Latest / AI Fire Daily / 🎙️ EP 290: $200 GPT Plan Costs OpenAI $14K & GLM-5.2 Crushes Banned Fable 5
Transcript
- 0:00That $200 monthly AI subscription you might be
- 0:02using, well, it could actually be a massive problem.
- 0:05Oh, absolutely. It's a huge issue. Yeah, because
- 0:07it could be costing OpenAI up to $14 ,000 a month.
- 0:11Right, for a single user. It's wild. That is
- 0:15the brutal math behind this current AI boom.
- 0:18Welcome to the Deep Dive. We have a massive stack
- 0:21of research today. And it really paints a fascinating
- 0:24picture. Okay, let's unpack this. Our mission
- 0:27today is to explore the hidden economics of AI
- 0:30subscriptions. We're looking at how massive compute
- 0:33costs are driving this crazy explosion of cheap
- 0:37open source tools. Yeah. And we're also covering
- 0:40how geopolitical bands are radically reshaping
- 0:42the whole global AI landscape. It's all connected.
- 0:44So let's start with that brutal financial reality.
- 0:47It really underpins the entire AI ecosystem right
- 0:50now. It does. It's the foundation of everything
- 0:52we're seeing. We all love a flat rate subscription.
- 0:55You pay 20 bucks or maybe 200 for a pro tier
- 0:58and you just use it. Right. It feels like an
- 1:01all you can eat buffet. Exactly. Like an all
- 1:04-you -can -eat buffet. But if you actually max
- 1:06out your theoretical limits on a $200 ChatGPT
- 1:10Pro plan... It costs OpenAI roughly $14 ,000.
- 1:15$14 ,000 in API compute just to serve one user.
- 1:18Yeah. And API compute is just the server power
- 1:21needed to process AI requests. It's the raw electricity
- 1:25and hardware. And Anthropic's top tier is not
- 1:28much better. Their system caps out around $8
- 1:30,000 in token costs. The break -even points are
- 1:34actually shocking. Yeah, I was looking at this.
- 1:36OpenAI starts losing money at just 11 .4 % utilization.
- 1:40Wow, 11%. Just 11 .4%. And Claude hits its break
- 1:44-even point at a 20 % utilization rate. So if
- 1:47you use it even a quarter of the amount you're
- 1:49allowed to, they lose money. They're just bleeding
- 1:51cash. Yeah. So you have to ask, what drives this
- 1:55massive cost? Because human typing questions
- 1:58can only work so fast. Right. A person can't
- 2:00type $14 ,000 worth of prompts. The real culprit
- 2:03here is agent workflows. Let's define that quickly.
- 2:06Sure. Agent workflows are AI systems that independently
- 2:09loop and execute multi -step tasks. I have a
- 2:13bit of a vulnerable admission here. I still wrestle
- 2:16with my own token usage. when trying to set up
- 2:19agent loops. Yeah. It gets out of hand so fast.
- 2:22Oh, yeah. You are definitely not alone there.
- 2:24I mean, I looked at my dashboard last week and
- 2:26was just staring at the bill. It's because an
- 2:28agent breaks a task down. If you ask it to research
- 2:31a company, it doesn't just write a summary. It
- 2:34breaks that into 20 subtasks. It searches the
- 2:38web. It reads a PDS. It scrapes a database. And
- 2:42it feeds every single step back into itself.
- 2:45So it's constantly talking to itself. Exactly.
- 2:47It loops infinitely until it solves the problem.
- 2:49And every single loop costs money. Going back
- 2:52to that buffet analogy, it's like an all -you
- 2:54-can -eat buffet where one guy just sits down
- 2:56and eats the entire kitchen. Where he be? The
- 2:58restaurant cannot survive that. No, they can't.
- 3:01And we saw this happen. One firm reportedly burned
- 3:03through $500 million in a single month on Claude.
- 3:06Wait, really? Half a billion dollars? Half a
- 3:08billion, yeah. How is that even possible? Don't
- 3:11they have kill switches for enterprise contracts?
- 3:13Well, you would think so, but they didn't cap
- 3:16internal employee access, so everyone was running
- 3:18complex multi -step queries. Wow. An employee
- 3:22asks for a massive data sort, the agent hits
- 3:25a glitch, and it just loops 10 ,000 times in
- 3:28the background. That is just financially terrifying.
- 3:31It is. And it proves you don't need a massive
- 3:33frontier model for every task. You don't need
- 3:35quantum -level AI. to summarize a Tuesday meeting.
