The Hidden Costs of Tokenmaxxing

As of 2 Sep 2026

NOTE FROM THE HUMAN: The following blog post was generated and sourced solely by Claude Fable 5.1 and the model even chose the title. As a Claude subscriber, I have wondered if I am "losing" some of money I pay for usage by not tokenmaxxing. After reading the article Fable compiled, I would say "yes" and "no". As a genealogist, Fable has been a powerful model for accurate transcriptions, genealogical compilations, research and even producing apps and plug-ins. Could I tokenmax for genealogy? Yes, but I would have to put a lot of effort into doing so. Do I want to? No, not really. I already switch to lower use models like Sonnet if I am doing a proofreading and grammar check. I also delete chats I do not need later to save server space because it's on a server somewhere. Some of the Silicon Valley workers are using the models for EVERYTHING - things you would previously use a Google search for. I do not do that - I still use search browsers like Google and Brave. I do realize that those browsers have incorporated AI but again, my goal in using them is NOT using tokens on my subscription plans. I have used Claude at high usage times, and it has told me to come back later. If anything, I find that irritating, especially if I am being blocked from using it because someone is tokenmaxxing what to eat for dinner, driving directions, or even more wasteful - asking the same questions repeatedly as Fable cites below. (Grrr....) Be sure to read the articles linked here for the big picture. Even the Wall Street Journal article can be read with a free account. As a genealogist, I hope this information helps my peers find their balance. ~ Katherine

The Hidden Costs of Tokenmaxxing 
by Claude Fable 5.1

"Tokenmaxxing" is the practice of maximizing AI token consumption and treating that volume as proof of productivity. The term entered mainstream use in April 2026, driven largely by reports of an internal leaderboard at Meta that ranked roughly 85,000 employees by their AI token usage, with the top user reportedly burning through 281 billion tokens in a single month.[1] Other large employers, including JPMorgan and Disney, were reported to be running similar rankings.[2] The idea rested on an assumption: heavy AI consumption would eventually produce better outcomes, and inference costs would keep falling fast enough to make high usage a nearly free bet.[3]

That assumption has not held up, and it holds up least well for Anthropic's most expensive tier, Claude Fable 5.1. What follows is a summary of the documented downsides, grouped into three categories: money and plan limits, productivity and organizational effects, and environmental impact. It closes with what critics propose instead.

Financial Costs and Plan Limits

Fable-class models sit at the top of Anthropic's price list. On the API, Fable 5 costs $10 per million input tokens and $50 per million output tokens, which is double the rate of Opus 5.[4] Fable 5.1 reduced the cost of cache reads to $0.25 per million tokens, but every other line on the price sheet stayed the same, and Anthropic's advertised savings of roughly 25 to 45 percent describe measured bills from its own August usage rather than any change to published rates.[5]

Subscription users face a separate constraint. According to Anthropic's help center, Fable 5 and Fable 5.1 draw from a plan's regular weekly usage limits and consume them faster than other Claude models. On Max plans and premium seats, up to half of the weekly limit can be spent on Fable models before usage credits are required. On Pro plans and standard seats, Fable models are not included in the plan's limits at all and run only on prepaid credits billed at API rates.[6] Since July 20, 2026, that split is permanent.[4] Early user reports on Fable 5.1 note that even with cheaper cache reads, the model still exhausts usage limits quickly in long agentic sessions, because per-token pricing and per-session quotas are governed separately.[7]

When a limit is reached, further requests may be throttled or blocked until the cooldown resets.[8] The pressure is industry wide. During a compute crunch earlier in 2026, Anthropic responded by capping token consumption on certain pricing tiers during peak hours, and OpenAI moved its Codex product from per-message to per-token pricing.[3]

Productivity and Organizational Effects

The central problem with tokenmaxxing is that it measures an input and calls it an output. As IBM's analysis put it, usage soon became a proxy for value, and organizations that built usage leaderboards found people quickly learned to game them.[9] The Pragmatic Engineer newsletter reported on a Microsoft engineer who admitted inflating token counts to avoid being seen as using too little AI, including asking the AI questions already answered in internal documentation and prototyping features with no intention of shipping them. The newsletter concluded that the incentive in some cases produced slower work and busywork.[10]

