If I walked into your office today and told you if you gave me $200 dollars a month you could have $14,000 in exchange what would you do? I assume that I would be getting a quick trip to the lobby and a stern talking to by the police along with people asking as to how I got into the building. I would probably say that response would be entirely justified.

However, I wasn’t lying: today, in the market, there exists an arbitrage opportunity that allows businesses and individuals to get exactly this deal. For example, SemiAnalysis found that a maxed-out ChatGPT Pro plan, run hard, can pull as much as $14,000 in API-equivalent value. The economics are absolutely mind boggling, it’s estimated a maxed Pro user burns $3,500 of actual compute against $200 of revenue, a 17.5x gap on cost alone.

In the world of technology, venture backed loss-led market capture is the norm. We’ve seen it with Uber, Amazon, and PayPal among many others. I don’t believe that any of them quite met the aggressive attempts that we are seeing with OpenAI and Anthropic though. The most concerning part is that it works and continues to work. Right now businesses are reconfiguring themselves around this massively discounted model and getting the shock of their lives when they truly discover just how much reality costs. Uber themselves, the undisputed masters of loss-led market capture fell for their own game blowing through their entire AI budget for 2026 in 4 months.

The number of businesses making this decision is growing:

The level of exposure to pricing shock is deep, these numbers represent millions of seats embedded in hundreds of thousands of organisations.

Startups aren’t immune from this either. Y Combinator's Spring 2025 batch was ~46% AI-agent startups (67 of 144). A quarter of the W25 batch had codebases ~95% AI-generated leaving them likely entirely dependent for ongoing feature and maintenance work.

To be honest though, I can’t blame them. I mean they’re getting one hell of a bargain right now. Unfortunately, it’s a bargain that’s Faustian at best. You’re getting the ability to do knowledge/engineering work for 1/17th of the cost it should be. In exchange you get to build your entire business on a shifting substrate. The only plan worth relying on is the one you have for when the subsidy ends.

I think taking this deal without properly preparing is extremely dangerous. Repricing shock and vendor concentration represent near-existential risks to ill-prepared businesses right now. With a shocking near-consistency across the board tweaks, adjustments and retirements are being made to pricing models in real-time:

Pricing alone is not the killer in this situation. Businesses will vote with their wallets. They can and will shop around and let the market handle the rest. Thus, there is an escape hatch if pricing were the only risk in the balance. Unfortunately this escape hatch is rusty at best, if not entirely painted on. Menlo Ventures’ (and note Menlo is an Anthropic investor, so apply some scepticism) 2025 survey indicated that around 88% of enterprise spend is with three companies: Anthropic at 40%, OpenAI at 27% and Google at 21%. This shows it’s really a market with 3 doors to choose from and every single one leads to the same hallway: a rent-seeking landlord repricing your unsustainable workloads.

Even if this weren’t the case I don’t think that switching a model or provider is in any way the hard part. From the perspective of API’s the big three are basically indistinguishable and making a swap is often a single line change. The same Menlo survey found that only 11% of teams switched vendors in a year, and most that moved simply upgraded to a newer model from an existing provider. If the model layer was a limiting factor this would be a much smaller problem.

The lock-in isn’t the model. It’s what you build around and on top of it. It’s the prompts that are tuned to model specific quirks, the agent harness that supports a single tool-calling format, the eval suite that assumes a model’s behaviour and the context layer plumbed around a particular window size. The knowledge you only gain building on a model and fixing it in anger. None of this translates cleanly from model to model even on the same provider. You don’t migrate a model, you migrate scaffolding that is shaped and supported by that model.

This is where the trap closes especially in the startup space. That quarter of YC startups with their 95% AI generated codebases? Those are all codebases that no one fully understands, that require the tool that wrote it to keep extending and maintaining it. Your agreement with Anthropic and OpenAI, made upon terms that likely will no longer exist, is entirely load bearing. When the price moves and it will, there is no way you’re bringing that back in-house in any reasonable or sustainable timeframe.

So the two risks are not separate after all. Repricing is the event. Concentration is why you cannot dodge it. And the scaffolding you built to move fast on cheap tokens is the reason you are still standing in the doorway when the rent goes up.

To be clear: I am not advocating that you don’t use the tools. Nor am I advocating for you to stop taking advantage of the arbitrage opportunity. The deal is real, the productivity is unreal and sitting it out is a massive missed advantage. The point I am making is that you can take a free taste and not get hooked. Here are a few things that separate businesses that live and those that melt.

Price based on reality before you commit. Regardless of the plan quota price your workflow at metered API rates and budget against that number. This is especially relevant if you’re building products that rely on these. The gap is the subsidy and the subsidy is being withdrawn. If the economics only work at the promotional price, you haven’t got a business you’ve got a countdown.

Put a gateway between you and the provider. Route everything through a point you control, OpenRouter, LiteLLM, Cloudflare AI Gateway. This makes provider changes substantially easier. You also get spend visibility and attribution so you can readily manage your future costs. This is a massive leverage point because you’re attacking the problem at the layer where switching is cheap.

Keep the scaffolding portable. The lock-in lives in the prompts, the harness, and the evals, so treat them as your core assets to protect. Abstract tool-calling format. Keep an eval suite that runs against multiple models and optimise for more than just your current choice this might also help you find a cheaper option. Resist tuning a task to a single model beyond what that task genuinely requires. Portability isn’t free, but it’s substantially cheaper than a forced migration on a timetable you don’t control.

Turn on hard spending caps. Tokenmaxxing is a big-tech luxury. You are statistically not from big tech. Set hard limits and run as if the caps are already binding. Set limits, run as if they are there. On most platforms the spending caps are off by default. This is how you accidentally get a 4/5/6 figure bill. Wire spend alerts and budgets the same way you wire your cloud infrastructure because it’s exactly the same.

Negotiate for continuity, not just price. At renewal the terms that matter are the ones that survive repricing. This means: caps that throttle rather than bill above a ceiling, protection against midterm renegotiation and some guarantee of version and price continuity for the life of the contract. While these vendors are pre-IPO they need your logo more than margin, so now is the time to ask.

Keep or get someone that understands the code. This is especially for vibe-coded startups out there. If no one can maintain the thing without the tool then you’ve not built a company, you’ve built a very expensive dependency on someone else’s subsidised pricing model.

The first taste is always free. It’s not a mistake, it’s a business model. The frontier labs are not running a charity, they are acquiring a market, and the discount exists to get you hooked. Take the taste, just don’t go for the ride.