- Everyday AI
- Posts
- Ep 813: AI Cost Control 101: Why Your Chatbot Bill Is Becoming a Board-Level Problem (Start Here Series Vol 31)
Ep 813: AI Cost Control 101: Why Your Chatbot Bill Is Becoming a Board-Level Problem (Start Here Series Vol 31)
DeepSeek is building its own AI chip, OpenAI launched GPT-Realtime-2.1 mini, and Amazon is raising $25 billion to fund its AI expansion. And more.
Everyday AI
July 07, 2026
👉 Subscribe Here | 🗣 Hire Us To Speak | 🤝 Partner with Us | 🤖 Grow with GenAI
In Partnership With Slack
Meet the All-New Slackbot: Your Personal AI Agent for Work
Most AI tools only know what you type into them. The all-new Slackbot is different, less like a tool, more like a teammate. Built right into Slack, Slackbot already knows your team's conversations, decisions, and the work you've built up over time, so there's no re-explaining yourself.
Slackbot can now read, write, and act across the other apps your team already uses, and it only ever sees what you're already allowed to see, with your data never used to train the model. Check out Slackbot today.
Sup y’all 👋
Our weekly ‘Working Wednesdays’ is tomorrow, where we go hands-on with a new-ish AI feature.
What should we cover for tomorrow’s AI Working Wednesday show? |
See ya tomorrow!
✌️
Jordan .
Outsmart The Future
Today in Everyday AI
8 minute read
🎙 Daily Podcast Episode: AI costs are climbing fast. In Start Here Series Vol. 31, we break down how businesses can control AI spending without sacrificing performance. Give today’s show a watch/read/listen.
🕵️♂️ Fresh Finds: OpenClaw rolled out major mobile updates, Palantir partnered with NVIDIA for secure government AI, and OpenAI is rebranding GPT Live 1. And more. Read on for Fresh Finds.
🗞 Byte Sized Daily AI News: DeepSeek is building its own AI chip, OpenAI launched GPT-Realtime-2.1 mini, and Amazon is raising $25 billion to fund its AI expansion. And more. Read on for Byte Sized News.
💪 Leverage AI: The days of unlimited AI are ending. Companies now need a real strategy to control AI costs without slowing innovation. Keep reading for that!
↩️ Don’t miss out: Miss our last newsletter? We covered: Broadcom extended its Apple chip deal through 2031, Microsoft is cutting 4,800 jobs, and Meta's new Watermelon AI model is now in training. And more. Check it here!
Ep 813: AI Cost Control 101: Why Your Chatbot Bill Is Becoming a Board-Level Problem (Start Here Series Vol 31)
AI’s all-you-can-eat era is ending. 🍲
For years, one subscription felt like unlimited access to frontier models.
But that business model for the AI labs apparently breaks when agents can now run for days, use tools, retry work and burn through tokens.
And with Anthropic's powerful Fable 5 model exiting subscription tiers today and moving to API only pricing, it's as imperative of a time as ever to figure out your AI spend strategy.
Frontier AI is becoming a metered utility. On today's show, we teach you how to deal with it.
Also on the pod today:
• Fable 5 subscription access ending ⏳
• API metering replaces flat fees 💸
• Tesla’s $200/week AI cap 🚗
Listen on our site:
Click to listen
Subscribe and listen on your favorite podcast platform
Listen on:
Here’s our favorite AI finds from across the web:
New AI Tool Spotlight – AnySearch is The Search Infrastructure Your AI Can Trust, Typeahead is AI Autocomplete that works everywhere you type, AirKaren is the AI that fights airline customer service
OpenClaw Mobile — OpenClaw just dropped big mobile updates, adding chat, talk, photos, and improved QR and security.
NVIDIA and Palantir — Palantir's teaming up with Nvidia lets U.S. agencies customize AI models and keep their data secure.
GPT Live 1 — OpenAI just rebranded its upcoming voice feature from GPT Bidi 1 to GPT Live 1.
