Matthew Berman opens by claiming that the best AI users in the world have stopped obsessing over prompt engineering and instead focus on which tokens they use for which task. The video is a practical primer on "tokenomics": what a token is, why not all tokens are created equal, how to combine models to raise quality while cutting cost, and the larger market forces deciding how intelligence will be priced. The promise is straightforward: increase the quality of your AI usage while decreasing the cost.
What: A token is the basic unit of a large language model like ChatGPT or Claude. You can think of tokens roughly as words. In the example "The cow jumped over the moon," each word is one token, for six total. But words do not always map to a single token: "programmatic" becomes two tokens, "program" and "matic."
Why: Models work by predicting the next token. Given "the cow jumped over the," the model ideally predicts "moon." Understanding tokens is the foundation for understanding cost, because usage is billed per token.
How: Berman demonstrates with the ChatGPT tokenizer, which shows how text is broken into colored token chunks.
What: Providers price tokens very differently. In the video's examples, GPT-5.6 ("Soul"), OpenAI's most recent and most powerful release, is $5 per million input tokens and $30 per million output tokens. Claude Fable from Anthropic is even more expensive. A new Chinese open-source model, Kimi ("Kimmy K3"), came in at $3 per million input and $15 per million output, about half the price of GPT-5.6.
Why: Half the sticker price sounds like half the cost, but it is not that simple. Using Artificial Analysis data, Berman shows cost to complete a task is nearly equal for Kimi (about 95 cents) and GPT-5.6 (about $1.04), while Claude Fable is far pricier ($2.75). Kimi is half price per token but takes roughly twice as many tokens to solve the same task, so the effective cost is the same.
How: The key metric is intelligence density: how much intelligence you squeeze into each token. Berman's take is that OpenAI/ChatGPT lead on token density, while Fable takes many more tokens to solve the same problem as GPT-5.6.
What: Different models approach problems differently, write code differently, and excel at different things. A strong technique for higher-quality code is letting two models review each other's work.
Why: When Claude (Opus 4.7) authored a pull request and GPT-5.5 reviewed it, the reviewer found more bugs than if Claude had reviewed its own code. Models even cluster differently by error type: for example the video notes Claude was more prone to "missing behavior" issues while Codex was more prone to "semantic intent" issues.
How: The video's sponsor, a code-review tool (Graphite), ships a feature that runs exactly this cross-model review; Berman notes teams like Nvidia, PostHog, Zapier, and Substack use it. A second axis to weigh is output speed: if a cheaper model takes an hour where a faster model takes 30 minutes, you may be losing time even while saving on tokens.
What: Berman's recommended workflow for the hardest tasks uses three roles. A top frontier model does the planning (for example Fable writes a full spec after reading the codebase). A cheap, fast, but capable model does the execution, the actual code writing (he names Grok 4.5 or Cursor's Composer model). A different frontier model (GPT-5.6) does the final review against the spec.
Why: The hardest part is high-level thinking, seeing around corners, so that goes to the best model. The bulk of tokens, however, are spent during execution, and output tokens are much more expensive than input tokens. You do not need a frontier model to write code once a good plan exists, so running execution on a cheap model captures most of the savings.
How: The input/output mix flips by stage: planning reads the whole codebase (heavy input, roughly 1M input, 100K output), execution mostly writes code (light input, heavy output), and review is mostly reading with little output. In Berman's cost example, Fable-only was about $81, GPT-5.6-only was cheaper, and the recommended model mixture came in as the clear winner at about $25.55 while getting the best results. Bonus: using a fast cheap executor also finishes in less overall time, since the biggest frontier models tend to be slower.
What: The future of how intelligence gets priced is being decided right now, framed as a battle between closed-source and open-source models, and a race between the US and China. The biggest, best models come from essentially two US labs, OpenAI and Anthropic, and their tokens carry the highest prices and margins.
Why: Selling tokens at high margins is the closed-source business model. Open source changes the dynamic: because anyone can download, run, and optimize a model like Kimi, every hyperscaler and data center competes to serve it at the lowest price, which drives inference costs down and benefits consumers.
How: When open source wins, token margins collapse and profit shifts elsewhere in the stack, to Nvidia and chip makers, to hyperscalers, and to the application layer, because cheaper tokens mean people use more tokens. If closed-source frontier labs win, you keep getting the best tokens but pay a premium; if open source wins, you pay less per token and the closed labs must compete on things other than token margins.
The through-line is that AI value comes from matching the right token to the right task, not from chasing the lowest sticker price or a single "best" model. Judge models by cost to complete a task and by intelligence density, split hard work into planning, execution, and review across the right models, and watch the open-source vs closed-source contest that will decide how much you pay for intelligence. Do this, Berman argues, and you raise quality while cutting cost.