If you priced an AI automation eight months ago and passed on it because it cost too much or refused the task, this article answers whether that decision is still valid, and how often you should re-test it.
A client told me last month that she had already looked into AI for her intake process. She said it like a closed door.
I asked when. January.
She had priced it, decided it was too expensive for the volume she runs, and moved on. That was a reasonable call in January. The problem is that she was still living inside a January quote in August, and nobody had sent her a correction.
Here is the direct answer: yes, you should re-price it, and the odds are meaningfully in your favor. Independent trackers show the cost of a fixed level of AI capability falling somewhere between roughly 10x and 900x per year depending on the task, and three separate pricing moves landed in the first days of September 2026 alone. Your old quote is almost certainly stale. So is your old list of things AI refused to do.
The expensive mistake is not choosing the wrong tool. It is treating a verdict as permanent when the thing you judged keeps changing underneath it.
I know something about permanent verdicts. I have been bankrupt. I have been to prison. I spent years operating off a sentence I had written about myself that had stopped being accurate long before I bothered to check.
That is the same failure mode, just cheaper when it happens to a software decision.
Key Takeaways
- The cost of reaching a fixed level of AI capability has fallen between roughly 10x and 900x per year depending on the task, according to a16z and Epoch AI, which means any pricing decision older than a quarter deserves a re-test.
- Three pricing changes landed in early September 2026 with almost no mainstream coverage, including a 75 percent cut to Claude Fable 5.1 cache reads and up to 66 percent lower cost for Gemini video analysis.
- Not everything is getting cheaper: Artificial Analysis measured Fable 5.1 at 20 percent more per task than Fable 5 at max effort, and Google's Gemini Flash promotional rates double on January 1, 2027.
- Refusals move too. Anthropic reports its biology safeguards now fire about 85 percent less often on benign questions, so work that used to get declined may now run.
- Keep a written document of every automation you rejected on cost and every task AI refused, then re-test the list every quarter.
The problem is that we file our judgments and never reopen them
You evaluated a tool. You did the math. You said no.
That was work. Real work. You read the pricing page, you estimated volume, you maybe ran a test, and you reached a conclusion. Reaching a conclusion feels like finishing.
And then the conclusion becomes furniture. It sits in your head as a fact about the world rather than as a measurement taken on a specific Tuesday.
This is not a character flaw. It is one of the best documented patterns in decision research. William Samuelson and Richard Zeckhauser's 1988 paper "Status Quo Bias in Decision Making," published in the Journal of Risk and Uncertainty, ran experiments on portfolio allocation, retirement plans, and job choices, and found that people disproportionately chose whichever option was labeled as the one currently in effect (Samuelson and Zeckhauser, 1988). The label alone moved the decision.
McKinsey put a number on what that looks like inside a company. In an October 2025 piece, they describe a healthcare company that built a full self-paced generative AI training program. More than 90 percent of attendees rated it highly. Six weeks later, fewer than 10 percent had actually adopted the tools in daily work. Some had not even tried them, including people whose jobs would have gotten easier (McKinsey, "Bias Busters: The cognitive brakes on the gen AI accelerator").
McKinsey's framing is the part I keep coming back to. Status quo bias makes us overweight the risks of adopting a tool and underweight the risks of inertia.
Inertia does not send you an invoice. That is exactly why it is expensive.
Candidly, the small business numbers back this up. The U.S. Census Bureau's Business Trends and Outlook Survey found overall AI use among American businesses hovering between 17 and 20 percent from December 2025 through May 2026. Use rose among firms with at least 20 employees. Among firms with fewer than 20 employees, it did not change significantly, and fewer than 20 percent of firms with four or fewer employees reported using AI at all (U.S. Census Bureau, May 2026).
The smallest businesses are the ones sitting still. Not because they are less capable. Because nobody is paid to walk into their office and tell them the price changed.
The evidence: prices moved four times while you were not looking
Start with the long trend, because the long trend is the part that should change how you operate.
Stanford's Human-Centered AI Institute reported in its 2025 AI Index that the cost of running a model performing at GPT-3.5 level on MMLU fell from $20.00 per million tokens in November 2022 to $0.07 per million tokens by October 2024. That is a drop of more than 280-fold in about 18 months (Stanford HAI, 2025 AI Index Report).
Epoch AI measured the same phenomenon across six benchmarks and found the price of hitting a fixed capability milestone falling between 9x and 900x per year, with the price to match GPT-4 on PhD-level science questions falling about 40x per year (Epoch AI, March 2025). Andreessen Horowitz calls the broader pattern LLMflation and pegs it at roughly 10x per year for equivalent performance, noting that a GPT-3 quality model cost $60 per million tokens in November 2021 and about $0.06 by late 2024 (a16z).