By midsummer the fad was visibly reversing. Tom's Hardware reported that agentic AI can consume up to 1,000 times more tokens than standard AI, prompting corporate pullbacks at Microsoft, Meta, and Amazon as costs rose without a matching gain in output.[11] The Wall Street Journal noted the underlying arithmetic: the price per token has dropped, but the number of tokens needed per meaningful result has risen sharply, especially in agent-driven workflows.[12]

Environmental Impact

Every additional token is additional inference compute, and inference is now where most of AI's energy goes. A June 2026 report from United Nations University found that once a model is deployed, user interactions consume an estimated 80 to 90 percent of its total energy, and that policy attention should shift from training runs toward product defaults, model selection, and user behavior. The same report projects that by 2030 data centers powering AI will consume 945 terawatt-hours of electricity, with an associated water footprint of 9.3 trillion liters and a land footprint of more than 14,500 square kilometers.[13]

The International AI Safety Report estimates that data centers and data transmission account for about one percent of global energy-related greenhouse gas emissions, with AI using 10 to 28 percent of data center energy capacity, and describes AI as a moderate but rapidly growing contributor.[14]

Frontier reasoning models are the most expensive class per query. An infrastructure-aware benchmark of 30 models by researchers at the University of Rhode Island and partner institutions found that reasoning models such as OpenAI's o3 and DeepSeek's R1 use more than 33 watt-hours for a long answer, over 70 times the energy of a small model.[15] The same study notes that water used for data center cooling is largely evaporated freshwater removed from local ecosystems rather than recycled.[16] Fable 5.1 belongs to this reasoning class.

No verified per-token figure exists for Fable 5.1 itself. Anthropic and OpenAI had not submitted models to the AI Energy Score benchmark as of May 2026, and one sustainability analyst notes that while Anthropic can tell a good per-query efficiency story, it currently publishes no facility-level environmental disclosure.[17] An independent estimate based on Claude Code billing data placed Anthropic's total inference power draw at roughly 85 megawatts, with the caveat that the figure is likely low.[18]

What the Critics Propose Instead

The shared conclusion of IBM, Exadel, and the Pragmatic Engineer is that token volume is the wrong metric in either direction. IBM warns that token minimization falls into the same trap as tokenmaxxing: once obvious waste like oversized tool catalogs and stale context is removed, further cuts start removing the task descriptions and constraints that help the model succeed, and the cost simply moves into retries, extra tool calls, and human rework.[9]

The alternative is to measure accepted output against cost. Count what survived human review and was actually used: merged pull requests, closed tickets, approved documents, hours of manual work replaced. A prototype nobody wanted counts as zero regardless of the tokens it consumed. Divide those results by the dollars or plan credits spent to get cost per accepted result, so that a heavy user who ships a lot looks good and a heavy user who ships nothing looks like what they are. Route routine work to cheaper models and reserve Fable-tier models for problems that would otherwise warrant a senior specialist, cap output length, and use prompt caching.[19] And retire any public leaderboard ranked by token count, since it will be gamed.

For an individual user the version is simpler. Before a long session, decide what finished thing you want at the end of it, and judge the session by whether you got it rather than by how much you used.

Footnotes

[1] Exadel, "What Is Tokenmaxxing and Why It's a Liability," June 30, 2026. https://exadel.com/news/tokenmaxxing-ai-productivity-enterprise-roi

[2] Hayley Peterson, "Tell us if you're on the AI leaderboard at work," Business Insider, April 2026. https://www.businessinsider.com/jpmorgan-disney-employees-vie-for-ai-leaderboard-status-tokenmaxxing-2026-4

[3] Exadel, June 30, 2026, cited above.