Codex Updates — Codex Remote just got a major upgrade in ChatGPT for iOS, with new thread management tools and file search.
Gemini Business Upgrades — Google Gemini now lets you link your Business Profile, pulling in real-world data to give you smarter, personalized insights.
Gemini Spark Upgrades — Gemini Spark lets you get custom game analysis emails after your favorite team plays, but users say Google's slow updates and poor support are dragging it down.
Google and Apptronik — Google DeepMind is putting Apptronik’s Apollo 2 robot to work in real factories, feeding its real-world data straight into Gemini.
Claude Voice Mode Updates — Anthropic’s Claude Voice Mode is getting a new UI and updated voices, plus a persistent model selector.
GPT-5.6 Projection — Polymarket odds shot up to 83% that OpenAI drops GPT-5.6 this Thursday.
Samsung Investors — Samsung posted its highest profit in years, but investors wanted more after the AI chip surge.
Elon Musk on AI — Elon Musk says AI could make work optional by creating abundance, but Michael Burry thinks we’ll see a major economic shakeup first.
Grok Voice — Grok just dropped 25+ new multilingual voices for testing on xAI.
AI Movies — Hollywood's first AI-generated actor, Tilly Norwood, is getting a full-length feature film called Misaligned.
Claude Hidden Workspace — Researchers discovered that Claude has a hidden "workspace" in its brain, letting it think silently and reason with concepts it doesn't say out loud.
AI and Illinois — Illinois just passed a law aimed at making AI companies more accountable for how they use your data.
1. DeepSeek is reportedly building its own AI inference chip to cut reliance on Nvidia and Huawei ⚙️
According to Reuters, DeepSeek is in early talks with chip-design, foundry and memory partners as it works on a custom chip for running AI models, not training them.
The move would be a major shift for China’s best-known AI startup, giving it more control over the hardware behind its models while U.S. export limits continue to squeeze access to Nvidia’s top chips.
2. OpenAI adds GPT-Realtime-2.1 mini for cheaper, faster voice AI ⚡
OpenAI’s GPT-Realtime-2.1 mini is now positioned as a lower-cost, faster option for realtime voice apps, with support for text, audio, and image input, plus text and audio output.
The model is built for WebRTC, WebSocket, and SIP connections, and its improved alphanumeric recognition means it should be better at handling details like codes, names, and numbers in live conversations.
3. Amazon seeks at least $25 billion in bonds to bankroll AI buildout 💸
Amazon is moving to raise at least $25 billion in the US bond market, Bloomberg reported, joining Meta, Nvidia, Oracle and others in a fast-growing debt push to fund massive AI infrastructure.
The rush has helped tilt the investment-grade market toward stronger-rated tech borrowers, but it is also concentrating investor exposure around one big bet: that AI spending will eventually produce enough savings and revenue to justify the borrowing.
4. Apple’s new Home AI features will require 2TB iCloud+ ☁️
Apple’s latest macOS Golden Gate beta notes confirm that the new Apple Home AI tools announced at WWDC 2026 will sit behind the 2TB iCloud+ plan or Apple One Premier.
That means users paying $10 a month for 2TB iCloud+ or $37.95 for Premier will get access, while lower iCloud+ tiers remain limited by camera count and appear to miss the advanced video analysis. The key feature is AI-powered HomeKit Secure Video review, which can analyze recorded clips for people, objects, and events to make security footage easier to search and understand
5. China weighs AI model export curbs 🇨🇳
According to Reuters, China is considering new limits on overseas access to its most advanced AI models, after officials met with Alibaba, ByteDance and Z.ai in the past month.
The move would treat top AI systems as strategic technology, potentially restricting both closed-source and open-weight models while adding tougher penalties for leaks or theft.
For three and a half years, frontier AI subscriptions were the best deal in business.
Pay a monthly fee, run models (now agents) for hours, never see the REAL bill.
That era ends today.