Now the last two weeks. On September 1, 2026, Anthropic released Claude Fable 5.1 and cut the cache read price from $1.00 to $0.25 per million tokens, a 75 percent reduction, while leaving input at $10 and output at $50 (Anthropic). Anthropic estimates that reduces typical workload cost by about 25 percent and highly agentic workload cost by up to about 45 percent, as reported by VentureBeat.
Google shipped agentic video understanding in Gemini, where the model inspects only the relevant clips instead of ingesting the whole file, cutting token use by up to 88 percent and cost by up to 66 percent (Google). Alibaba shipped an upgraded Qwen3.8-Max on September 2 and held pricing flat at $2 per million input and $6 per million output, which is a free capability upgrade for anyone already on it, as reported by CellCog.
And the refusals moved. Anthropic reports that Fable 5.1's cyber safeguards produce roughly 60 percent fewer interventions per Claude Code session, and that its updated biology safeguards fire about 85 percent less often on benign elementary biology and medical questions, a change that applies to Fable 5 as well (Anthropic). If a health coach or a security consultant got stonewalled last spring, the wall moved.
Here is the part most coverage skipped, and I am not going to skip it.
Not everything got cheaper. Artificial Analysis, which does independent benchmarking, found that Fable 5.1 at max effort costs $3.76 per task on their Intelligence Index, which is 20 percent more than Fable 5 at $3.14, because the new model burns roughly 1.7x the output tokens. The cache cut saves about $1.40 per task; without it the number would have been about $5.16 (Artificial Analysis).
Digital Applied's teardown makes the same point about who actually benefits: the per-token list price did not move at all, only the cache read line item, so a workload with no reusable prefix saves exactly nothing (Digital Applied).
And prices go up too. Google's published pricing lists Gemini 3.6 and 3.7 Flash at $0.75 input and $3.75 output per million tokens through December 31, 2026, then $1.50 and $7.50 starting January 1, 2027 (Google AI pricing). That is a doubling, on the calendar, in writing.
So the honest claim is not "AI is getting cheaper." The honest claim is "AI pricing is volatile in both directions and your snapshot expires."
What I actually do about it: a rejection log
I keep a document. It is not clever. It is a table.
Every time I decide against an automation because of cost, and every time a model refuses to do something I legitimately needed, it goes in the document. Date, task, tool tested, the specific reason for the no, and the number that would have made it a yes.
That last column is the one that does the work.
Most people record the decision. Almost nobody records the threshold. If your note says "too expensive," you have to redo the entire evaluation to know whether anything changed. If your note says "would work at under $40 a month at 900 transcripts," you can check that in ninety seconds.
I learned to think this way from watching operations people run reorder points. You do not re-evaluate your whole supply chain every week. You set a trigger and you watch the trigger. Same idea, applied to your own past judgments.
The second column that earns its space is the refusal log. It is separate from cost, because it fails differently. A cost no is arithmetic. A refusal no is policy, and vendor policy changes on its own schedule with almost no announcement. Anthropic's 85 percent reduction in biology false positives is a real example of a category of work that used to bounce and now runs.
Then I calendar it. First week of January, April, July, October. Ninety minutes. I open the document, I re-price the top five rows, and I close it.
Three things happen at that meeting, reliably.
Some rows flip to yes. Those go straight into the build queue.
Some rows are still no, and I update the threshold number, which takes two minutes because I wrote down what I was measuring.
And some rows turn out to be the wrong question entirely, because the thing I wanted to automate got absorbed into a tool I already pay for. That has happened to me more than once, and I would have missed it if I were not looking on purpose.
The system is not the document. The system is that the document has a review date. A rejection log nobody reopens is just a nicer way to be wrong.
Practical steps: build the log this week
1. Set up the document in the next twenty minutes. A Google Sheet is fine. Columns: date, task, tool tested, no because cost or no because refusal, the specific number or the specific refusal message, and the threshold that would flip it. Do not make this pretty. Make it exist.
2. Backfill from memory, then from your inbox. Search your email and your notes for the last twelve months of tools you demoed and did not buy. Search for "pricing," "trial ends," and the names of vendors you remember evaluating. Ten rows is plenty to start.
3. Write down the threshold, not just the verdict. For every row, finish this sentence: "I would do this if it cost less than X per month at Y volume." If you cannot finish it, you did not actually evaluate the tool, you reacted to a price tag. Go finish the evaluation.
4. Log every refusal the day it happens. When a model declines a task, paste the exact refusal text into the sheet. Vendors update safeguards quietly and describe the changes in developer notes, not press releases. Your log is the only record you will have.
