A practical answer to the question most small business owners have not thought to ask yet: if the AI tool you built your process around gets shut down next month, how much of your work survives, and what should you be doing right now to make sure the answer is “almost all of it”?
Your AI Tool Has An Expiration Date And Nobody Sent You The Memo
Five days from now, on August 9, OpenAI’s Atlas browser stops working. Not “gets worse.” Stops. On August 26, the o3 model retires from ChatGPT after a 90-day sunset. On August 30, the official DALL-E GPT goes away. And on August 31, Anthropic’s promotional pricing on Claude Sonnet 5 ends, which means anyone running volume through that model wakes up on September 1 paying 50 percent more per token.
If you built a piece of your business on any of those, this month is either a Tuesday or a fire drill. Which one depends entirely on a decision you probably made without realizing you were making it.
So here is the direct answer. What happens to your AI workflows when a model gets retired depends on where your work actually lives. If your prompts, your background context, and your process documentation are stored in your own files, a deprecation notice costs you an afternoon of testing and maybe a small quality adjustment. If they live inside one vendor’s chat history, custom GPT builder, or proprietary interface, you are rebuilding from memory under time pressure.
That is the whole thing. Portability is not a technical concept. It is a filing decision.
I want to be honest about what changed here, because I do not think the reason is what most people assume. It is not that vendors got less trustworthy. It is that the pace of model releases and retirements accelerated past the point where any of us can keep up by reacting. Three labs shipped models at trillion-parameter scale in the last three weeks. OpenAI is retiring three things this month. Anthropic is repricing. That is one month.
You cannot out-monitor this. You can only build so it does not matter.
Key Takeaways
- Model deprecation is now a routine operating event, not a rare disruption, and your systems should be built to absorb it.
- Your prompts, context files, and process documentation are the actual asset; the model is a replaceable component underneath them.
- Open-weight models from Alibaba, Moonshot AI, and DeepSeek have made switching genuinely viable for small operators, not just enterprises.
- Model and API fees are rarely your biggest AI expense; labor and integration are, which is exactly what a rebuild costs you.
- Portability is a filing and documentation practice, and you can implement it this week without technical skills.
The Problem Nobody Warned You About When You Started
Here is how it happens, and I say this having done it myself.
You find a tool that works. You spend a few weeks getting good at it. You write a prompt that finally produces the client onboarding summary you want, and you save it where? In the chat. You build a custom GPT that knows your service offerings and your tone, and the instructions for it live where? In the GPT builder. You upload your brand guide, your pricing sheet, and three sample proposals, and those files sit where? Inside that one product.
None of that felt like a mistake at the time. It felt like using the tool the way the tool wanted to be used. That is the trap. Every vendor designs an interface that makes it easiest to store your work inside their walls, because that is good business for them. It is not malicious. It is just not aligned with your interests.
I have lost work this way. Not catastrophically, but enough to learn from. I had prompts I had refined over dozens of iterations that existed nowhere except a conversation thread I could no longer find. I had a whole process for a recurring deliverable that lived in my head plus one custom assistant, and when I wanted to move it, I discovered I could not actually describe the process without the tool in front of me. That is an uncomfortable thing to learn about your own business.
And I want to acknowledge the difficulty honestly, because the advice “document everything” is easy to give and annoying to receive. Documenting your AI workflows is boring. It produces no visible result today. It feels like busywork compared to actually shipping client work. Nobody has ever gotten excited about writing down what they already know how to do.
But here is the reframe that changed how I think about it.
You are not documenting for the future. You are building a portable asset. The prompt library, the context files, the written process: those are the only parts of your AI operation that appreciate in value. The model underneath gets faster and cheaper every quarter, and eventually gets turned off. Your documented process gets better every time you use it, and it works on whatever comes next. One of those is a rental. The other is equity.
Stop thinking of it as backup. Start thinking of it as the actual product.
What The Data Actually Says About Switching Costs
Let me give you the numbers, with sources, because this is where the argument either holds up or it does not.
Model prices are falling fast, and that is not a rumor. The Stanford Institute for Human-Centered AI reported in its AI Index that the cost of querying a model at GPT-3.5-equivalent capability fell from about $20 per million tokens in November 2022 to roughly $0.07 per million tokens by October 2024. That is more than a 280-fold reduction in under two years for the same level of capability. Prices for a fixed capability level keep dropping. The floor is still moving.
