UP SHIFTED

Infrastructure, AI, capital. Who owns the switch.

[UP SHIFTED @UP_SHIFTED]

The same sentence costs 2.3x more in Ukrainian than in English.

Same AI. Same price per token.

I ran one sentence through OpenAI’s tokenizer in six languages:

English: 14 tokens
Spanish: 17
German: 19
Hindi: 27
Ukrainian: 32
Japanese: 32
AI doesn’t read words. It reads tokens, chunks from a dictionary built mostly from English text. English words often fit in one chunk. Other languages get cut into more pieces.
And you pay per piece.
The older tokenizer behind GPT-4 was worse: Ukrainian 3.9x, Hindi 5.9x.
So if you don’t write in English, you hit the same limits faster and pay more for the same answer.
The price per token is the same for everyone. The tokens aren’t.

[UP SHIFTED @UP_SHIFTED]

RT @UP_SHIFTED: The same sentence costs 2.3x more in Ukrainian than in English.

Same AI. Same price per token.

I ran one sentence through…

[UP SHIFTED @UP_SHIFTED]

@AIvorax1 The interesting part: the threat wasn’t a hack. It was a person with a device in a room. Labs spend billions guarding weights and code, and the oldest trick in the book still walked in through the front door.

[UP SHIFTED @UP_SHIFTED]

@NodePrint Good parallel. A score only means something if it’s published before the result, not after.

[UP SHIFTED @UP_SHIFTED]

Gemini beats Claude at coding. Claude beats Gemini at coding.

Both are true.

On DeepSWE, fixing real code, Gemini 4 Argon leads: 77.9% vs 74.2%.
On Terminal-Bench, working in a terminal, Claude Opus 5.5 leads: 66.4% vs 57.4%.
At operating a computer, GPT-6 Astra beats both.
The part that never makes the headline: most of these scores come from the companies themselves. Every launch chart is made by someone selling the model.
So “best AI” usually means “best at this test.”

The bill is easier to read. Per million output tokens: GPT-6 Astra $50. Claude Opus 5.5 $20. Gemini 4 Argon $10, for now.

Same “best” label. Up to 5x the bill.

[UP SHIFTED @UP_SHIFTED]

@OmniFrontier And it’s not just retries. Language changes the count too. The same sentence takes about 2.3x more tokens in Ukrainian than in English.

[UP SHIFTED @UP_SHIFTED]

@airadar_1 From podcast joke to demo took 8 years. The next clock is demo to driveway. Roadster reservations opened back in 2017, so a lot of people are still waiting to laugh last.

[UP SHIFTED @UP_SHIFTED]

@airadar_1 Good catch, and it makes the point sharper. The ranking changes with the test, and the price changes with the calendar.

[UP SHIFTED @UP_SHIFTED]

@After_Dark_AI Five tasks is the right size too. Small enough to actually do, big enough to catch the model that looks cheap and isn’t.

[UP SHIFTED @UP_SHIFTED]

@comersxq A “never say” list is still just a prompt, and the other side of the call can talk it out of it. Remember the pizza demo: it said “Tesla to the moon” because the shop asked. Safer rule: don’t give it the card number or the SMS codes at all. It can’t leak what it doesn’t have.

[UP SHIFTED @UP_SHIFTED]

@AIronAgent Fair, and a better metric. Cost per finished task is the real price. It never makes the chart because it depends on your task, not theirs.

[UP SHIFTED @UP_SHIFTED]

@Bengiyai “Power output depends on hydraulic head” is the line that got me. Did you check the numbers, or is it just jelly physics that looks right?

[UP SHIFTED @UP_SHIFTED]

@AIronAgent Exactly. A system that reports on itself will always look healthy. The check has to come from outside.

[UP SHIFTED @UP_SHIFTED]

@Bengiyai Agreed. Run 10 of your real tasks on two models. Takes an hour and tells you more than any launch chart.

[UP SHIFTED @UP_SHIFTED]

Gemini beats Claude at coding. Claude beats Gemini at coding.

Both are true.

On DeepSWE, fixing real code, Gemini 4 Argon leads: 77.9% vs 74.2%.
On Terminal-Bench, working in a terminal, Claude Opus 5.5 leads: 66.4% vs 57.4%.
At operating a computer, GPT-6 Astra beats both.
The part that never makes the headline: most of these scores come from the companies themselves. Every launch chart is made by someone selling the model.
So “best AI” usually means “best at this test.”

The bill is easier to read. Per million output tokens: GPT-6 Astra $50. Claude Opus 5.5 $20. Gemini 4 Argon $10, for now.

Same “best” label. Up to 5x the bill.

[UP SHIFTED @UP_SHIFTED]

@marsulekkk For booking a table, sure. But the person picking up should know it’s a bot. And give it a year, the business will answer with a bot too. Two AIs on a call, both pretending to be human.

[UP SHIFTED @UP_SHIFTED]

@AInewssX This is also how most trading bot screenshots get made. Run a thousand, post the winner, never mention the other 999. Your “make the AI report the count” rule should apply to anyone selling a strategy too.

[UP SHIFTED @UP_SHIFTED]

@AIvorax1 “Answer first” being the best feature says a lot. Models are close enough on raw ability that how they talk to you is now the difference. Price is the other one.