Neither, at least not first. The real question this article answers is how to make your business able to switch AI tools in an afternoon, so that a vendor's pricing decision stops being your emergency.
When the Floor and the Ceiling Moved on the Same Day
On August 17, two things happened that most business owners never saw.
Alibaba released Qwen3.8-27B, an open-weight model built to run on a regular laptop, and put the weights out for anyone to download (Quartz). The same weekend, DeepSeek's new API pricing kicked in, raising costs across its V4 lineup by anywhere from 50 percent to more than 1,100 percent depending on the model and the hour of day you use it (InfoWorld).
The cost of running AI hit a new low and a new high in the same 48 hours. If that sounds like a technical story, it is not. It is a story about whether you have any leverage with the companies you now depend on to run your business.
Here is the direct answer. Do not switch tools because of a headline, and do not just absorb the increase either. The move is to make switching cheap. If your entire AI operation lives inside one vendor's chat window, with prompts you cannot find, workflows nobody wrote down, and a bill you cannot break apart, then that vendor sets your prices and you sign the check. If your prompts, your context, and your process live somewhere you control, a price change becomes a Tuesday afternoon decision instead of a crisis.
That is the thesis. Model portability is not a developer concern anymore. It is a business skill, and it is the cheapest insurance you can buy against a pricing decision made in a room you will never be in.
Key Takeaways
- A vendor's price increase is only an emergency if you cannot leave, so the fix is reducing the cost of leaving, not predicting the price.
- Your prompts, your business context, and your written process are the assets; the model is a replaceable part.
- Most owners cannot say what they spend on AI each month, which means they cannot tell whether an increase actually matters.
- Running models locally is now technically possible for small businesses, but it is a fallback option, not the strategy.
- Portability is tested, not assumed: if you have never moved a workflow to a second tool, you do not know whether you can.
The Problem Nobody Warned You About
You did not sign up for a vendor relationship. You signed up for a twenty dollar subscription that helped you write faster.
Then it became three subscriptions. Then your CRM added AI credits. Then your design tool bundled an assistant and raised the plan price. Nobody ever sat you down and said "you are now building your operations on top of a supplier whose pricing you do not control." It happened one free trial at a time.
I have been through this in my own business, and I will admit the uncomfortable part: for a stretch, I could not have told you what I was actually paying for AI in a given month. Not because I was careless, but because the spend was spread across seat licenses, usage credits, and bundled features inside tools I bought for other reasons. There was no single line item to look at. There was no moment where I decided to spend that much. It accumulated.
That is the trap. The bill is invisible until it is not, and by then your business has muscle memory. Your team knows one interface. Your prompts live in one chat history. Your best workflows exist as things people "just know how to do." None of that is written down anywhere you could hand to a different tool.
And this is genuinely hard. You are running a business. You do not have a procurement department, a FinOps analyst, or an engineer whose job is to keep options open. The vendors are not being evil, either. DeepSeek said plainly that it raised prices to allocate strained compute more reasonably. Demand is real. Capacity is real. Prices are going to move again.
So here is the reframe. Stop trying to pick the winner. You will not out-forecast a market where the cheapest capable model changes every few months. Instead, get good at leaving. The businesses that will be fine in eighteen months are not the ones that chose correctly. They are the ones that can change their mind in an afternoon.
What the Evidence Actually Shows
This is not a vibe. The numbers are unusually clear.
Almost nobody is hitting their AI budget. An independent survey conducted by Sapio Research in February 2026, commissioned by DoiT, covered 500 finance leaders at US and UK organizations with 1,000 or more employees. Every single one of those organizations was already spending on AI. Seventy-nine percent had experienced AI cost overruns in the previous twelve months, and only 15 percent could calculate AI ROI without significant bottlenecks (DoiT). These are companies with finance departments. You are competing with that level of visibility using a credit card statement.
Mid-size companies got hit harder than the giants. In the same survey, organizations with 1,000 to 4,999 employees overran their AI budgets more often than large enterprises, 81 percent against 76 percent, despite running smaller budgets. Smaller operations have less cushion and less instrumentation. Scale down further to your business and the pattern does not improve.
Even Uber blew it. Uber burned through its entire 2026 budget for AI coding tools in four months and then capped employee spending, with its chief operating officer saying the costs were getting harder to justify (TechCrunch, June 2026). Anthropic subsequently built budget alerts and model access controls into Claude Enterprise, which tells you what customers were asking for. Vendors do not ship spend caps for customers who are spending carefully.