5. Put four ninety-minute reviews on the calendar right now. First week of each quarter. Invite whoever runs operations with you. If it is only you, still put it on the calendar, because the whole point is that this cannot depend on remembering.
6. Re-test with a real task, not a demo. Take an actual piece of last week's work and run it. Measure cost per completed task, including the retries, not the advertised per-token rate. Artificial Analysis found a model that got cheaper per token and more expensive per task in the same release. Per-task is the only number that pays your bills.
7. Use a prompt to run the review so you are not starting from a blank page. Paste your sheet into your assistant of choice with this:
[The Job] Review my log of AI automations I previously rejected and tell me which ones are worth re-testing this quarter.
[The Background] I run a small business and I keep a running document of automations I decided against, along with the specific cost threshold or refusal that stopped me. Current AI pricing changes faster than my review cycle, and I want to catch the ones that flipped. I am not technical and I need plain language.
[The Deliverable] A ranked list of no more than five rows to re-test, each with the reason I should re-test it now, the specific number or capability I should verify, and a one sentence test I can run in under fifteen minutes. Flag any row where the economics probably got worse, not better.
[The Questions] Ask me any questions you have.
Frequently Asked Questions
How often should I re-test an AI tool I already rejected?
Quarterly is the right cadence for most small businesses. Given that Epoch AI measured price declines for fixed capability ranging from 9x to 900x per year, a twelve month gap is far too long, and monthly reviews create more overhead than they return. Ninety days catches the meaningful moves without eating your calendar.
Is AI actually getting cheaper across the board?
No. The cost of reaching a fixed capability level has fallen dramatically, but frontier models stay expensive and some prices rise. Artificial Analysis measured Claude Fable 5.1 at 20 percent more per task than its predecessor at max effort, and Google's Gemini Flash promotional pricing doubles on January 1, 2027. Direction varies by workload.
What does "cost per task" mean and why does it matter more than token price?
Cost per task is what you actually pay to get one finished piece of work, including retries, tool calls, and every token the model generates while reasoning. A model can cut its per-token price and still cost you more per task if it produces more output. Your bill tracks tasks, not tokens.
If a model refused my request last year, will it refuse it now?
Maybe not. Anthropic reports its updated biology safeguards fire roughly 85 percent less often on benign elementary biology and medical questions, and Fable 5.1's cyber safeguards produce about 60 percent fewer interventions per coding session. Refusal policies get retuned regularly. Re-test the exact request rather than assuming.
I am not technical. Can I really do this myself?
Yes. This is a spreadsheet and a recurring calendar entry, not engineering. The skill required is writing down what you decided and why, in enough detail that your future self can check it. If you can track invoices, you can track rejected automations.
The close
My client's January decision was correct in January.
That is the part worth sitting with. She was not careless. She did the analysis, reached the right answer for the conditions in front of her, and then the conditions changed four times without telling her.
You are almost certainly carrying one of these right now. A workflow you wanted automated. A price that did not work. A task the model declined, and the small embarrassment of asking twice, so you never asked again.
Go get it out of your head and put it on paper. Write down the number that would have made it a yes.
Then put a date on it.
Because here is the thing: you are not competing against people who are smarter than you or better funded than you. You are competing against people who reopen their own decisions on a schedule.
That is a habit, not a talent. Anyone can build it, and it costs you ninety minutes a quarter.
If you want the conversation about which of your rejected automations to reopen first, come find me. I would rather help you re-price a no than watch you defend one.
The judgment you filed away was true on the day you made it. Do not let it become a fact about your business.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he teaches non-technical entrepreneurs to use AI to amplify the skill and experience they already have. He runs the AI Insiders membership and live AI trainings, and speaks on AI for business owners. He keeps his own rejection log in a plain Google Sheet, and the oldest row on it flipped from no to yes about five months after he wrote it, which is the only reason this article exists.
Sources
- Stanford HAI, The 2025 AI Index Report
- Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks
- Andreessen Horowitz, Welcome to LLMflation
- Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1
- Anthropic, Improving Fable 5's biology safeguards
- VentureBeat, Anthropic's Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads
- Artificial Analysis, Claude Fable 5.1 tops the Artificial Analysis Intelligence Index
- Digital Applied, What Claude Fable 5.1 Costs, and What It Breaks
- Google, Introducing agentic video understanding with Gemini
- Google, Gemini API pricing
- CellCog, Qwen3.8-Max-0902: Same Price, Much Better at Coding and Office Work
- McKinsey, Bias Busters: The cognitive brakes on the gen AI accelerator
- Samuelson and Zeckhauser, Status quo bias in decision making, Journal of Risk and Uncertainty, 1988
- U.S. Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users
- AI Breakfast, Anthropic rolls out Fable 5.1 after sandbox escapes