You describe your workflows by capability instead of by brand name, you keep your prompts portable, and you test a second option before you are forced to use one.
On Saturday, Stripe finalized a deal to acquire OpenRouter for more than seven billion dollars.
OpenRouter does not train models. It does not own a frontier lab. It is a routing layer, a piece of plumbing that sits between developers and roughly 400 models from OpenAI, Anthropic, Google, Meta, DeepSeek, and others. It serves about 8 million developers and reportedly moved around 1.5 quadrillion tokens in the past year.
Stripe paid more than five times the valuation OpenRouter carried three months earlier, and it did so a few months after buying Metronome, a usage based billing company.
I want to answer the question in the headline directly before I explain why that deal is the reason to ask it now. You avoid AI vendor lock in by doing three things: describe every workflow in terms of the capability it needs rather than the product you happen to use, write your prompts so they contain nothing product specific, and run one real job on a second provider before you are ever forced to. None of that requires a developer. All of it can be done in a week.
The reason to do it this month rather than next year is that the market just told you where the leverage sits, and it is not with you.
Key Takeaways
- Stripe paid over seven billion dollars for a model routing layer, not a model, which signals that the durable value sits in who chooses and meters the model.
- DeepSeek raised API prices by as much as 1,100 percent on August 16, 2026, two days after a major release.
- Research reported in 2026 found 81 percent of US enterprise executives concerned about dependency on a single AI vendor, and 45 percent saying lock in has already blocked them from adopting better tools.
- Anthropic’s 42 minute outage on August 16 took down web, Code, and Cowork while API traffic largely survived, showing that dependency is not evenly distributed across a single vendor’s own surfaces.
- Portability is built by writing capability based workflows and product neutral prompts, not by building technical infrastructure.
Lock In Is Invisible Until You Try to Leave
Most small business owners think they are not locked in because they are not on an annual contract.
That is the wrong definition. Lock in has almost nothing to do with contracts at this scale. It is the accumulation of small, sensible decisions that each made your life easier and collectively made leaving expensive.
Here is what it actually looks like in a business your size. You built a custom assistant inside one product. Your prompts use that product’s specific features. Your team learned that product’s interface and nothing else. Your files live in its project folders. Your automations call its API. Every one of those was the right call in isolation. Together they form a wall.
Research reported through 2026 found that 81 percent of US enterprise executives were at least somewhat concerned about dependency on a specific AI vendor, and 47 percent said losing their primary vendor would disrupt a key business function. Forty five percent said lock in had already prevented them from adopting a better tool.
Those are enterprises with procurement departments and legal teams. If they are stuck, so are you, and you found out the same way they did, which is by trying to move.
I do not think most owners are careless about this. I think the problem is that the cost of lock in is entirely in the future and entirely invisible in the present. It never shows up on an invoice. It shows up on the one day you need to move quickly and discover you cannot.
But what if the reason vendor risk feels abstract is that we keep imagining a company going out of business, when the far more likely version is a price change on a Sunday?
Four Things That Happened in the Last Week
One. The router sold for more than the models it routes to are worth to most buyers.
Stripe’s acquisition of OpenRouter is the clearest possible market signal about where value is accruing. The company that decides which model runs, counts what it used, and sends the bill is worth more than seven billion dollars. Not the model. The chooser.
Nvidia made the same bet in a different way in the same week, open sourcing NeMo Switchyard, a routing library it claims cuts task cost to roughly a third while preserving accuracy. Cognition reported a 28 percent mean cost reduction after integrating it into Devin Desktop.
Two very different companies, one conclusion. Flexibility between models is where the money is.
Two. A provider raised prices by up to 1,100 percent with two days notice.
DeepSeek shipped V4-Pro on August 14 and announced on August 16 that API prices would rise by as much as 1,100 percent, with new peak and off peak tiers.
If your margin was built on that pricing, your margin belonged to DeepSeek. This is not a criticism of DeepSeek, which is entitled to price its product. It is an observation about what happens to anyone whose cost structure sits inside someone else’s pricing decision.
