AI Content Has Run Out of Words, but Not Subscription Tiers

Lede

AI deserves credit when it improves, but apparently every update also requires a demolished brain and another monthly payment.

Hermit Off Script

AI content is exhausting me, and it isn’t because the technology is uninteresting. It’s because so much of the coverage I encounter seems to have settled on two descriptions: insane and mind-blowing. A new model is insane. A new feature is insane. A small improvement arrives and, somehow, my mind is supposed to be blown again. Either these creators don’t know any other words, or they think those words are enough to make people watch. They have achieved the opposite with me. I’ve started avoiding that content because the language makes a potentially intelligent discussion sound stupid before it has even started. Tell me what changed. Show me the difference. Show me what didn’t work. I don’t need instructions about how amazed I should be. I can manage that part myself. Then there is the money. Some demonstrations leave me wondering whether I’m supposed to have a few hundred dollars spare every month for subscriptions, tokens and additional services. Cheaper access exists, so I won’t pretend everyone needs the most expensive package. But the price of entry and the cost of repeating an elaborate workflow are different questions. A demonstration without its full bill tells me less than the presenter seems to think. [OpenAI; Claude] And then there is Musk. I don’t need to pretend Grok is useless to decide that I don’t want to fund the business behind it. His public support for Germany’s far-right AfD is documented, including his appearance at its January 2025 campaign rally. [Reuters; AP] I regard that political support as extreme and damaging. That is my judgement, and it affects where my money goes. However capable Grok becomes, technical performance doesn’t settle the question of whom I want to support. I think more customers should make that distinction. If someone uses their influence in ways I believe undermine democracy and freedom, losing customers should be a consequence. If enough customers withdraw their money and a business fails, I won’t regard its impressive software as an argument that everyone should have kept paying. I can’t claim the whole world has forgotten. But it often feels as though a shiny release is enough to make yesterday’s objections inconvenient. I’ll reconsider the software when it improves. I won’t automatically reconsider my principles.

P.S. Update – 12 September 2026
I originally drafted this article in July 2026, but it remained unpublished. I’ve now revisited and updated it for its first publication today, 12 September 2026, taking account of newer AI releases and my own experience using the tools. The original rant reflects what frustrated me then; this update considers what has changed since. Genuine improvements deserve recognition, even when the marketing remains irritating. Calling every release “insane” still isn’t a review.
The newer releases require a fairer distinction than simply calling everything a small update. Grok 4.6 arrived on 12 August 2026, Anthropic announced Claude Fable 5.1 on 1 September, and OpenAI introduced GPT-6 Astra on 3 September. Their published descriptions concentrate on longer, multi-step work, coding and completing more demanding tasks. Those are claims worth testing, not dismissing because the promotional language is irritating. [SpaceXAI; Anthropic; OpenAI]
My own experience making the images for this article provides a smaller, more useful example. I liked the artwork. Then I asked for Musk’s face on the prohibition sign in the third panel. One revision turned him into the main figure, changing the meaning of the scene. Further corrections were needed to get the sign right and keep it within the panel. The result could be impressive while the editing process still needed supervision. Both observations belong in an honest review.
That is the test I want applied to new releases, including OpenAI’s. Can the product finish the work? Can it make one correction without changing something else? Does it check the result against the original request? What does the finished, checked result cost once retries and human attention are counted?
OpenAI also acknowledges that Astra’s additional safety checks can slow, pause or stop legitimate work. That is a practical limitation worth reporting alongside capability claims, not an argument for removing safeguards. The product a customer can actually use matters more than the best performance achieved in a research setup. [OpenAI]
I am happy to recognise better technology. I would also like better coverage of it. The models are learning to do more; the headline department could try learning a third adjective.

What does not make sense

• Calling a minor feature and a substantial capability gain equally “insane” tells the viewer almost nothing about either.
• Presenting a successful demonstration without showing retries, corrections or the final bill.
• Treating a benchmark score as a guarantee that an ordinary customer will get the same result with their own tools and settings.
• Praising an assistant for saving time while leaving the user’s checking and repair work out of the calculation.
• Dismissing genuine progress because the advertising is irritating. That is still allowing the advertising to do the thinking.
• Assuming that a technically excellent product entitles its owner to a customer’s money, regardless of that customer’s principles.

Sense check / The numbers

Updated on 12 September 2026. Prices below are published US-dollar figures, not UK checkout quotations. Subscription access and metered software access are different products.

