![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
) - and the development/furtherment of public facing 'normie LLMs'/chatbots/etc do help further these as well edit: it's the transformers really it's using from LLMs that are then customized, not the LLMs themselvesThis.
I agree, but AI is something different—the things you mention are tools that can be used by humans, whereas AI will be a tool that uses humans, and depending on how it develops, this could become “scary” in some aspects.
So...buy more F?
At no point the word fired/offest is in the article, the way it is phrased make it sound they are paying a fortune to bring back for a short time people that retired with full benefit already, which is common in many industry and usually as little to do with AI, just baby boomer retiring leaving gapping instutional knowledge hole, we see it in old school climatisation system and many other sector.
My apologies. My attention span for reading anything from The Verge is very short.
AI is also a tool that is used by humans.
Paul Kurgman isn't all economist...I can't stand the guy
Companies across tech, entertainment, banking, and many other industries are throttling their employees’ use of AI and pleading with workers to use less powerful models to stop AI costs from spiraling out of control, according to leaked Slack chats, screenshots of internal dashboards, emails, and more material obtained by 404 Media from half a dozen companies including Atlassian, Adobe, and Amazon. In at least one case, AI spending has tripled to more than $15 million a month.
The news shows the looming fallout from companies adopting AI as quickly as possible, and AI providers’ moves to charge enterprises based on how much they use AI rather than a flat fee. Emails obtained by 404 Media even show some companies cutting off access to some AI models altogether in an attempt to stop burning through their AI tokens, and big tech companies like Adobe are ending unlimited access to Claude.
“A lot of people had ideas about how to adjust workflows with lower-reasoning models for certain tasks in order to mitigate token consumption,” an Adobe employee told 404 Media. “But I am not sure that they fully absorbed the news, and I'm not sure the full ramifications are going to be clear to everyone until it goes into effect.” 404 Media granted multiple employees at companies using AI anonymity because they weren’t permitted to speak to the press.
Citi, for example, has shut off access to Claude’s and ChatGPT’s latest models entirely, according to an internal Citi email obtained by 404 Media. That includes Claude Opus 4.6 and 4.7, and GPT-5.5.
Do you know anything else about token spend inside companies? We would love to hear from you. Using a non-work device, you can message Joseph securely on Signal at joseph.404 or Emanuel at emanuel.404
“These models consume significantly more AI Credits per interaction and have been the primary driver of elevated enterprise consumption,” the email reads. The email says Citi disabled the models on June 24 and plans to re-enable them on July 1.
Before shutting off access, Citi sent employees another email asking them to not use the more powerful models unless they absolutely had to.
“Action needed: Choose the right model for the task (reduce Opus 4.7),” one section of the email reads, referring to one of Claude’s more recent, and token hungry, models. Since AI tokens are now pooled across Citi, the email says, developers with heavier AI-assisted workflows draw more from the shared pools, while lighter users ideally contribute their unused portion, freeing it up for the developers who may need their tokens. “We need everyone to be intentional about model selection to ensure fair access for all users across the enterprise.”
The email points again to Opus 4.7, saying, “Every interaction with Opus 4.7 (and other models in its class such as GPT 5.5) consumes significantly more credits than standard or mid-tier models.” It then provides a breakdown of what Citi employees should use each model for: GPT-5.3-Codex for quick questions, explanations, or simple code generation; the same model or Claude Sonnet 4.6 for code review and “standard chat;” then higher models like Claude Sonnet 4.6 for “architectural reasoning.”
Citi’s changes come directly in response to GitHub moving from a flat subscription model to a usage-based billing one in June, according to the email. The email says Citi is also monitoring daily Copilot usage to find “excessive or anomalous usage patterns early” and has budget controls in place. Citi told 404 Media it has not disabled models and the company is not taking steps to curb usage by allocating workers a certain number of AI tokens. This is despite the email and other screenshots clearly showing Citi blocking access to certain models.
Atlassian, the company behind the popular software product development tool Jira, recently ended unlimited use of AI tools at the company and introduced a dashboard where employees can track how much their AI use costs the company. 404 Media has seen the dashboard, which shows Atlassian went from spending $5 million on things like AWS, Google Cloud, and OpenAI LLMs in the month of August 2025, to more than $15 million in May 2026. The company is on track to spend more than $120 million on AI tools for the fiscal year, the dashboard shows. Atlassian told 404 Media these numbers don’t accurately reflect its AI usage, but declined to say which of the figures were wrong and how.