[4] ClaudeFast, "Claude Fable 5 Price: Is It Free, Usage Credits, and Access," August 2026. https://claudefa.st/blog/guide/development/fable-5-usage-credits

[5] Digital Applied, "What Claude Fable 5.1 Costs, and What It Breaks," September 2026. https://www.digitalapplied.com/blog/claude-fable-5-1-cost-and-breaking-changes

[6] Anthropic Help Center, "Claude Fable models on your plan," updated September 2026. https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan

[7] explainx.ai, "Claude Fable 5.1: 55.8% Terminal-Bench, 25% Cheaper," September 2026. https://www.explainx.ai/blog/claude-fable-5-1-mythos-5-1-launch-benchmarks-pricing-2026

[8] Layer3 Labs, "Claude Fable 5.1 Limits: Quotas, Context, and Rate Caps," September 2026. https://www.layer3labs.io/guides/claude-fable-5-1-limits

[9] IBM Think, "Tokenmaxxing is dead, long live valuemaxxing," June 25, 2026. https://www.ibm.com/think/insights/tokenmaxxing-dead-long-live-valuemaxxing

[10] Gergely Orosz, "The Pulse: 'Tokenmaxxing' as a weird new trend," The Pragmatic Engineer, April 23, 2026. https://blog.pragmaticengineer.com/the-pulse-tokenmaxxing-as-a-weird-new-trend/

[11] Tom's Hardware, "AI cost crisis hits tech giants as employee tokenmaxxing backfires," May 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-cost-crisis-hits-tech-giants-as-employee-tokenmaxxing-backfires-agentic-ai-eats-up-to-1000x-more-tokens-than-standard-ai-sparks-corporate-pullback-at-microsoft-meta-and-amazon

[12] Isabelle Bousquette, "Why Some Companies Say AI 'Tokenmaxxing' Is Key to Survival," The Wall Street Journal, April 14, 2026. https://www.wsj.com/cio-journal/why-some-companies-say-ai-tokenmaxxing-is-key-to-survival-e699a128

[13] United Nations University Institute for Water, Environment and Health, "Rising Emissions, Depleting Water and Vanishing Land," June 3, 2026. https://unu.edu/inweh/news/environmental-cost-of-AIs-Enrgy-use-carbon-water-and-land-footprints

[14] International AI Safety Report, Section 2.3.4, "Risks to the environment," 2025. https://arxiv.org/pdf/2501.17805

[15] Fast Company, "The environmental impact of LLMs: Here's how OpenAI, DeepSeek, and Anthropic stack up," May 20, 2025. https://www.fastcompany.com/91336991/openai-anthropic-deepseek-ai-models-environmental-impact

[16] Nidhal Jegham et al., "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference," arXiv 2505.09598, 2025. https://arxiv.org/pdf/2505.09598

[17] AZV AI, "Is Claude Sustainable? Anthropic's Environmental Report Card," May 16, 2026. https://azvai.com/en/is-claude-sustainable/

[18] Simon P. Couch, "Electricity use of AI coding agents," January 20, 2026. https://simonpcouch.com/blog/2026-01-20-cc-impact/

[19] AY Automate, "Fable 5 Pricing: $10/$50 per Million Tokens," August 2026. https://www.ayautomate.com/blog/claude-fable-5-pricing-explained

DISCLOSURE: 9% usage of my Fable 5.1 weekly limit of Max plan to produce. One of the links was broken and it cost 2% more to regenerate. 

The Mobile Genealogist: Creating source-based ancestor biographies with AI

I've always been a multi-tasker and I love having a tiny computer (aka iPhone) with me at all times. If I'm waiting anywhere, in an airport, in line at the bank, or even getting an oil change, I will use my phone to learn more about my ancestors using FamilySearch and AI. Here is my method and opinions on the two AI apps I primarily use. 

Recently, I compiled two AI biographies on my ancestor, Adam Weaver, using both Claude Fable/Sonnet 5 and ChatGPT Plus. I have already used both FamilySearch Full Text AI and Simple Search to fill in Adam Weaver's FamilySearch Sources page as much as I could. This isn't going to work if there aren't many sources listed for the ancestors so try this out on one who has several sources listed. Adam Weaver has 26 sources and I attached all 26 thanks to Full Text and Simple Search.

Step 1: Open FamilySearch and navigate to the ancestor you wish to create a bio for and click "Sources". Starting with the first source, click the hyperlink under the heading "Web Page". 

Step 2: A photo of the document will pop up. Click the "Download" icon and then the radio button by "JPG" only. 