The alternatives are real and they are named. On August 3, 2026, Alibaba released Qwen3.8-Max, a 2.4 trillion parameter mixture-of-experts model with a one million token context window, priced at $2 per million input tokens and $6 per million output on Alibaba Cloud Model Studio, with open weights promised the following week. On July 16, 2026, Moonshot AI released Kimi K3 at 2.8 trillion parameters, the largest open-weight model publicly released to date, with full weights published on July 27 under a modified MIT license. DeepSeek’s V4-Flash is on Hugging Face under MIT license at $0.14 per million input tokens and $0.28 per million output. That last number is worth sitting with. Claude Sonnet 5 goes to $3 input and $15 output on September 1.
Open models are competitive, though not uniformly ahead. I want to be careful here because the popular narrative overstates it. The Stanford AI Index 2026 found the leading closed model ahead of the leading open model by 3.3 percent as of March 2026, which is actually a wider gap than the 0.5 percent measured in August 2024. So the gap did not close in a straight line; it narrowed, then widened again. What is true is that the gap is small enough that for most small business tasks, writing, summarizing, drafting, analyzing, the difference is not something you will feel. Tom’s Hardware reported Kimi K3 outperforming leading closed models in the Frontend Code Arena benchmark, which is one benchmark, not a coronation, but it tells you the competition is genuine.
And here is the number that reframes the whole conversation. Multiple 2026 AI implementation cost analyses put labor and integration at 60 to 75 percent of total AI project cost, with API and model fees typically landing between 5 and 20 percent. The model bill is almost never your biggest line item.
Read that one more time, because it inverts the usual advice. If model fees are a small slice and labor is the big slice, then the expensive part of a deprecation is not paying more per token. It is the hours you spend rebuilding. Which means the highest-return investment you can make is not chasing the cheapest model. It is reducing the labor cost of switching to zero.
That is what portability buys you. Not a cheaper bill. A cheaper move.
The Four Layers You Must Own
Here is the framework I use, and it is deliberately simple because complicated systems do not get maintained.
Every AI workflow you run has four layers. Three of them belong to you. One of them does not, and never will.
Layer one is intent. This is the plain-language description of what the workflow is supposed to accomplish and what “good” looks like. Not the prompt. The purpose. “Turn a discovery call transcript into a one-page proposal outline that a client can approve without a follow-up call.” Intent is model-agnostic by nature. It never expires.
Layer two is context. These are the reference materials the model needs in order to do the job the way you want it done: your brand voice notes, your service descriptions, your pricing structure, your examples of past work that hit the mark. This is your most valuable and most commonly trapped layer. When people say “my custom GPT knows my business,” what they mean is that their context files are locked inside a vendor product.
Layer three is process. The actual sequence. What you feed in, in what order, what you check after each step, what you do when the output is wrong. Written as if you were handing it to a new hire. Because functionally, you are: every new model is a new hire.
Layer four is the model. This is the only layer you do not own and should not try to. It is a component. It is supposed to be swappable. When you treat the model as the workflow, you have built your house on rented land.
The proof case for this framework is happening in public right now. Look at what open weights actually did to the switching decision. Before mid-2026, moving off a major vendor meant moving to another major vendor with the same lock-in shape. Now Moonshot AI has published 2.8 trillion parameter weights under a modified MIT license and DeepSeek has V4-Flash on Hugging Face under MIT. You do not need to run those yourself, and I would not recommend most small business owners try. But their existence means dozens of hosting providers can serve them, which means the price and the access are no longer set by one company’s roadmap. That is what changed. Portability went from an ideal to an available option.
As for tools, you do not need anything exotic. A folder of plain markdown or text files in Google Drive, Dropbox, or Notion will do it. Plain text is the point. Plain text opens anywhere, pastes into anything, and will still be readable in ten years. Fancy formats are just a smaller version of the same lock-in problem.
How To Make Your AI Workflows Portable This Week
Seven steps. None of them require a developer. Budget about three hours total.
1. List your live workflows. Write down every repeated task where you currently use AI, from proposal drafts to social captions to meeting summaries. Most small business owners find between four and ten. You cannot protect what you have not named, and the list itself is usually shorter and more manageable than people fear.
