Would my agentic workflows degrade gracefully, or fall over completely? Would my writing and research pipelines still produce something useful, even if not as fluent? Would my side projects survive a 5x price hike on a single API?

For most of my setup, the honest answer right now is: not really. And I doubt I’m alone.

Where local models start to look interesting

This is where the conversation about local, on-device AI stops feeling like a hobbyist curiosity and starts feeling like resilience planning.

The hardware story has shifted faster than most people realise. According to IDC, more than 370 million generative-AI-capable smartphones shipped in 2025 alone — roughly 30% of the market — and the forecast points to that share crossing 70% by 2028. On the laptop side, Apple’s M-series and Qualcomm’s Snapdragon X chips now ship with neural engines in the 38–45 TOPS range. A modern phone or laptop can run a 4–8 billion parameter model at 15–30 tokens per second — fast enough for real, interactive use.

Local models won’t match a frontier hosted model on raw capability. They don’t need to. The point isn’t to replace what the big labs offer. The point is to make sure your workflow has a floor — a level of usefulness that holds even if the cloud layer becomes more expensive, more rate-limited, or simply unavailable for an afternoon.

Designing workflows like you design backups

The mental model I’ve started using is the one any IT person already applies to data: don’t keep all of it in one place, and don’t trust any single provider to be there forever.

Practically, that means a few things. Separate the irreplaceable from the convenient. The context, prompts, project notes, and structured workflows you’ve built up are yours — keep them in formats and folders that aren’t tied to one vendor’s UI. Identify which steps actually need a frontier model, and which would be perfectly fine running on something local. Summarising a transcript, tagging emails, generating a first-pass draft — these don’t always demand the most expensive intelligence in the room. Have a “good enough” fallback for the workflows that matter most to you, even if you never have to use it.

This isn’t about pessimism. I’m still bullish on what hosted AI can do, and I’ll still be paying for it. It’s about building systems that respect the fact that we’re in a young, fast-moving, financially unsettled industry.

Food for thought

The most important thing the April pricing experiment did for me wasn’t to change my opinion of any one company. It was to nudge me from a consumer mindset to an architect’s mindset — to start asking questions I should have been asking from day one.

So I’ll pass the question along: if your most relied-on AI provider went dark for a week — or goes up 5 times in price overnight — what would still work in your setup? And what would you have to rebuild from scratch?

How are you thinking about resilience in your own AI workflows? Are you already experimenting with local models, multi-provider setups, or other ways of staying flexible? I’d love to hear what’s working — and what isn’t — for people building real things on top of these tools.

#AI #LocalLLM #AIInfrastructure #ResilientWorkflows #IndieHackers #FutureOfWork #AIStrategy