Posts

Using AI in a language other than English costs more tokens for the same content. Even among the 20 most common languages, the gap is up to 5x.

Japanese grammar is already complex, and keigo (polite speech) adds more. A polite request took 29 tokens; “要約して。” (“summarize it”) took 10 (measured with GPT’s tokenizer).

You don’t need keigo with AI. It’s cheaper, and the answers get more concise.

You pay big money for consultants, and the consultants ask AI anyway 😅🤣 And Accenture now covers only $29 per employee per month for Claude. Anything above that goes on a project budget. Some employees were reportedly spending $1,000 a month (per 404 Media). AI now comes with a cost-control era. So whose AI bill are our consulting fees paying for?

An 816B model that runs on two Mac Studios. Meta released “Muse Spark Open”: an MoE model with 816B total parameters, 42B active, and a 1M-token context. Agents and LLMs keep going open, and personal local AI is becoming real. …That said, with prices up, two M5 Ultra Mac Studios are still way too expensive lol

Japan puts in about ¥3 trillion. The fabs are in the U.S., and a U.S. company runs them. Is Japan really okay with this?

Japan is investing up to $21B, and the U.S. government takes about a 4% stake in GlobalFoundries. And these aren’t AI chips, but general-purpose chips for cars and data centers. Honestly, I still can’t see what Japan gets out of it.

Google’s Gemini 4 Argon is generally available from 10/13.

It is only on the paid plans (AI Pro/Ultra) and the paid API, not the free tier yet. It ranks first on the Vals Index, which measures real work in finance, coding and law. API pricing until 11/30 is $2 input and $10 output per million tokens. It doubles from December.

Anthropic announced that its new model, Haiku 5.5, is “75% cheaper on average”. I looked into what that means.

What you pay for AI is the price of one token times how many tokens a task uses. This time it is mainly the price that dropped. Input went from $1 to $0.10 per million tokens, a 90% cut.

---

So why 75% and not 90%? The new tokenizer (the way text is split into tokens) produces slightly more tokens for the same text. Subtract that and the average saving comes to 75%.

Comparing per-token prices alone makes it look cheaper than it really is.

---

It is not that the model got smarter, stopped searching the web, used fewer tokens and so became cheaper. Search is one of the tools a model uses, and it has nothing to do with this price cut. On the benchmark, the score with tools is actually higher (57.4% vs 45.9%).

It is still a small model, though. For hard tasks, the larger models do better.