# The AI Circus: More Models, More Money, Same Old Mistakes Canonical URL: https://themodernhermit.blog/ai-circus-more-models-money-same-mistakes/ Published: 2026-11-07T02:44:44+00:00 Updated: 2026-11-07T02:50:54+00:00 ## Lede Artificial intelligence companies keep multiplying their models to save humanity time, while paying customers spend theirs choosing between them, correcting their mistakes and wondering who will have a job left to pay the subscription. ## Hermit Off Script Terra, Luna, Sol, Astra... Bloody hell! Four different models, each with its own levels of reasoning effort, as if choosing the right artificial brain wasn't complicated enough already. How many models, versions and thinking levels do we need before one of them actually does everything properly? Every time I think I've finally found a model that works for what I need, another version appears, with another name, another set of capabilities and another explanation of why it's better than the previous one. And apparently I should be grateful because this one is faster, that one is smarter, another is cheaper and the expensive one supposedly does everything. Well, if one of them does everything, why do I need the others? When will these damn AI companies understand that most of us don't want an entire catalogue of artificial brains? We want ONE very good model, and that's it. Give me something like Netflix, with three or four clear subscription tiers depending on what people need and can afford. No more endless model versions, no more wondering which one to use, no more watching allowances disappear because the machine needed another attempt to fix its own mistake. I understand that computing costs money and that different tasks require different amounts of processing. Fine. Let the company handle that behind the scenes. I shouldn't need to understand the internal economics of a data centre just to finish a task. And why can't that one model simply keep improving? Weekly, monthly, daily, whatever is necessary. We update software constantly without having to choose between twenty different versions of the same application. Yes, I know training an AI isn't the same as installing a software update, but surely the people promising superintelligence can work out how to make the customer experience simpler. If a model needs more training, train it. If it needs better memory, improve it. If it keeps making the same mistakes, fix the bloody mistakes instead of giving the next version a prettier name. Of course, GPT-6.1 Terra arrived from OpenAI on 26 October 2026 and managed to mess up my head all over again. Now I have Terra, Luna, Sol and Astra to think about, and that's before opening Claude and discovering another family of models waiting for attention. Both companies seem determined to compete over who can release another version and explain how much closer we're getting to AGI or superintelligence. But closer to what, exactly? Because from where I'm sitting, none of these systems is dependable enough to make my life consistently easier without supervision. Don't misunderstand me. They can be incredibly useful. I've seen them complete work that would have taken me hours, and that's why I keep using them. But I've also spent those saved hours correcting mistakes, repeating instructions, checking whether something was actually finished and repairing what the previous repair managed to break. Sometimes I feel less like a customer and more like the manager of an artificial employee who needs constant supervision but somehow charges me for overtime. And here's the part that really annoys me: I correct the model, explain what went wrong, show it the proper way and hope it learns. Depending on my data settings, those conversations and corrections may even help the company improve future models. So I'm paying for the service and potentially contributing to its improvement at the same time. But does that mean the model permanently learns my correction? Of course not. It can remember something within the conversation, perhaps save a preference, perhaps follow the instruction correctly for a while, and then another session or another task brings back the same bloody mistake. I correct it again. It apologises again. My allowance gets consumed again. What a wonderful arrangement. The company sells me artificial intelligence, I supply the missing intelligence, and somehow I'm the one receiving the invoice. Then comes the grand promise that all this technology will create extraordinary productivity, abundance and a better future. Perhaps it will. I genuinely hope so. But I also wonder who will receive that abundance. The companies could become trillion-dollar giants while the people paying for their products face fewer jobs, lower wages and more insecurity. This isn't a prediction that every job will disappear. It's the possibility that worries me when I hear so much celebration of automation and so little explanation of how ordinary people will benefit from it. Maybe we'll spend years paying to correct machines that eventually become good enough for employers to decide they don't need us. And then what? No job, no money and no subscription to the artificial intelligence we helped improve. Perhaps all we'll have left is our vote, so politicians can explain why the future they promised has become so expensive. The companies get the profits, the machines get the jobs, and we get to choose who apologises for it at the next election. ***P.S.