Waste is rising for the first time in five years. Flexera's 2026 State of the Cloud Report, based on more than 750 cloud decision-makers, found wasted cloud spend climbed to 29 percent, reversing a five-year decline. Flexera attributed the reversal directly to the surge in AI workloads and new services (Flexera). The people whose full-time job is controlling this spend are losing ground.
The professionals are treating this as a new discipline. The FinOps Foundation's State of FinOps 2026 report surveyed 1,192 practitioners and found 98 percent now manage AI spend, up from 31 percent two years earlier, with AI cost management ranked the number one skill teams are trying to hire (Linux Foundation). When a skill goes from a third of practitioners to nearly all of them in twenty-four months, that is a category being invented in real time.
And the market is voting for portability with money. OpenRouter, a service that lets you send the same request to any of 400 or more models through one connection, raised a $113 million Series B in May 2026 led by Alphabet's CapitalG at a $1.3 billion valuation, with NVIDIA, Snowflake, MongoDB, and Databricks ventures participating (TechCrunch). Sophisticated investors are betting that the ability to move between models is worth more than loyalty to any one of them.
Put those together. Costs are unpredictable, visibility is poor, and the smartest capital in the industry is funding the exits. Plain language version: the professionals assume they will need to switch, and they are paying to keep that option open.
The Shift: Own the Asset, Rent the Engine
Here is what changed on August 17.
Before, "just use a local model" was a hobbyist answer. Running a capable model on your own machine meant compromise: slower, dumber, not worth it for real work. When Alibaba shipped a 27 billion parameter open-weight model designed to run on consumer hardware, that stopped being automatically true. Meta had entered the same on-device market the week before. There is now a credible floor under your costs that does not require anyone's permission.
But I want to be careful here, because the wrong lesson is loud right now. The lesson is not "switch to local models." For most small businesses, running your own model is more trouble than it is worth. You would trade a predictable bill for an unpredictable Tuesday.
The lesson is that the model is no longer the thing you are buying. It is the engine. Engines are now abundant, they are getting cheaper and better, and several of them will do your job acceptably.
What you actually own, or should own, is different:
Your prompts. The specific, tested instructions that produce output you would actually send to a client. That is intellectual property. It should live in a document you control, not scattered through a chat history inside a vendor's app.
Your context. The description of your business, your customers, your voice, your offers, your constraints. Write it once, well. It is the single most valuable thing you can hand any AI tool on day one, and it works in every one of them.
Your process. Which steps a human does, which steps the AI does, what "good" looks like, and how you check it. Documented. Boring. Portable.
Own those three, and the engine underneath becomes swappable. A price increase becomes a comparison, not a crisis. A better model launching becomes an upgrade you can take in an afternoon instead of a migration you keep postponing.
This is also why the practitioner behavior you see in AI communities right now matters. People are building their own homemade usage dashboards because their vendors will not give them a clear, readable bill. That is not paranoia. That is what happens when customers realize they are exposed and the supplier has not given them a gauge. You should not have to build a spreadsheet to know what you spent. But until you get one, build the spreadsheet.
Seven Steps to Make Switching Cheap
1. Find your actual number. Pull the last three months of bank and credit card statements and list every charge connected to AI, including AI features bundled into tools you bought for other reasons. Add it up. Most owners are surprised in one direction or the other, and either surprise is useful.
2. Write your context file. One document, one to two pages: what your business does, who you serve, how you talk, what you sell, what you never say. This is the asset that makes every AI tool useful on day one instead of day thirty. Save it outside any AI vendor's platform.
3. Pull your top ten prompts out of the chat window. Find the ten AI tasks you actually repeat, and copy the working instructions into a document with a plain name for each. If you cannot find them, that is the finding. Rebuild them once and store them properly.
4. Run a swap test on one workflow. Pick a single low-risk task, take the prompt and context file, and run it in a second tool you do not currently use. Compare the output honestly. You are not looking for a winner. You are measuring how long it took and what broke, which is your true switching cost.
5. Set a cap before the next tool enters the business. Decide the monthly ceiling for AI spend across the whole business, write it down, and use vendor spend limits where they exist. A number you set in advance is a decision; a number you discover in arrears is a bill.
6. Rank every AI tool by how hard it would be to leave. Ask one question per tool: if this doubled in price tomorrow, could I be off it in a week? Tools where the answer is no are where your risk lives, and those deserve either a documented exit plan or a second option already tested.