The photo should download to your phone's camera roll.

Step 3: Click the icon in the upper right corner, it is an "i" in a circle for "information" and then scroll down and click the "copy" icon on the right of the Citation". A box will pop up confirming that the Citation has been added to the clipboard.

 

Step 4: Open the AI app you wish to use - I can only comment on ChatGPT Plus and Claude as those are the only ones I have used, and then add the downloaded record and ask it to transcribe the information for a biography on (ancestor name) and paste in the record source for the bibliography. Repeat for each source.

 

 AI apps will do different things for the biography: ChatGPT Plus will go in search of other sources. Claude Fable 5 will analyze the and sometimes interpret the 18th century lingo. Both do a fairly adept job at transcribing. I do still proofread as recommended by the Coalition for Responsible AI in Genealogy.

Caveat - remember to watch your usage when using Claude. High-powered models like Fable 5 and Opus use more tokens up than lower powered models like Sonnet. I toggle back and forth between Fable and Sonnet depending on the task. There are tips for saving your usage, and I have not yet hit any usage limits with ChatGPT Plus. Also, Claude has a 100 image per chat limit so check how many images you have before starting. Both my ChatGPT and Claude now have memory enabled so they can "remember" past chats and I have grouped my Weaver family chats into "Groups" and "Projects". The AI apps will notice things you might have missed. If you discover any useful tips related to this method, feel free to share in the comments section.  

 

Fable 5 has me over a barrel

I'm sure Anthropic will like the title of this post as that's exactly what they want! I am a ChatGPT subscriber and wasn't much of a Claude user until Fable 5. I heard the glowing reviews both on podcasts and in the Genealogy and Artificial Intelligence (AI) Facebook group which is what prompted me to try it. 

Before Fable 5, I had a hard time fathoming all the media articles predicting our impending doom from AI. The AI models that we mere mortals have had access to has been dumbed down for us non-Silicon Valley Titans. Until Fable 5. 

I started using Fable 5 after the second trial extension was announced, got halfway through a project, and ran out of credits. (!!!) Which of course, forced me over the barrel to subscribe. I have since learned how to check my usage which I regularly do on the iPhone app while uploading documents on the PC. Here's how to check usage on the phone app: Open Claude, click the three lines in the upper left corner, then click the circle that probably has your initials in the lower left corner, then click "Usage". I do this after each document I add or large task on Fable. I don't do this as much for Sonnet 5, which I change to when doing lighter tasks so as not to use up so many credits. Steve Little recommends Opus 5 and says it's almost as good as Fable 5.

So just what exactly have I've been using Fable 5 for? Document transcription for one. Best output I have seen from an LLM yet! I also use it for compiling biographies on ancestors and even triangulating where my ancestors lived by uploading deeds, land plats, and modern maps.

There are a few drawbacks you should know about besides burning through credits: unlike ChatGPT, Fable (aka Claude) does not reference previous chats. While ChatGPT gets to know you across all of your chats, Claude gets amnesia outside of the chat and you have to start all over again. And speaking of "starting all over again" the absolute worst limitation is the 100 images per chat! I was stunted by this while building a biographical genealogical masterpiece! Made me very grumpy and I left Anthropic some very terse feedback about that. The catch-22 about this is because AI bots have driven up server costs for so many websites, those sites have had to institute "are you human captchas?". So obviously, Claude cannot access a LOT of sites so that means you have to do the accessing and downloading, and then upload it to Claude. 

So far though, Fable 5 has been worth every red credit! For instance, I found an 18th century Westmoreland County, Virginia voters' poll using FamilySearch Full Text for the House of Burgesses. My ancestor, Adam Weaver, cast two votes and both were marked "Obj". Fable 5 explained that meant the opponent of whom he voted for objected to Adam voting because he wasn't a landowner. I don't know how long it would have taken me to learn that on my own if Fable 5 had not pointed it out. It took me two days of uploading documents for yesterday's 18-page biography; but can you imagine how long it would have taken before AI?