“I’ve seen a lot of people complaining that they changed their workflow to maximize AI usage, and now they can run out in 2-3 days, especially when using agents or similar or using the latest Claude model. Lots of angsty messages in Slack like ‘now how do I do my job,’” an Atlassian employee told us. “For what it's worth I think it’s insane they were allowing huge amounts of spending on it before, it was only a matter of time before that had to end.”
Inside GitHub things are a bit different. Employees don’t have a limit on token spend, but workers were recently told the company is looking into decreasing token spend by using open source models, a GitHub employee said. The employee told us that GitHub plans to test user-based billing, meaning budgeting AI tool use to individual people instead of teams, projects, or unlimited usage.
At Adobe, unlimited Claude access is not being renewed and will expire on June 30, an Adobe employee said. Workers there were told instead, in essence, try to get everything you can done before that date.
As 404 Media previously reported, Amazon recently shut down an internal company leaderboard which ranked employees based on how much they used AI tools at work. Multiple Amazon employees told us they suspect Amazon shut down the leaderboard because it was encouraging wasteful and expensive AI usage. After Amazon shut down the leaderboard, 404 Media saw a discussion on Amazon’s internal Slack where an employee shared a screenshot showing they had hit a token limit employees seemingly didn’t know existed previously.
“Crazy, we go from no more leaderboard to actual usage limits in two weeks,” one Amazon employee said in a reply on Slack.
An Amazon spokesperson told 404 Media in an email “We encourage employees to use and experiment with AI, and our guidance around AI usage hasn't changed.”
Other companies have burned through their AI tokens. An employee at an entertainment company told 404 Media, “We hit our limit for ChatGPT token use this month for the first time. One developer used almost half the entire company’s allocated pool with no obvious ROI [return on investment].”
Last week 404 Media reported consulting giant Accenture found that much token usage, or ‘chewing,’ is not from supercharged engineers creating lots of code, but people converting PDFs into presentation slides. Accenture is seeing “soaring token spend” among its clients, according to leaked audio 404 Media obtained.
There is an obvious irony—or cold calculation—in Accenture pointing this out. In the audio, senior Accenture staff explained they told their clients to adopt AI as quickly as possible. Now that AI costs have skyrocketed or become unpredictable, Accenture is positioning itself also as the solution to that problem, with one of the employees saying Accenture has a new opportunity regarding its clients: “to really think about token economics.”
Accenture continues to use AI internally for trivial projects, though. Screenshots obtained by 404 Media show an internal tool that lets employees predict which team will win the World Cup. The tool was made with AI, a source with knowledge of the tool said.
“They are still trying to ram AI down our throats at all levels and areas of work,” the source said. “Everyone seems to be trying to outdo each other in finding new ways to waste water and no one is telling us to slow down.”
Adobe, GitHub, and Accenture did not respond to requests for comment.
The other day I talked to a department lead who was very eager to adopt AI and introduce it to his employees. He did not have a technical background, and his entire department consists of experts in an area completely unrelated to AI.
He asked me what he should know about AI as a people’s leader, and what the most frequent reasons an AI use case fails are.
AI is the industry of lies. (image generated by chatGPT)
Since that conversation, I have been thinking about it. What should a people’s leader practically know? Obviously, not that AI is so transformative and revolutionary — they have been hearing that in dozens of presentations. In fact, AI has been somewhat forced on employees:
Accenture now tracks how often senior employees utilize artificial intelligence on a weekly basis, according to recent reports. The firm links these adoption metrics to promotion opportunities for veteran staff, to ensure they embrace the growing role of technology in the workplace.
In recent years, AI has turned into an industry of lies: benchmaxxing, hallucinated consulting reports, overpromising, underdelivering, AI experts who have never heard of a loss function — barely any other area is as filled with lies as AI.
It is not easy to comb through these braids of lies.
What are the real things one needs to know? That is what this article is about.
In this article, I will refer to LLMs as AI, because that is what is frequently understood by it. So further on, we talk exclusively about state-of-the-art LLM-based approaches.