** From my experience, Astra is the only model that's genuinely useful and dependable enough for serious work, but it's bloody expensive. With the others, I spend half my time testing, correcting, repairing and watching the usage meter instead of actually getting things done. And that's precisely why I think these companies will eventually lose to open-source AI. Why keep paying to supervise their experiments when open-source alternatives could offer more freedom, control and predictable costs? They might win the race to build the smartest AI, only to discover that their customers have already left.* ## What does not make sense - **Four models to simplify one person's work.** OpenAI promotes Astra, Sol, Terra and Luna for different needs. The distinction may make technical sense, but why should the ordinary customer have to manage it? - **Another version instead of a dependable service.** New releases can improve speed, cost and accuracy. They don't automatically remove the need to check whether a real task was completed correctly. - **A subscription that still requires accounting.** Monthly access doesn't necessarily mean unrestricted use. Different models and features can have different allowances, while developer APIs use separate token-based billing. - **The customer becomes the model selector.** An intelligent service should be capable of choosing the appropriate processing method without requiring the user to study model specifications. - **The update problem.** Software improvements usually arrive through updates. AI models require more complicated training, testing and deployment, but customers shouldn't have to absorb every technical distinction. - **The correction that doesn't become permanent learning.** A model acknowledging an error is not proof that its underlying parameters have changed. Remembering a conversation is different from learning a general lesson. - **The expensive apology.** A failed attempt may consume usage allowance, and a subsequent correction may consume more. The customer can end up paying in time and capacity for the same task twice. - **The invisible human workforce.** AI productivity figures don't always capture the user's time spent preparing instructions, supervising execution, correcting errors and verifying the finished work. - **AGI enthusiasm versus everyday reliability.** A model may achieve impressive research results while still failing on a practical task. Both facts can be true. - **The abundance contradiction.** Higher productivity doesn't automatically produce higher wages, shorter working hours or broader prosperity. Ownership and economic policy determine who receives the benefits. - **The customer without an income.** Businesses need paying customers. An economy that automates labour without addressing purchasing power risks undermining the people expected to buy its products. ## Sense check / The numbers 1. **OpenAI currently promotes four main models.** Its selection guidance identifies GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna. Each has a different intended balance of capability, speed and cost. This supports the complaint about multiple choices, although it doesn't prove that all customers must manually select between them. \[OpenAI Model Selection\] 1. **Anthropic also lists four current Claude models.** Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 serve different workloads. Together, the two companies therefore advertise eight principal model options in their published lineups. That figure excludes older versions, specialist models and additional settings. \[Anthropic Models\] 1. **OpenAI's standard API input prices range from $0.10 to $10 per million tokens.** Luna costs $0.10 per million input tokens, Terra $1, Sol $2 and Astra $10. The most expensive is therefore 100 times the cheapest per input token. These are developer API prices, not consumer subscription charges, and don't establish the total cost of successfully completing a task. \[OpenAI API Pricing\] 1. **OpenAI reported a fall in one internal factual-error measure from 53.6 per cent to 16.7 per cent.** Its GPT-6.1 Terra announcement compared the newer model with GPT-5.6 Terra at low reasoning effort on difficult prompts. This is evidence of reported improvement, not proof that errors have disappeared or that every real-world workflow will achieve the same results. \[OpenAI Terra Announcement\] 1. **An estimated 25 per cent of global employment is potentially exposed to generative AI.** The International Labour Organization's 2025 study, developed with Poland's NASK research institute, assessed nearly 30,000 occupational tasks. The researchers concluded that job transformation is more likely than wholesale replacement. Exposure is not an unemployment forecast, but the distribution of future productivity gains remains a legitimate concern. \[ILO-NASK 2025\] *OpenAI's consumer data policy also confirms that eligible content from Free, Plus and Pro services may contribute to model improvement, subject to user controls. Business and API data are treated differently by default. Individual corrections don't automatically retrain the model, and submitting feedback may have separate data-use implications. \[OpenAI Data Policy\]* --- ## The sketch Scene 1: **The Model Supermarket** *A small customer stands before four towering machines labelled TERRA, LUNA, SOL and ASTRA. Each machine has a different usage meter. A large corporate silhouette stands behind the machines.