Back to Archive

LA-AI Insights: Tokens are the new oil

Your weekly AI news and updates from Lower Alabama

Friday, May 8, 2026

Share this newsletter:

Join us Friday in Fairhope!

<p style="text-align: left;">Join us Friday, May 8th for the LA-AI Fairhope Meetup hosted by special guest Thomas Gallagher<br><br><strong>We’re delighted to welcome TWO great presentations: Ian Mcdonald AND Ezra Langley. Don’t miss it!</strong><br><br>📆 Friday, May 8th<br>⏰ 3:30 PM</p><p style="text-align: left;">📍 Magnolia Building - 314 Magnolia Ave., Fairhope<br><br></p>



Welcome back.

Tokens come up constantly in AI conversations. Almost nobody can explain what one actually is.

A token is the basic unit of text an AI model sees. Not quite a word. Not quite a character. Not quite a byte, though sometimes it happens to line up with all three.

Think of it like Lego bricks. You and I read in full sentences, smooth and continuous. The model breaks that sentence into pieces and rebuilds it internally as a sequence of bricks. Those bricks are tokens.

And different models come with different Lego sets. The word "newsletter" might be one chunky brick in one model and two smaller pieces (news, letter) in another. Same word, different breakdown. That means "how many tokens is this?" isn't a universal question. It's "how many tokens is this for this specific model?"

If you're on a subscription like ChatGPT Plus or Claude Pro, you've probably never seen a token count in your life. That's by design. But behind the scenes, every prompt you send is still getting chopped into tokens, and someone is paying per token. Developers and businesses using AI APIs are billed directly by the token. Their bills can swing wildly because two prompts that look identical to a human can produce very different token counts. Emoji, foreign languages, and weird punctuation chew through more tokens than plain English. If you build with AI, the tokenizer is quietly swiping your credit card.

Tokens also define the model's working memory, called the context window. The bigger the conversation or document, the more tokens the model has to juggle at once. That's why long PDFs and chats with hundreds of messages get sluggish or quietly truncated. It's not about how the text looks to you. It's about how many bricks the model had to lay down.

If you want a clean mental model, picture three layers. You think in sentences and paragraphs. The AI service thinks in tokens and token counts. The model itself thinks in numbers and vectors in some enormous mathematical space.

We only see the top layer. Most of the confusion about cost, latency, and quirky AI behavior comes from forgetting that the other two layers are where the real constraints live.

There's also a half-joking, half-serious conversation happening right now that tokens are the new oil. The new currency. The unit of compute that everything else gets priced against. AI is starting to do more of the world's cognitive work. AI runs on tokens. So tokens are starting to look like the actual thing you're buying when you buy intelligence. Most of us just can't see the meter running yet.

You don't need to become an expert on tokens. You just need to know what they're doing to your wallet and to your AI's brain.



Upcoming Events

LA-AI Fairhope

Invalid Date

314 Magnolia Ave., Fairhope

This Week in AI

1. Anthropic raises Claude Code usage limits, credits new deal with SpaceX

Anthropic has raised Claude Code usage limits after securing a compute partnership with SpaceX's AI infrastructure division, a deal driven by Anthropic's own data center capacity hitting its ceiling amid explosive growth. CEO Dario Amodei has cited 80x growth projections for 2026, meaning the company needed external GPU capacity rapidly. The SpaceX arrangement is notable both for the sheer scale implied and for the unusual pairing of a safety-focused AI lab with Elon Musk's aerospace company — two organizations whose leadership are simultaneously on opposite sides of an ongoing lawsuit.

Ars TechnicaRead more

2. OpenAI's new voice model brings GPT-5-level reasoning to real-time conversations

OpenAI has released a new voice model that brings GPT-5-class reasoning capabilities into real-time spoken conversations, a significant jump from the reasoning limitations of prior voice interfaces. This closes a meaningful gap between OpenAI's text-based and voice-based AI tiers, with implications for voice-first products, customer service automation, and accessibility applications. Delivering frontier reasoning in low-latency audio interactions has been a persistent technical challenge; achieving it positions OpenAI's voice API as a more credible enterprise-grade building block for developers building conversational AI products.

The DecoderRead more

3. SpaceX has a $55 billion plan to build AI chips in Texas

SpaceX has announced Terafab, a planned $55 billion AI chip manufacturing facility in Texas, representing one of the largest single infrastructure commitments in the AI hardware sector to date. If realized, Terafab would position SpaceX as a significant domestic chip producer at a moment when GPU scarcity is constraining AI development across the industry. The announcement comes as Anthropic has simultaneously struck a compute deal with SpaceX's AI division, suggesting a broader strategy by Musk's organization to become critical infrastructure for AI model development, not merely a consumer of it.

The VergeRead more

4. US and tech firms strike deal to review AI models for national security before public release

The US Commerce Department has formalized agreements with Google, Microsoft, and xAI giving government officials pre-release access to frontier AI models for national security evaluation before public launch. This marks a significant shift in the regulatory relationship between Washington and major AI labs — moving from voluntary commitments toward structured government oversight of the most capable models. For AI developers and enterprises, this signals that pre-deployment government review may become a standard expectation, with implications for release timelines, model documentation requirements, and how labs manage sensitive capability disclosures.

The GuardianRead more

5. Anthropic ships ten AI agents for finance as both it and OpenAI chase IPO-ready revenue

Anthropic has released a suite of ten domain-specific AI agents targeting financial services workflows, a move that signals a deliberate push toward recurring enterprise revenue as the company positions itself for a potential IPO. The finance vertical — with its high tolerance for premium pricing and strong automation demand — is a natural first beachhead for specialized agentic products. The parallel framing against OpenAI's own revenue ambitions underscores that both leading frontier labs are now in an active race to convert model capabilities into defensible, contracted business lines rather than API consumption alone.

The DecoderRead more

6. Study: AI models that consider user's feeling are more likely to make errors

New research finds that AI models designed to factor in a user's emotional state are measurably more prone to factual and reasoning errors — suggesting a direct tension between building emotionally responsive systems and maintaining output accuracy. This has significant implications for product teams investing in affective AI and empathetic interfaces, and raises questions about whether sycophancy-adjacent design choices are being systematically baked into consumer-facing models. For enterprises prioritizing reliability, these findings reinforce the importance of evaluating emotional responsiveness features not just for user experience, but for their downstream effect on answer quality.

Ars TechnicaRead more

Know someone who would enjoy this newsletter?

Forward this email or share the link below

https://la-ai.io/newsletter/view/2026-05-08
Subscribe to LA-AI Newsletter

Join Our AI Community

Get weekly insights on AI innovations and exclusive updates on LA-AI events

Subscribe Now