Friends, today we bring you a clash of the titans: "money power" versus "value for money."

Here's what happened: a user named Vyom recently got into building games with AI. He pulled out Anthropic's flagship Opus 5, fed it a prompt of about 2,000 words, and asked it to generate a playable American-cartoon boat racing game. The model delivered — INK TIDE: a speedboat, a stretch of water, a racing mechanic, playable.

The price tag? 690 million tokens.

Friends, what does 690 million tokens mean? At roughly 2,000 tokens per Chinese article, that's enough to write 345,000 articles. This wasn't making a game — this was donating to Anthropic.

And then? Within two days of the game's release, another user recreated it in Codex with GPT-5.6 Sol — for $5.

Five dollars. The price of one fast-food combo. Possibly with change.

This Comparison Is Brutal for Anthropic

Let's do the math. 690 million tokens at Opus 5 pricing is a number that gives CFOs heart palpitations. GPT-5.6 cloning the same game costs $5. The gap isn't describable as "price-performance" anymore — it's the difference between farming with a golden hoe and farming with a tractor.

To be fair, the two scenarios aren't fully equivalent: Vyom used a single 2,000-word prompt to have Opus 5 generate a complete game in one go — the "brute force" school. The cloner stood on top of a finished product, so the task was naturally much smaller. But that's exactly the point — this comparison proves the "cost black box" of AI coding can no longer be hidden.

Before, the model discussion was about benchmarks, leaderboards, "my model is smarter than yours." Now? One user builds something with 690 million tokens, another clones it for $5. In real tasks, cost differences between models can span several orders of magnitude. Sure, Claude Opus 5 can write code — but that token counter spins like a taxi meter.

What stings more: this isn't Anthropic's first loss on value. Is Claude Code expensive? Yes. Is Codex cheap? Yes. As companies start scrimping on AI budgets, more and more will pick "good enough and cheap" over "great and pricey."

Three Friendly Reminders for the AI World

Friends, laugh all you want, but this incident carries three takeaways.

First, do the token math before you start. For any big task, ask whether the job really needs a flagship model at full throttle. Building a game demo or an internal tool doesn't require the priciest model as your mule — you don't take the Lamborghini to buy groceries.

Second, learn to tier your tasks. Heavy, hard, long-context reasoning — use the expensive model. Repetitive, templated grunt work — use cheap or open-source models. Tiering models can cut costs by an order of magnitude. The cloning user is the perfect demonstration.

Third, stop believing "expensive means better." The 690M-token INK TIDE and the $5 INK TIDE probably play identically. Benchmark leaderboards are reference material; real-task price-performance is the real deal. Being rich lets you be reckless — just don't let your boss see the bill.

Fairness note: Anthropic's models are genuinely capable — turning 690M tokens into a complete playable game was unthinkable two years ago. But the cruelty of the AI world is this: the moment you prove you're the strongest, someone else proves at a tenth of the price that they're good enough.

That's all the roasting for today. More tomorrow.