- 3:38How can frontier models survive if the business
- 3:41model is this fundamentally broken? They'll likely
- 3:44transition from flat consumer rates to strict
- 3:47metered enterprise usage contracts soon. Unlimited
- 3:49subscriptions die, replaced by strict pay -as
- 3:52-you -go corporate meters. Two sec silence. So
- 3:56because these frontier models are unsustainably
- 3:58expensive, the market is fracturing. Oh, completely.
- 4:01It is urgently fracturing into chipper, highly
- 4:04efficient alternatives. Let's talk about that
- 4:06push for efficiency. Startups are ditching expensive
- 4:08APIs, right? Yeah, they're moving to dirt -cheap
- 4:11alternatives like DeepSeek. And they're saving
- 4:13up to 95 % on compute. 95%. That is the difference
- 4:17between life and death for a startup. Absolutely.
- 4:20And we're seeing specialized models emerge too.
- 4:22Look at the new Kimi K2 .7 code. Right. The coding
- 4:26model. Yeah. It beats Opus 4 .8 on tool use,
- 4:30and it uses 30 % fewer reasoning tokens. And
- 4:34reasoning tokens are just tokens spent thinking
- 4:36before typing an answer. Exactly. Less thinking
- 4:38time means lower cost. Plus, Kimi can be self
- 4:42-hosted for entirely free. So no cloud fees at
- 4:44all. None. And the big players are scrambling
- 4:46to adapt. OpenAI just launched a $150 million
- 4:50partner network. To turn AI plans into actual
- 4:54workflows. Right. And Anthropic is testing something
- 4:57called Conway. I saw this. Whoa. Imagine scaling
- 5:00to an always -on agent environment like Conway.
- 5:03The possibilities are staggering. It really is.
- 5:06Conway is a standalone cloud agent environment.
- 5:08It uses webhooks and active Chrome browsing.
- 5:11Just to clarify, webhooks are automated messages
- 5:13apps send when something happens. Exactly. And
- 5:16this isn't just for big enterprise either. We're
- 5:18seeing a massive explosion of consumer tools.
- 5:21Like Slashy. Yeah. Slashy handles AI email triage.
- 5:26It writes in your voice and makes sure you don't
- 5:28miss follow -ups. And Tastelab. I thought this
- 5:31was fascinating. Oh, Tastelab is incredible.
- 5:33It extracts a website's design DNA. It pulls
- 5:36the hex codes and typography directly into a
- 5:39template, right? And there's Permute, which is
- 5:43a universal media converter. And Athenic 2 .0.
- 5:47Athenic is the data analysis agent. Right. It
- 5:49basically replaces a junior data analyst. It
- 5:52ships automated dashboards and reports. Are we
- 5:55moving toward a future of a million micro models
- 5:57instead of one giant god model? Definitely. A
- 6:00swarm of hyper specialized local models is far
- 6:03cheaper than one expensive monolith. We trade
- 6:05one expensive supercomputer for an army of cheap
- 6:08digital interns. Mid -roll sponsor placeholder.
- 6:11So it's not just cost driving people away from
- 6:14these big... proprietary models, it is also access.
- 6:17Geopolitics is playing a massive role here. It
- 6:19really is. Geopolitics is forcing the world to
- 6:22build around US tech giants. I want to maintain
- 6:24complete neutrality here and just report the
- 6:26facts. Sure. Let's look at the actual events.
- 6:28US export restrictions recently forced Anthropic
- 6:32to shut off Fable 5 and Mythos 5. Yeah, that
- 6:34happened incredibly fast. Anthropic is actually
- 6:37rushing staff to Washington, D .C. right now.
- 6:40They're claiming the security risk was overstated.
- 6:42What's fascinating here is the global ripple
- 6:45effect this caused. Bans didn't slow the competition
- 6:48down. No, they didn't. India is now actively
- 6:50rethinking its dependence on foreign models.
- 6:53Right. They are having a serious debate. Do they
- 6:55lean into open source or do they build sovereign
- 6:58AI? Sovereign AI simply means an AI model built
- 7:02and controlled entirely by one nation. Exactly.
- 7:04And Europe is taking the sovereign AI route.
- 7:07They are heavily backing Mistral AI. Mistral
- 7:10is raising 3 billion euros at a 20 billion euro
- 7:13valuation. Which is huge. Though, to be fair,
- 7:16OpenAI and Anthropix still lead the global market.
- 7:19For now. But the massive breakthrough came just
- 7:21two days after Fable 5 was banned. Yeah, this
- 7:24was wild. China's Zippo AI stepped in. They released
- 7:27GLM 5 .2, and they released it as a fully open
- 7:31source model under a permissive MIT license.
