Every edition of Cinco
One essay a week, every Tuesday at 7am, in five movements: with charts, executed code, judgments and sources. The archive is public and permanent.
Verified: what changes when a third party measures
The ARC Prize verified Gemini 4 Argon on the day it shipped; Vals AI published its Terminal-Bench measurement; ANPD tested three Brazilian AI systems. In every case, the third party did not end the doubt — it showed where it lies. Five movements on measuring what someone else claims.
- IVerified, with two yardsticks
- IIWho is better at the terminal depends on who measures
- IIIThe same name, six points more
- IVThe week preview became production
- VThe sandbox found what the benchmark does not
The floor fell to ten cents, and Macaíba slipped to November
Claude Haiku 5.5 arrived at a tenth of its predecessor's price, with a step hidden at 100 thousand tokens. Mistral opened a trillion-parameter model without saying under what license. A leaderboard corrected three tasks. And the PBIA supercomputer got a new date and new rules.
- ITen cents up to the step
- IIThe ten-cent model that solves terminal tasks
- IIIA trillion parameters, still without a license
- IVThe leaderboard that corrects itself
- VMacaíba postponed, and price came to count for less
The price list became the same; the bill did not
Three labs launched models with the same price list, to the cent. On the official Terminal-Bench leaderboard, the cost per solved task of those three runs from US$ 3.30 to US$ 35.78. Five movements on everything the per-token price stopped saying.
- IThe same model, four prices
- IISol ties Astra for a fifth
- IIIThree labs, the same price list
- IVSixty days or six months
- VThe calibration that held for three of four
The most expensive stopped being the best
Anthropic launched Opus 5.5 cheaper than Opus 5 and with the index's highest score; OpenAI delivered the same score for half the price. The top of the frontier now charges US$ 20, and the US$ 50 models fell five points behind. Five movements on what price stopped saying.
- IFirst place costs US$ 20
- IIThe same score for half
- IIISeventy point six, according to whom?
- IVThinking that won't switch off, and cache at 5%
- VThe second-best open model does not make the panel
The strongest open model costs 87 cents
Xiaomi published under MIT the open model ranked highest on the Artificial Analysis index, at US$ 0.87 per million output tokens. The same week, Grok 4.7 arrived with the same score at US$ 6. Five movements on the price an agent really pays — and on who measures what.
- IForty-six points for 87 cents
- IIThree numbers for the same model
- IIIA first place that lasts a week
- IVThe price that matters is the cache's
- VThe width of the scissors is a choice of method
What you call is not what answers
DeepSeek announced that `deepseek-v4-pro` would be served by another model, at another price; OpenAI turned the agent harness into a managed product; and the yardstick every comparison uses changed scale the Monday before. Five movements on what changes without your code changing.
- IThe alias that changes model
- IIThe harness became a product
- IIIEveryone lost ten points
- IVTwo scoreboards, two measures of the same model
- VThe call the panel did not count
The scissors: the frontier got pricier and the open model became pennies
In two weeks, two Chinese labs opened 320- and 770-billion-parameter models under MIT and Apache; OpenAI launched GPT-6 Astra at US$ 50 per million tokens and admitted that reading its reasoning has become less reliable; Fudan showed that saturated benchmarks hide half the bugs; and Brazil finally published a tender. Five movements on what intelligence costs — and what the price hides.
- IThe weight of the weights
- IIFifteen cents
- IIIThe number that no longer separates
- IVReading the reasoning
- VSeven thousand petaflops
Get the next Cinco
Tuesday, 7am BRT, in your inbox. Five items, with sources. No daily newsletter, no promotions, and unsubscribing is one click.