2. Create one folder with one file per workflow. Name it something obvious like “AI Workflows.” Inside, one plain text or markdown file per task. Keep it in whatever cloud drive you already use so it syncs and backs up without you thinking about it.
3. Copy your working prompts out of the tools right now. Go into your chat histories, your custom GPTs, your saved projects, and paste every prompt you actually reuse into the matching file. Do not edit them yet. Just get them out. This single step takes about forty minutes and eliminates most of your exposure.
4. Write the intent line above every prompt. One sentence describing what the workflow accomplishes and what a good output looks like. This is what lets you rebuild the prompt for a new model instead of hoping the old text still works, because prompts are model-specific but intent is not.
5. Gather your context files into the same folder. Brand voice notes, service descriptions, pricing, your three best example outputs. If a custom assistant currently “knows” something about your business, that knowledge needs to exist as a file you own. Re-uploading files to a new tool takes minutes; reconstructing them from scratch takes days.
6. Document the process steps in plain language. For each workflow, write the sequence as numbered steps a new hire could follow. Include what you check before you send the output to a client. This is the layer people skip, and it is the layer that actually costs the most to rebuild.
7. Run one deliberate portability test per quarter. Pick one workflow, take your files, and run it on a different model than usual. Compare the output. You are not looking for a winner. You are confirming that your documentation is complete enough to travel, and finding the gaps while nothing is on fire.
Frequently Asked Questions
Do I need to switch away from ChatGPT or Claude to do this?
No. Portability is not about leaving a vendor. It is about being able to leave without pain. Keep using whatever works best for you today. The point is that your prompts and context live in your own files, so the choice stays yours instead of being made for you by a sunset date.
How do I know when a model I use is being retired?
Every major provider publishes deprecation pages and release notes; OpenAI and Anthropic both maintain them. Subscribe or check monthly. That said, monitoring is a weak defense on its own. Build portable and the notice becomes information rather than an emergency.
Should I be running open-weight models myself to save money?
Almost certainly not. Self-hosting means infrastructure, security, and maintenance work that will cost you more in labor than you save in fees. The value of open weights for a small business is competitive pressure and hosted access through third parties, not do-it-yourself deployment.
Is a cheaper model going to hurt my output quality?
Sometimes, and it depends on the task. Stanford’s AI Index put the top closed model 3.3 percent ahead of the top open model as of March 2026. For drafting, summarizing, and routine analysis, most people will not notice. For your highest-stakes deliverables, test before you switch anything.
What if I am not technical enough to manage all this?
You do not need to be. Everything in the seven steps above is copying text into files and writing plain sentences about how you work. There is no code, no API, no configuration. If you can write a training document for a new employee, you can do this.
Your Process Is The Asset. Act Like It.
Come back to the calendar for a second. August 9, August 26, August 30, August 31. Four dates in one month where something you might depend on changes or disappears. None of those were your decision. None of them were announced with you in mind.
That is going to keep happening. Not because anyone is out to get you, but because this industry is moving faster than its own products can hold still. The model that is best this quarter will not be best next quarter, and the interface you love will eventually get folded into something else. That is the terrain now. Getting angry about it is not a strategy.
So here is the most direct thing I can say to you.
Stop measuring your AI setup by which tool you are using. Start measuring it by how long it would take you to rebuild if that tool vanished tonight. If the honest answer is more than a day, you have work to do this week, and it is not glamorous work. It is copying prompts into a folder and writing down what you already know. Three hours, tops.
You are not behind because you have not adopted the newest model. You are exposed because your best thinking is stored in somebody else’s product.
Get it out. Put it in a folder with your name on it. Then let the models fight it out underneath you, because that fight is not yours to win.
The tools are rented. The process is yours.
About the Author
Jonathan Mast is the Founder and CEO of White Beard Strategies, where he provides AI coaching and mentorship to entrepreneurs and small business owners who want practical systems instead of hype. He is the creator of the Perfect Prompt Framework, a method for writing prompts that produce consistent, business-grade output regardless of which model you run them on, and he speaks regularly on applied AI for small business. Jonathan has spent years building operating systems for businesses that had to survive their founder taking a vacation, which turns out to be excellent preparation for building AI workflows that survive their vendor changing the rules.