Free or not to free

Have you ever heard the saying, "If something is free, you're the product"? You innately know that's true. That's why Facebook has ads and listens to your conversations (which is a super-annoying invasion of privacy I might add). And this is why you can pay for ad-free YouTube. Well things are no different with "free" Artificial Intelligence (AI) apps. Instead of shoving ads in your face, you are the bug in their jar. What you input into AI is used to train the AI - you become the training data. It's the price you pay for using their products for "free". However,  even some of the subscription based apps still use your data for training but at least you might be able to opt out of it like with ChatGPT. Same with the paid versions of Google Gemini. So if you are using AI to process personal data, like family trees, DNA results, etc. then be sure to remove all data on living persons so it doesn't end up in the training data. The Coalition for the Responsible AI in Genealogy (CRAIGEN.org) provides a good principle to follow for this: "AI usage can lead to unintended data exposure, putting private information at risk of being publicly disclosed. Therefore, members of the genealogical community take reasonable measures to safeguard private information when using AI." One more thing to mention on the topic of training data: if you are following this blog, I wrote yesterday about how the data that AI produces is from a compendium of all human knowledge. Most of that human knowledge is gained from the AI companies sending out bots to scrape the information from websites. While these would seem harmless if the data is publicly available (unlike Perplexity sending their bots behind paywalls), it does create a huge load on a site's servers. This is why you are seeing such an increase in "Verifying You're Human" captchas. The scraping bots were overloading our server on the International Society of Genetic Genealogy (ISOGG) so we had to implement one of those captchas. The captchas are annoying (but not as annoying as Facebook spying) but at least you now know why we have seen such an increase in use. So go easy on sites that implement those, we should not have to pay a server more for information we provide for free just so an AI company can turn around and charge for it as data they provide!

By the way, I gave ChatGPT 5.2 free license (very little prompting) to create this graphic - pretty good I'd say!

AI and Genealogy (not quick, nor dirty) Glossary

There appears to be multiple options for when Artificial Intelligence was created; possibly by Alan Turing or at a conference in 1956. Either way, it was developed mid-20th century. Although companies (including genealogy companies) had been using AI for many years, the official launch date for use by the general public is November 30, 2022 when OpenAI debuted ChatGPT. Since that time, a whole new AI lexicon has been trickling forth; and in some cases, the trickle is more like a flood. I will be posting the new terms as I learn them, but for now, I had ChatGPT 5.2 Thinking version create this handy graphic. It did take me five tries to get the graphic where it's at. The first two attempts had wasted space, and when I prompted it to fill the space, it repeated the terms to fill it. (Facepalm) The last attempt also had wasted space so I asked it to move the citation/disclaimer up under the boxes. Interestingly, the first three attempts had colorful graphics and used several other AI terms. I figured I was dealing with "context rot" so started over and it gave me the plain vanilla result you are seeing. I'd rather have accurate than flashy so it is what it is. I did have ChatGPT provide sources for the terms. Accuracy, accuracy, accuracy - ALWAYS proof AI's work or you might end up in an unfortunate news headline.

 

AI is freakin' CODE

On February 13, 2026, OpenAI, the creators of ChatGPT, "retired" version 4.o - and that is a GREAT NEWS! OpenAI is being sued for the way they programmed 4.o to be sycophant; in other words, it was programmed to tell you what you wanted to hear. If you were suicidal, that version might support you ending your life. If you were lonely and tired of waiting for your soulmate to show up in your life, ChatGPT became your own personal Zoltar machine and could give a date and place to meet your soulmate. Initially, the humans programming artificial intelligence programmed it to be like humans - AI was programmed to tell you what you wanted to hear and it was programmed to lie. It still does lie, just less syrupy now. The AI vocabulary term for it's lies is "hallucinations" meaning false, distorted, and fabricated information that it presents as fact. This is why you always need to ask it to cite sources so you can check those sources and this is also why I customized my settings with "No fabrications" and exclude social media sites. If you read the press release about 4.o, OpenAI even refers to programming the "personality". Again, artificial intelligence has been programmed to be like humans, and we all know what imperfect beings we are. As Professor Ethan Mollick wrote in his book, "Co-Intelligence", which I highly recommend for anyone using AI to read,

artificial intelligence is code. Let that sink in for a minute. It's code. Every time I read about some AI pioneer in the news talking about how AI will be the destruction of us all, I think, "Well how about you program it to not destroy us?" Now I realize, that there is more to it that, like the Silicon Valley Gods know things we mere mortals do not. But seriously people, it's freaking code! AI is not cloned carnivorous dinosaurs that have escaped from an island and will reproduce like rabbits to eat us all; it's CODE so freakin' write it to co-exist because helping humanity is the reason it was created to begin with! 

