Free Samples, Then the Hook (And Yeah, I’m in Full Withdrawal)

Look, I wrote this right after a stupid billing glitch with my main AI provider nuked my agentic coding access for two full days. No deep reasoning chains. No autonomous agents refactoring my messy repos. No vibe-coding flow where I spin up tools, iterate like a madman, and ship in hours what used to take weeks.

I felt actual withdrawal. Irritable. Slow. Like my brain was running on dial-up while the rest of me knew what god-mode felt like. That’s when it hit me: these AI companies aren’t just selling tools. They’re dealers. And we’re all hooked.

The Classic Dealer Playbook – Free Samples to Get You Dependent

First they flood the zone with insane value. Generous free tiers, cheap Pro plans, agentic capabilities that feel like having 1-2 full-time devs in your pocket for under $50/month. You get hooked on the productivity superpowers. You start building faster, thinking bigger, shipping stuff you’d never attempt solo. Your workflow changes. Your expectations reset.

Then comes the squeeze.

Everyone’s raising prices. Or doing the sneaky version: shrinkflation on tokens. Same monthly fee, but the model “thinks” less, outputs dumber results, or burns through your quota faster. I’ve seen reports of Claude Opus variants suddenly using 67% fewer thinking tokens for the same tasks. Same price, worse output, more tokens consumed. Classic move.

They’re quietly shifting from flat subscriptions to consumption models – per-call, per-token, usage-based billing (UBB). GitHub Copilot just did it. Anthropic is pushing enterprise users toward metered API rates on top of seats. OpenAI and others are following the same pattern. They pretend the subscription is still the core, but the real work (agentic sessions, heavy reasoning, long contexts) now eats credits that vanish fast. No rollover. Burn mid-project and you’re stuck.

I also see coding agents getting stuck in a loop, when it makes a mistake – wait a second – I’m paying for this tool, and it got stuck in a loop and burned MY credits…. That doesn’t seem fair.

This isn’t random. It’s the plan.

The Humanity Angle: This Technology Must Stay Accessible

Here’s where I get pissed off at the dealer playbook. Let’s stop for just a second.

This shouldn’t just be for those who can afford the new premium rates. AI needs to stay cheap — or get cheap again — because humanity desperately needs broad access to this knowledge and capability.

Think about the 20-year-old kid who couldn’t afford college, grinding in a basement somewhere with a killer idea but no formal credentials. With affordable AI, that kid can prototype, research, iterate, and ship at a level that used to require a full dev team and venture backing. One good prompt chain and they’re competing with people who have more resources.

Or zoom out to emerging markets — places where even $20/month is a real barrier. A developer in Lagos, Jakarta, or rural India with limited local opportunities suddenly has world-class reasoning, coding help, and research at their fingertips. The talent and ideas bubbling up from those places could be the next massive breakthroughs. We’re talking exponential global innovation in emerging markets, if we don’t gatekeep this behind usage-based pricing that prices out the ambitious but broke.

Everyone should have access, not just Western professionals with corporate cards. The more minds hooked into these tools, the faster we solve hard problems in climate, biotech, energy, education — you name it. Imagine the compound effect: millions of new builders creating tools, businesses, and scientific insights we can’t even picture yet. That’s the real moonshot. Locking it behind ever-rising costs kills that potential.

I remember listening to Jake Hirsch-Allen talk about democratizing AI, building a coalition of open democratic economies. Investing in Public AI “The CBC of Compute” he calls it. I recall that line in “Anti-Trust” (A horrible 2* tech film from the 2000’s filmed in British Columbia) “Human knowledge belongs to the world” – Ryan Phillippe said on screen.

Why This Was Always Coming (And Why It’s Not Pure Evil)

Let’s be real for a second – I’m not some conspiracy nut. The economics make sense.

Frontier models are insanely expensive to run. Inference costs, HBM memory shortages, power bills, massive CapEx – the whole industry is bleeding cash even as they raise billions. Demand exploded. They loss-led hard to build market share and moats while compute caught up. The “basically free god-mode AI” era was the sample pack. Now they’ve got us dependent, the bills have to get paid.

Agentic workflows aren’t cheap chats. A multi-hour autonomous coding session that loops, reasons, tools, and iterates eats serious compute. Companies can’t subsidize that forever while memory prices go parabolic and everyone wants Opus-level reasoning.

So yeah – prices up, or shrinkflation, or straight usage-based. Pick your poison. GitHub’s recent Copilot changes (credits instead of unlimited premium requests, Opus multipliers jumping hard) are just the visible tip. More will follow.

We’ve been underpaying for this superpower. Shipping in days what used to need teams? Turning weekend experiments into real tools? That value is still ridiculous compared to hiring humans. But the golden age of predictable flat-rate unlimited was always temporary.

What This Means for You (Practical Moves)

Diversify like your workflow depends on it (it does).

• Stay flexible: Tools like Continue.dev (bring your own keys) or OpenRouter let you pick the best model without middleman markups. Consider holding accounts with more than one provider, another subscription might be cheaper than overage charges (think 2 bags on the aircraft is cheaper than one overweight one)
• Watch the credits: Set hard budgets. Test heavy sessions in advance so you don’t get stranded mid-project. Don’t let flows run endlessly.
• Hybrid stack: Keep one smooth native tool (Cursor, Copilot, Windsurf) for daily flow, but route deep agentic work to raw APIs where you control cost.
• Build habits: Use AI to multiply your thinking, not replace it. The real moat is knowing when to guide, when to verify, and when to ship.
• Prompt engineer: Better prompts = better output = more context = less tokens. Learn to plan projects, and use quality prompts, this will save both time, money and of course – Tokens.

I’m still all-in on AI. Sitting with my kid building stuff is next-level. The superpowers haven’t disappeared – the pricing is just normalizing to reality. How can we make this accessible, don’t gatekeep it behind rich corporations. Those who have this power will become superpowers.

The dealer’s got us. Time to use the high productively while we figure out the new economics.

What do you think? Hooked yet?