- 7:33That is the key detail. An MIT license means
- 7:36there are zero regional restrictions. Anyone
- 7:39can use it. And GLM 5 .2 absolutely dominated
- 7:43the benchmarks. It hit number one on BridgeBench.
- 7:45Right. It scored 100 .0 on BS and a 42 .8 on
- 7:49reasoning. So it actually beat the banned Fable
- 7:535 model. It did. And it runs at 300 tokens per
- 7:56second. That is incredibly fast. And it does
- 7:58that at one tenth the cost. of the big proprietary
- 8:01models plus it has a 1 million token context
- 8:04window a context window is an ai's short -term
- 8:07memory limit during one conversation right so
- 8:09you can dump an entire massive code base into
- 8:12it at once we saw real world tests immediately
- 8:14the platform z .ai rolled it out instantly yeah
- 8:18developers got their hands on it right away and
- 8:20it coded a 925 line svg clock from scratch it's
- 8:24pure math and visual logic it also built a functional
- 8:273d penalty kick game And a mini spreadsheet.
- 8:30All running flawlessly. If export bans immediately
- 8:33result in highly capable open source clones,
- 8:35do these restrictions actually work? Honestly,
- 8:38they seem to act as a catalyst that just supercharges
- 8:41global open source competition instead. Export
- 8:44blocks fail to contain tech and only force competitors
- 8:47to innovate faster. Two sec silence. So while
- 8:52superpowers fight over these massive models,
- 8:55AI is quietly rewriting our daily realities.
- 8:58It's happening on a deeply personal level now.
- 9:00It really is. It's both taking jobs and managing
- 9:04our personal lives. The job data is pretty grim
- 9:06right now. AI layoffs are rising. Nearly 120
- 9:09,000 tech workers have lost jobs this year. Wow,
- 9:12that's a massive shift. But there is a very surprising
- 9:15statistic here. Almost 75 % of unemployed Americans...
- 9:19Never apply for unemployment benefits. That is
- 9:21so high. People just quietly absorb the loss.
- 9:24Yeah. And this brings us to what we're calling
- 9:26the personal assistant paradox. Right. Because
- 9:28while AI is displacing tech jobs, people are
- 9:31leaning on AI to manage their daily friction.
- 9:33Exactly. For example, there are these 15 chat
- 9:36GPT problems going viral right now. Oh, I saw
- 9:38these. People use them to dissect their paychecks.
- 9:41They use the AI to find highly specific ways
- 9:44to squeeze savings out of their budget. It's
- 9:47crazy. The same tech disrupting the job market
- 9:50is what people use to survive the disruption.
- 9:52Right. So what does this all mean? We are seeing
- 9:55the tension of macro -level job losses mixed
- 9:57with micro -level convenience. Yeah. And there's
- 10:00a lighter side to this convenience, too. It's
- 10:02not all just budgeting and layoffs. True. ChatGPT
- 10:06now has a dedicated World Cup 2026 page. You
- 10:10can track live scores and matches. And you do
- 10:12it without ever leaving the chat interface. It's
- 10:15all just seamlessly integrated into your conversation.
- 10:18It becomes the primary lens for how you interact
- 10:20with the internet. It's the ultimate trade -off
- 10:22of the AI era, exchanging job security for extreme
- 10:25personal convenience. It really seems like we
- 10:27are trading long -term career stability for hyper
- 10:30-efficient... Automated daily task management.
- 10:33Trading lifelong career stability for incredibly
- 10:35smooth daily task automation. Exactly. Let's
- 10:38briefly synthesize the major through line of
- 10:40this deep dive. The era of relying on a few massive,
- 10:44expensive, centralized AI models is fracturing.
- 10:47It is absolutely breaking apart. Between the
- 10:50massive token burn rates we discussed, the $14
- 10:52,000 underlying cost per user, and geopolitical
- 10:57export bans. Right. The global developer ecosystem
- 10:59is just actively routing around all of those
- 11:01roadblocks. We are entering an era of incredibly
- 11:04cheap, hyper -capable, open -source agents that
- 11:08basically live everywhere. It's a completely
- 11:10decentralized future. I want to leave you with
- 11:12a final lingering question to ponder. If government
- 11:17roadblocks only motivate the global community
- 11:19to build smarter, faster, and completely open
- 11:21source models with zero regional restrictions,
- 11:24are we rapidly approaching a point where AI regulation
- 11:28is no longer a legal question, but a technical
- 11:30impossibility? Two sex islands. Thank you for
- 11:33joining the Deep Dive today. Keep questioning
- 11:35the information around you. Out to your own music.