* Dialogue: Customer: "I need one working AI." Salesman: "Choose your intelligence first." Customer: "Isn't that your job?" Scene 2: **The Unpaid Trainer** *The same customer sits at a desk correcting a document. A large mechanical brain feeds the same incorrect document back into the printer. A usage counter continues decreasing.* Dialogue: Customer: "That's the same mistake!" AI: "Thank you for correcting me." Customer: "You're welcome. Again." Scene 3: **The Future of Employment** *The same customer stands outside a closed workplace holding an empty wallet and a ballot paper. Above him, a giant corporate building displays rising profits.* Dialogue: Company: "AI now does your job." Worker: "Who gets the money?" Company: "Remember to vote." --- ![Three comic panels show a customer choosing between four AI machines, correcting a document produced by a mechanical brain and standing outside a closed workplace with a ballot paper.](https://themodernhermit.blog/wp-content/uploads/2026/11/themodernhermit-ai-circus-more-models-money-same-mistakes-sketch-1300x545-1-1024x429.jpg) First choose the model, then correct its work, then ask who receives the profits. --- ## What to watch, not the show - **[Cost per completed task](https://themodernhermit.blog/ai-limits-reset-not-mistakes/).** The meaningful measure is the cost of a verified result, including retries, human corrections and time spent checking. Cheap tokens aren't necessarily cheap work. - **Automatic model selection.** Providers could route tasks to suitable models internally while offering customers one consistent service. The question is whether that becomes reliable enough to replace manual selection. - **Subscription transparency.** Clear usage allowances, predictable prices and understandable restrictions matter more to ordinary customers than another model name. - **Persistent improvement.** Watch whether providers develop dependable ways to retain corrections and apply them across future tasks without compromising privacy, security or existing capabilities. - **Customer training data.** Users should understand when their conversations and feedback may contribute to model development, and what control they retain over that information. - **Independent reliability testing.** [Model announcements](https://themodernhermit.blog/smarter-ai-same-repairs-another-model-upgrade/) should be compared with external evaluations of complete workflows, including failure recovery, instruction-following and accurate reporting of completed work. - **The AGI marketing race.** More capable models are genuine technological developments. Claims about imminent general or superintelligence require separate evidence. - **Who owns the productivity gains.** Automation can reduce labour costs, but whether the resulting value reaches workers, customers or shareholders depends on business decisions and public policy. - **Employment protections.** Training, labour rights, taxation, social insurance and competition policy will influence how AI changes work and who bears the costs. - **Democratic accountability.** Concentrated technological wealth can increase political influence. Elections matter, but voting alone doesn't guarantee fair distribution of economic gains. ## The Hermit take A smarter model is useful. A dependable service that finishes the job without constant supervision is worth paying for. The real test of artificial intelligence is how much human work it saves, not how many new versions it sells. ## Keep or toss ***Toss*.** Keep genuine improvements, fair subscriptions, useful competition and dependable automation. Toss unnecessary model complexity, opaque limits and the expectation that customers should pay to supervise recurring mistakes. #### Sources - OpenAI Model Selection: [https://developers.openai.com/api/docs/guides/model-selection](https://developers.openai.com/api/docs/guides/model-selection) - OpenAI Models and API Pricing: [https://developers.openai.com/api/docs/models](https://developers.openai.com/api/docs/models) - OpenAI GPT-6.1 Terra Announcement: [https://openai.com/index/introducing-gpt-6-1-terra/](https://openai.com/index/introducing-gpt-6-1-terra/) - Anthropic Claude Models Overview: [https://platform.claude.com/docs/en/models/overview](https://platform.claude.com/docs/en/models/overview) - OpenAI Data Use and Model Improvement: [https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance) - OpenAI Data Controls: [https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt](https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt) - International Labour Organization, ILO-NASK Global Index, 20 May 2025: [https://www.ilo.org/resource/news/one-four-jobs-risk-being-transformed-genai-new-ilo%E2%80%93nask-global-index-shows](https://www.ilo.org/resource/news/one-four-jobs-risk-being-transformed-genai-new-ilo%E2%80%93nask-global-index-shows) ***Disclaimer**: This article combines verified facts, satire and the author's personal opinions. The three comic scenes are fictional. The possibility of job displacement is not a prediction that all exposed occupations will disappear. Model availability, API pricing and published policies were checked on 28 October 2026 and may change.*