Three. Availability is not uniform even inside one vendor.
Anthropic confirmed an outage beginning around 21:58 UTC on August 16 that took Claude.ai, Claude Code, Claude Cowork, and the platform site offline for roughly 42 minutes. The Console and the API stayed up.
That detail matters more than the outage itself. Businesses running on the API largely kept working. Businesses running on the web apps did not. Same vendor, same incident, completely different exposure depending on which surface you had built on. Dependency is more granular than most owners have mapped.
Four. Deprecation is a standing operational cost now.
Today, August 17, Google retires three Imagen 4 model IDs and directs developers to a Gemini image model instead. The replacement is not a drop in swap. The generate_images() method is gone entirely, so image generation now runs through a different method. Anyone still calling the old endpoints has to re test prompt adherence, aspect ratios, latency, and quotas before their pipelines break.
Nobody signed a contract that ended. The product simply moved, and everyone downstream had to move with it.
Read those four together and the picture is not that any one vendor is unreliable. It is that building your operation as a permanent extension of one vendor’s current product is a strategy with no floor under it.
Describe Capabilities, Not Products
The single most useful change I made in my own business was rewriting how I describe my own workflows.
I used to document my process by naming the tools. Step one, open this product. Step two, use this feature. Step three, run this specific assistant.
Now every step names the capability required. Step one, produce a structured summary of the source material. Step two, draft in the house voice using the reference documents. Step three, check the draft against the claims list.
That sounds like a documentation exercise. It is actually the whole strategy, and here is why.
The moment your process is written in capability terms, three things become possible that were not possible before. You can evaluate a new tool against a real specification instead of a vibe. You can hand the process to someone who does not use your tools. And you can swap a component without rewriting the process, because the process never mentioned the component.
The second piece is prompt portability. Most people’s best prompts are quietly welded to one product. They reference specific features, specific formatting conventions, specific behaviors that only exist in one place. Those prompts are genuinely valuable assets, and if they only run in one product, they are that product’s asset rather than yours.
Rewriting them in neutral language costs a little quality in some cases. I want to be honest about that rather than pretend portability is free. Product specific features exist because they work. What you get in exchange is that your library becomes something you own, stored somewhere you control, usable in whatever tool is best six months from now.
The third piece is the one almost nobody does, and it is the one that actually protects you. Run one real job on a second provider. Not a test prompt. A real job, with your real quality bar, judged the way you would judge the original.
Do it on a normal Tuesday when nothing is wrong. Because the alternative is doing it during an outage, or the week after a price change, when you are making a consequential technical decision under time pressure with your judgment degraded by stress. That is how businesses end up on a worse tool permanently.
There is a business principle underneath all of this that predates AI by a century. Never let a single supplier own a step you cannot perform another way. Restaurants know this. Manufacturers know this. Somehow the software era convinced us that dependency was fine as long as the product was good, and the last twelve months have been a fairly persistent argument that it is not.
Practical Steps
1. Inventory every workflow and the provider it depends on.
List every process in your business that involves AI, and next to each one write the provider, the specific product surface, and the revenue that depends on it working. Most owners find one workflow carrying far more weight than they realized.
2. Rate each workflow easy, moderate, or hard to move.
The rating is not about the model. It is about what you built around it. Custom assistants, stored files, integrations, and team habits all raise the difficulty. Be honest, because an optimistic rating here is a plan that fails on the day you need it.
3. Rewrite your process documentation in capability terms.
Go through your written process and remove every product name from the steps. Replace each with the capability required. Where you cannot describe a step without naming the product, you have found genuine lock in and you now know exactly where it lives.
4. Make your core prompts product neutral.
Take your five most used prompts and strip out anything that only works in one place. Then store them somewhere outside any AI vendor’s product, in whatever document system you already use, with a naming convention and version dates.
Here is the prompt that does the rewriting:
[The Job]
Rewrite my core business prompts so they work on any major AI model without editing.