1. OpenAI’s Astra launch report gives GPT-6 Astra a score of 57.9 per cent on Terminal-Bench 4.0, compared with 37.3 per cent for GPT-5.6 Sol. That is evidence of a substantial reported improvement on this coding benchmark. OpenAI also says its evaluation scores use the maximum at any effort level and that research or API results can differ from production ChatGPT. A benchmark deserves its conditions, not merely a victory caption. [OpenAI]
2. Published standard API prices for GPT-6 Astra and Claude Fable 5.1 are $10 per million input tokens and $50 per million output tokens. Grok 4.6 starts at $2 input and $6 output. Tokens are the pieces of text the models process and generate. These are unit prices, not quotations for completing a project; different amounts of processing, caching and retries can change the final comparison. [OpenAI; Anthropic; SpaceXAI]
3. OpenAI lists $100 and $200 monthly Pro tiers, with the same core capabilities but different usage allowances. However, from 10 September 2026, new subscriptions and upgrades to the $200 tier are temporarily paused. Existing $200 subscribers continue as usual, and the $100 tier remains available. Being shown an expensive plan and being able to buy it are now separate details worth checking. [OpenAI]
4. Useful access does not universally require hundreds of dollars: ChatGPT Plus and monthly Claude Pro are each listed at $20, while SuperGrok is $30 and SuperGrok Plus is $100. These are not equivalent packages or usage limits. The criticism should concern the cost of the demonstrated workflow, rather than imply that every useful AI interaction needs a premium budget. [OpenAI; Claude; SpaceXAI]
5. On 25 January 2025, Musk addressed an AfD campaign event by video. Reuters reported an audience of 4,500, and AP independently reported his remote appearance and support. This supports the factual basis for discussing his political conduct. Whether that conduct rules out buying his products remains a consumer’s judgement, not a benchmark result. [Reuters; AP]

The sketch

Scene 1: The adjective machine
A silhouetted creator films a demonstration beside an oversized screen covered in excited AI headlines. A much smaller box contains the actual description of what changed.
Dialogue:
Creator: “Mind-blowing!”
Viewer: “What changed?”

Scene 2: The affordable future
The same viewer stands before an AI machine beneath a banner promising access for everyone. Receipts emerge for the model, additional tools and repeated attempts.
Dialogue:
Viewer: “Total cost?”
Machine: “That’ll be extra.”

Scene 3: Good model, wrong wallet
The viewer holds a closed wallet beside a screen displaying strong model performance. Musk appears only as a portrait inside a prohibition sign.
Dialogue:
Viewer: “Good model.”
Wallet: “Still closed.”

What to watch, not the show

• Whether the reviewer shows the whole task, including mistakes, corrections and the final usable result.
• Whether sponsorship, affiliate links, free access or unusually expensive settings affect the demonstration.
• Whether the quoted cost covers subscriptions, metered usage, extra services and the work needed to check the output.
• Whether a claimed improvement concerns the underlying model, the surrounding tools, a usage allowance or simply a better-looking first attempt.
• Who gets access, under what conditions, and who remains responsible when the result is wrong.
• Whether admiration for a product is quietly being treated as forgiveness for the conduct of the people behind it.

The Hermit take

Give genuine progress its credit, not a standing ovation for every patch.
A better model still doesn’t decide where my money goes.

Keep or toss

Keep / Toss.

Keep useful advances, honest testing and the freedom to choose.
Toss compulsory amazement, incomplete cost comparisons and the idea that good software cancels bad judgement.

Sources

• OpenAI GPT-6 Astra announcement and evaluation conditions: https://openai.com/index/gpt-6-astra/
• OpenAI release notes, including the 3 September 2026 Astra introduction: https://openai.com/products/release-notes/
• Anthropic Claude Fable announcements, access and API pricing: https://www.anthropic.com/claude/fable
• SpaceXAI Grok 4.6 announcement and pricing: https://x.ai/news/grok-4-6
• OpenAI Pro tiers and the September 2026 subscription pause: https://help-lb.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
• OpenAI ChatGPT Plus pricing and separate API billing: https://help-lb.openai.com/en/articles/6950777-what-is-chatgpt-plus
• Claude subscription pricing: https://claude.com/pricing
• Grok subscription pricing: https://x.ai/pricing
• Reuters on Musk’s 25 January 2025 AfD campaign appearance: https://www.reuters.com/world/europe/elon-musk-appears-video-german-far-right-campaign-event-2025-01-25/
• AP on the January 2025 demonstrations and Musk’s AfD rally appearance: https://apnews.com/article/germany-afd-protests-farright-elections-b318328d080b026424137653513e37ac
• Historical source for the existing comic’s July 2026 benchmark figures: https://x.ai/news/grok-4-5

Disclaimer: This article contains satire and personal opinion. Political judgements are the author’s; company-reported performance figures are identified as such. Comic dialogue is fictional.