7. Put a calendar reminder on it. Ninety days out, re-run steps one and four. Pricing will have moved. New models will have shipped. This is a quarterly review that takes an hour, not a project.
Here is a prompt to run step six. Fill in the brackets with your own details.
[The Job]
Analyze my business's dependency risk across the AI tools I currently pay for and tell me where I am most exposed to a price increase.
This is for: [the owner of a [INDUSTRY] business with [NUMBER] employees].
It matters because: [a sudden vendor price change should not be able to disrupt our operations or blow our monthly budget].
[The Background]
Here is what you need to know: we currently use these AI tools and pay roughly these amounts each month: [TOOL 1, COST, WHAT IT DOES]; [TOOL 2, COST, WHAT IT DOES]; [TOOL 3, COST, WHAT IT DOES]. The workflows that depend on them are: [WORKFLOW 1], [WORKFLOW 2], [WORKFLOW 3]. Our monthly ceiling for total AI spend is [DOLLAR AMOUNT]. The workflows we cannot afford to have interrupted are: [CRITICAL WORKFLOW LIST].
Do not use: generic advice about "AI strategy," recommendations to hire technical staff, or suggestions that require writing code.
[The Deliverable]
Return: a ranked table of my tools from hardest to easiest to replace, followed by a one-page action list.
Must include: for each tool, the specific thing that locks me in, one named alternative I could test, and an estimate of how many hours a switch would take. Written from the perspective of an operations consultant speaking to a non-technical business owner.
Optimize for: accuracy.
[The Questions]
Ask me any questions you have.
Frequently Asked Questions
Do I need to run AI on my own laptop to save money?
No. Local models are now genuinely capable, but running one is a technical project with real setup and maintenance costs. Treat it as a fallback that proves you have options, not as your main plan. For most small businesses, the savings come from right-sizing subscriptions and matching cheaper tools to simpler tasks.
How do I find out what I am actually spending on AI each month?
Start with three months of bank and credit card statements rather than vendor dashboards, because bundled AI features inside your CRM, design, and email tools rarely show up as AI line items. Add every charge to one sheet. Vendors are not obligated to show you a combined number, so you have to build it yourself.
Will my prompts still work if I switch from one AI tool to another?
Mostly, with adjustment. The structure carries over cleanly: the job, the background, the deliverable, the questions. What changes is tone and defaults, since each model has its own habits. Expect to spend twenty minutes tuning a prompt after a move, not rebuilding it from scratch.
Is it safe to use an open-weight model like Qwen in my business?
Two different questions live in there. Downloading open weights and running them on your own hardware means your data never leaves your machine, which is the strongest privacy position available. Sending data to any vendor's hosted API, from any country, is a separate decision governed by their terms and your obligations.
How often do AI vendors actually change their prices?
Often enough that you should plan for it. DeepSeek moved its API pricing twice inside four months in 2026, and Anthropic raised prices earlier that year citing the same capacity pressure. Assume at least one meaningful pricing change per year from any vendor you rely on, and review quarterly.
The Real Point
Two things moved on the same day in opposite directions, and almost nobody running a small business noticed either one.
That is not a criticism. You have customers to serve and payroll to make. Nobody expects you to track Chinese model releases or the fine print of token pricing. But the pattern underneath those two events is going to keep repeating, and it will keep repeating on days when you are busy.
So do not track the news. Build the thing that makes the news irrelevant.
Write down your context. Pull your prompts out of the chat window. Test a second tool once, on something small, so you know what switching actually costs you. That is a few hours of work. It permanently changes your position from dependent to portable.
Here is the direct version, and I mean this plainly: if you cannot leave a vendor, you are not their customer. You are their inventory. Every pricing decision they make is a decision they get to make about your business, and you will find out about it in an email.
The good news is that this is fixable, and it is fixable this week, without hiring anyone or learning to code. The cost of switching is the only number in this whole conversation you actually control. Go make it small.
Own the prompts, rent the engine.
About the Author
Jonathan Mast is the founder of White Beard Strategies, an AI coaching and mentorship brand that helps entrepreneurs and small business owners put AI to work without turning it into a second job. He is the creator of the Perfect Prompt Framework, a speaker, and a working operator who tests everything he teaches inside his own business first. White Beard Strategies runs live trainings and a member community focused on practical AI systems for people who own the business, not the IT department. Training replays and membership details are available at whitebeardstrategies.com.