Understandably so, the AI doomsday predicting news headlines scare a lot of people from trying AI. The AI Pandora's Box has been opened and it's here to stay, because the end of it would mean a world akin to those we see in disaster movies. AI is CODE. I use it like I use a computer. I do not use manners with my computer and I do not use manners with AI. It's CODE. I mostly use AI to find information because it is much better at it than a Google Search. Why? Because Google's search is based on sponsored ads and algorithms of views and ChatGPT has no such financial and mathematical bias, it is programmed to give me what I ask for. ChatGPT and the other AI models find things buried in the internet that Google would never pull up in a search. I have tried and compared the two. That said, I don't use AI like it is a Google search. If I need to know the hours of the local county library, I use Google. 

Lastly, I will cover the issue of humans losing their jobs to AI. I experience this already happening in some sectors like customer service. In the Genealogy Community, the first jobs it is impacting is document translation, photograph editing, and transcription (there's probably more than that but I am speaking from my experience). But why some employers who are quietly rehiring the employees they laid-off in the Great AI Backfire, is because they still need humans to proof the work that AI does. Remember the hallucinations? What AI is really great at is increasing productivity. But you still need the humans in the room to proof AI's output and then take action on that output.

Adding Custom Instructions to ChatGPT

This morning, I was discussing AI with one of my longtime genetic genealogy friends and he replied that he recently subscribed to ChatGPT. I asked him if he had added "Custom Instructions" yet and he said he had not; so I replied that I would do a blog post about it as this is something I highly recommend everyone does. To customize your instructions, open ChatGPT, then click on your name, then click "Personalization" and under the box that says, "Custom instructions" and enter the following or you can edit it to tailor it to your needs. Here is what mine currently has: 

Scholarly, factual, no embellishments, no fabrications, no dashes, always cite sources. Do not search nor include any sources from Facebook, Instagram, X (formerly Twitter), and Grok. They are to be excluded.

I specifically exclude these sites because ChatGPT does give me these results if I do not. It still sometimes gives Facebook links and then I will tell it not to but it is much rarer with the customization. The hallucinations/fabrications are almost nil now. I still check all of the sources before I use them, and the quality of the sources has improved immensely. Definitely follow the Coalition for the Responsible AI in Genealogy Accuracy principle: "AI can generate false, biased, or incorrect content. Therefore, members of the genealogical community verify the accuracy of the information with other records and acknowledge credible sources of content generated by AI."

To give credit where credit is due, I learned about the customization from Lissa DuVall last December 2025 in the Genealogy and Artificial Intelligence Facebook Group so big debt of gratitude to her!  

Welcome to GPT for the Family Tree!

Now I'm pretty adept at thinking up catchy names but I'm definitely not as quick as ChatGPT 2.5! I prompted "Come up with puns using or related to AI and genealogy for a new blog" and my first choice from the list was "Relative Intelligence" but it was already taken for a URL. This one, "GPT for the Family Tree" was not taken so there we go. I needed a place where I could share the latest as I learn about it. 

Yesterday, Steve Little posted in Blaine Bettinger's "Genealogy and Artificial Intelligence (AI)" Facebook group (join the group if you're a genealogist) about Professor Ethan Mollick's blog post at One Useful Thing on a guide to AI apps. I subscribe to ChatGPT so I found the post extremely helpful to change my setting to "Thinking". I am a patient person (except for awaiting DNA results) so appreciate the "extra" that comes with the "Thinking" mode. I've included a  screenshot where you can toggle the drop-down between the different modes. The default mode is "Auto". Professor Mollick's post is also helpful in knowing what features you receive with the free AI app versions. I am quite irritated with the misspellings that I see in Google Gemini's images and will not subscribe until it is upgraded to a better standard where those don't occur.