AI is not cheap
One needs to budget for it, and, without proper controls, an AI application that is actually useful can end up costing more than a full-time employee. The other problem is that the prices are not predictable. Three months ago, one might have budgeted roughly €1,100 a month for GitHub Copilot Enterprise for a team of 30 developers. Now, that same budget might not be enough for a single heavy user: since GitHub switched to usage-based billing in June 2026, costs are calculated per token at the API rate of whichever model handled the request — and as reasoning models become the default, individual monthly bills are already being reported in the hundreds. How much one actually needs to budget is also hard to predict: there is no reliable way to calculate the cost upfront.
$500M on consumption in a month
Many AI tools — particularly coding assistants and enterprise platforms — have moved from flat subscriptions to consumption-based pricing, which means one pays per million tokens. AI APIs are affected too even though they have always been consumption-based. The cost of those tokens has been going up. Anthropic's Claude Fable 5 ships with a new tokenizer that generates roughly 30% more tokens from the same input text, meaning the same system prompt and conversation history that cost 10,000 tokens yesterday costs 13,000 today, with no change in what you're sending. OpenAI's GPT-5.5 is priced at double GPT-5.4 on every billing line.
One can pay per million tokens up to $50.
How much can one do with a million tokens? That is something impossible to predict.
The price includes reasoning tokens — the internal deliberation of the LLM when trying to solve a problem. A response that looks like 500 tokens in the output may have consumed 50,000 or more. The price of every request depends to some extent on the skill of the programmer to word tasks in a way that helps the LLM reason less and produce the solution faster. But, in general, prompting skill is not the decisive factor. What matters more is how much context is sent with every request: the whole chat history, agent descriptions, tools, MCP servers, and system prompts — all of it is billed on every single call. But the most important factor is luck. If the LLM happens to generate the right reasoning traces quickly, you save money. If it does not, it can ramble indefinitely trying to self-correct. The choice of model also matters: some reason longer than others. The banned Fable could burn $50 or more from a single prompt easily.
No one can tell you that your employees need X tokens a month. The best way to figure this out is to calculate what is economically feasible for you: the trade-off between productivity and cost. If your developers run out of quota within three days, two things will happen: they will ship much slower for the rest of the month, or they will use their private accounts. 68% of employees already use personal AI accounts for work tasks, with 57% of them entering sensitive company data and the average cost of a shadow AI data breach is $4.2 million.
All in all, the important steps here are: make the prices and quotas transparent to your team; train developers in best practices for reducing token consumption — prompt optimization, context management, and caching alone can cut costs and consider alternative options, e.g., local deployment of a MiniMax model. A local model will not entirely replace proprietary models, but you can route simpler tasks to them to keep costs down.
Benefit from the heavy competition in AI
AI has become a highly competitive space. Anthropic, OpenAI, and Google are constantly challenging each other. Chinese providers are undermining Western prices massively — releasing capable models available locally, on the cloud, or via API. OpenAI is considering drastic price cuts to avoid losing enterprise market share to Anthropic. You should keep an eye on all of this, diversify your providers, and renegotiate offers.
I have met many companies that simply use the provider their developers picked one or two years ago. By switching, you could prevent shadow AI, have uninterrupted access to models, and save thousands. For example, GLM-5.2 — an open-weight model released three days ago by Z.ai — costs $1.40/1M input tokens and $4.40/1M output tokens. It just claimed the #1 spot on Design Arena, a crowdsourced HTML design benchmark, beating Claude Fable 5 (which, at $50/1M output tokens, is currently unavailable due to a US government export control order — though give it some time). As Andriy Burkov, author of The Hundred-Page Machine Learning Book series, put it on X: he’s been running GLM-5.2 with OpenCode instead of Codex for three days and “doesn’t see any difference” and has already cancelled his Anthropic subscription.
That said, it is still not that easy: force your developers to use a model too weak for their tasks and you will create massive technical debt. Replacing a strong proprietary model with a cheap open-weight alternative risks a far greater drop in product quality than the cost saving justifies. Consult your internal experts, introduce rigorous evaluation pipelines, and rely on validated research and reliable benchmarks. AI Realist offers a service to help companies optimise their AI stack and consumption but whether you choose us or not, go in with a clear understanding of the risks: underperforming models, data privacy constraints, the overhead of maintaining your own infrastructure, throughput, and more.
The big banks, money market and hedge funds, as well as insurance companies are leveraging their bets on the AI bubble. When the collateral behind the AI bubble is NVIDIA and Micron, hold onto your butts. It's a big circle jerk, and everyone else is in the middle.
Man, I wish I had the leverage to short like Michael Burry is doing for Micron.