This is for: me and my team at [YOUR BUSINESS].
It matters because: prompts tuned to one product’s quirks become a hidden switching cost I never agreed to.[The Background]
Here is what you need to know: my core prompts are [PASTE THREE TO FIVE], the product I wrote them for is [NAME], and the outputs I need are [DESCRIBE WHAT GOOD LOOKS LIKE].
Do not use: any syntax, tag, or feature that only exists in one provider’s product.[The Deliverable]
Return: each prompt rewritten, plus a short note on what I removed and what quality I lose by removing it.
Must include: plain language instructions only, no provider specific formatting.
Optimize for: accuracy.[The Questions]
Ask me any questions you have.
5. Run one real job on a second provider this month.
Pick your highest volume workflow. Run one genuine unit of work through an alternative. Score it against your written standard, not against your familiarity with the original. Write down what broke and what the smallest fix would be. You are not switching. You are buying an option.
6. Stress test your pricing against a cost shock.
Model what happens to your margin if your AI input cost doubles, triples, and rises tenfold. Find the point where the business stops being profitable. Then decide whether your pricing structure needs to change now, while you have time to have that conversation calmly with clients.
7. Write the trigger that tells you to move.
Decide in advance what causes a switch: a specific price threshold, a specific number of outage hours, or a specific quality failure. Write the number down. Decisions made in advance under calm conditions are better than decisions made at 10pm during an incident, and you will only ever get one of those two.
Frequently Asked Questions
Is vendor lock in really a problem for a business my size?
Yes, and arguably more so. Large companies have procurement leverage, negotiated contracts, and engineering teams who can migrate. A small business absorbs a price increase or an outage directly, with no cushion and no negotiating position.
Should I just use multiple AI tools for everything?
No. Running everything twice costs more and creates confusion about which output is authoritative. The goal is portability, meaning you could move if you needed to, not redundancy on every task. Maintain one primary tool and one tested alternative.
How much quality do I lose by making prompts product neutral?
Usually less than people expect, and it varies by task. Some product specific features genuinely improve output, particularly around file handling and structured formatting. Test it rather than assuming, and accept a small quality cost on the workflows where portability matters most.
What is the fastest thing I can do this week?
Write down which workflows depend on which provider, and run one real job on a second tool. Those two steps take a few hours and give you both a map of your exposure and one proven alternative, which is most of the protection available.
Does using an AI routing service solve this for me?
It helps with model switching, but it introduces its own dependency, which is exactly the layer Stripe just bought. Routing services are useful. They are not a substitute for describing your own workflows in capability terms.
The Close
Stripe did not pay seven billion dollars for a company that makes AI smarter. It paid seven billion dollars for the company that decides which AI runs and sends you the bill.
That should tell you something about where the power in this market actually sits, and it is not with the person who has spent the last two years building their entire operation inside one product’s interface.
I am not arguing you should distrust your provider. I use mine daily and I am glad it exists. I am arguing that the strength of your business should never depend on a decision somebody else makes on a Sunday afternoon in a pricing meeting you were not invited to.
Portability is not a technical project. It is a way of describing your own work that keeps the choice in your hands. Name the capability, not the brand. Keep your prompts yours. Test the alternative before you need it.
The businesses that get hurt this year will not be the ones that picked the wrong provider. They will be the ones that never noticed they had stopped choosing.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs build AI into their operations without building their operations into somebody else’s product. He is the creator of the Perfect Prompt Framework, a speaker, and the host of a community of business owners who care more about durable systems than shiny ones.
Sources
- Stripe to acquire OpenRouter, TechCrunch
- DeepSeek launches V4-Pro and raises API prices, Caixin Global
- Anthropic confirms Claude outage, BleepingComputer
- Google Imagen 4 API retirement, Kingy AI launch tracker
- Nemotron 3.5 Lightning and NeMo Switchyard, Nvidia
- Enterprise AI vendor lock-in and switching costs, Vaasblock research
- Why enterprises need a multi-model AI strategy, Kai Waehner