Anthropic’s Claude Fable 5 leaves included paid subscription access in hours. And that’s the latest big domino to fall.
GitHub Copilot moved to GitHub AI Credits, Google tightened Gemini around compute-based limits, Microsoft priced Copilot Cowork through credits, and Grok shifted users toward weekly pools and extra usage credits.
Frontier AI is mostly now a metered utility, and the subscription that covered everything prolly aint coming back. In its place: paying for metered usage, which could easily be 5-20X of a normal subscription.
Yikes.
So, yeah, your AI budget just became a board-level problem.
That’s what we tackled on today’s episode of Everyday AI: why tokenmaxxing crashed into token efficiency, how companies are using more AI without lighting money on fire, and the seven-step playbook to keep agentic work from becoming finance’s next group therapy session.
You’ve got a window. Let’s get after it.
1. Cheaper tokens, bigger bills ⚡
Token prices dropped roughly 98% over three years, and bills still got uglier.
Welcome to Jevons paradox, minus the econ class. When the unit cost falls, people use way more of the thing, and AI teams absolutely did.
Longer context windows. More retries. More tool calls. More agents running for eight to 16 hours while we all pretend subsidized subscriptions will last forever.
The Stanford Digital Economy Lab found agentic coding tasks can use 1,000x more tokens than code chat or reasoning chat, with the same task varying up to 30x between runs.
That makes forecasting brutal. It also makes “use more AI” feel weirdly dangerous unless leaders can see where the tokens go.
Uber reportedly burned through its 2026 AI coding budget in four months. Tesla capped employee AI tool spend at $200 a week, and UBS says 60% of interviewed enterprises are already throttling AI spend.
If the giants are reeling in budgets, smaller and mid-sized teams have a lane: move faster, but measure better.
Try This
Pull your team’s AI usage by tool, model, workflow, and project this week. Add outcome tags next to the spend: shipped code, customer response, research, internal doc, dead-end loop.
If you can’t answer what a project costs in tokens, don’t touch the budget yet. Blind cuts are how useful workflows die and zombie prompts survive.
2. Route smart, not expensive 🔥
Coinbase showed the grown-up playbook.
They moved default work toward open-weight models like GLM-5.2 and Kimi K2.7 Code, paired that with caching and difficulty-based routing, and cut spend to nearly half its peak while total token usage kept rising.
Translation: they didn’t win by using less AI.
They won by sending the boring, repeatable, lower-risk work to cheaper models and saving the heavy hitters for work that deserves the bill.
Model routers make this easier. OpenRouter Fusion points to the new control layer: fan a prompt across models, synthesize the answer, and stop pretending one premium model should do everything.
Do your privacy and data-sharing homework before connecting any third-party router. Boring? Sure. Necessary? Also yes.
Try This
Pick the three AI tasks your team runs every day, then test one cheaper open-weight model against your current default. Compare quality, cost, latency, and failure modes.
If you’re getting 85% to 95% of the value at a fraction of the cost, congrats. You just found budget to spend on the tasks that actually need frontier power.
3. Fine-tuning builds the real moat 🚀
Routing picks the right model for the moment.
Fine-tuning builds something competitors can’t copy with a credit card.
Microsoft Foundry now offers fine-tuning as a service, and Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, showed the sharper version with Bridgewater. A tuned specialist beat a frontier model at about 14x lower cost.
That’s the executive unlock hiding in the boring stuff. If 80% of your team’s AI usage is one repeated task, a smaller tuned model may do it faster, cheaper, and more consistently than the fanciest frontier option.
Try This
Audit your workflows and find the single most repeated AI task. Support triage, financial modeling, compliance review, code review, sales research, and weekly reporting are all fair game.
Define the pass/fail bar, gather the best examples, and ask your cloud provider what fine-tuning path is realistic now. The ROI won’t come from prompting harder forever. It’ll come from turning repeated judgment into cheaper infrastructure.
Reply