The scissors: the frontiergot pricier and theopen model became pennies
In 2023, the most capable model in the world charged US$ 60 per million output tokens. In April 2025 it charged US$ 8. Last Thursday, GPT-6 Astra shipped at US$ 50 — the same price as Claude Fable 5.1, and two and a half times the model OpenAI launched in June. In the same two-week window, Z.ai put a 320-billion-parameter model, MIT-licensed, served according to the company only on Chinese chips, at US$ 0.50. The distance between the two ends is a hundredfold. That is what we call the scissors: the blades move apart, and what sits between them is the entire market of people who build with AI.
This edition's thesis is simple to state and hard to accept: price per token has stopped being the variable that decides between closed and open. For agentic work — code, terminal, operating systems — permissive open models have reached a capability band where the buying decision now depends on three other things: whether the model works in your domain (and the benchmarks every launch cites do not say), whether you can trust what it says it is thinking (GPT-6's own report says less and less), and where it runs (which, in Brazil, has become a R$ 959 million tender due on 8 October).
The five movements below follow that order: the weight of open weights, the price, the benchmark that no longer separates, the reasoning that will not be read, and the seven thousand petaflops of Macaíba. Each has a chart or a piece of executed code; the numbers are in the tables; the sources, numbered at the end. If you only have time for one: the second.
The five movements
- IOpen models
The weight of the weights
Tencent published a 770-billion-parameter mixture-of-experts under Apache 2.0. It costs to run like a 49-billion model — and needs memory for all 770. “Open”, at this size, is a word with an owner.
- IIInfra and cost
Fifteen cents
The closed frontier is back to what it cost in 2023. The open model that trails it closely costs a hundredth of that. The list price has become a positioning signal — and stopped being a measure of cost.
- IIIEvaluation
The number that no longer separates
SWE-bench Verified went from 33% to 96% in two years and stopped telling models apart. A Fudan group collected 119 real bugs from scientific software: the best agent solves fewer than half — and, with that sample, the top five are statistically the same.
- IVProduct · Interpretability
Reading the reasoning
GPT-6 Astra reset the ceiling of nearly every benchmark OpenAI chose to show. The system card, read to the end, says something else: chains of thought got shorter, and the company will have “significantly reduced confidence” in detecting some misaligned behaviors.
- VRegulation and Brazil
Seven thousand petaflops
After two years and three postponements, the Brazilian AI Plan's supercomputer has become a tender: R$ 959 million, 7,200 PFLOPS in FP16, Macaíba (RN), bids due 8 October. It is the first step with a number, a deadline and a document. It is also smaller, in money and in megawatts, than the plan promised.
Get the next Cinco
Tuesday, 7am BRT, in your inbox. Five items, with sources. No daily newsletter, no promotions, and unsubscribing is one click.
What we're building
- LiveWeekly edition
Cinco
One essay a week on the AI market, in five movements — with charts, executed code and numbered sources. Every Tuesday at 7am.
- LiveService for companies
Bespoke briefings
One question your company has about AI, answered in ten sourced pages and an hour of conversation — from what the Archive has already read.
- In constructionAgents product
Ofício
Agents that carry out the repetitive work of Brazilian companies — with verification at every step and a human in the loop.
- PlannedSemi-annual publication
GenAI BR Radar
What to adopt, trial, assess and avoid in AI, every six months, with a page per item and what changed since the last edition.
- PlannedAnnual report
State of AI Brasil
Annual survey of AI professionals in Brazil: what they use, what they earn, what is changing — open, with published data.
- PlannedProfiles and reading
Sala
One profile per reader: what you saved, the themes you follow, the trails you did — private by default, public if you want.
- ConceptMonthly index
Inference Cost Index
What it costs, in reais, to actually run AI: a fixed basket of tasks, repriced every month across open and closed models, served here and abroad.
- ConceptLiving database
Map of AI in Brazil
Who does AI in the country — labs, companies, public programs, datacenters, laws — every entry sourced and dated, on a map that updates with Cinco.
- And the whole catalogueAll projects→
What changed since the last edition
- Tencent open-sourced a 770B MoE under Apache 2.0 — and it edges Kimi K3 on code with a quarter of the parameters
- GLM-5.3-Flash: an MIT-licensed 320B model served entirely on Chinese chips, at US$ 0.15 per million tokens
- On real scientific software, the best coding agent solves fewer than half the bugs
- GPT-6 Astra: same price, new benchmark ceiling — and a system card that says reasoning is getting harder to monitor
- The government published the R$ 959 million tender for the Rio Grande do Norte AI supercomputer — bids due 8 October
- ReAct: interleaving reasoning and action is what made the first language agents work
Archive by theme
Archive→- Open models
Open weights, licenses, technical reports and what you can run yourself.
- Agents
Systems that plan, use tools and act — and how to tell when they work.
- Evaluation
Benchmarks, contamination, leaderboards and the gap between the score and real use.
- Interpretability
What happens inside the model: features, circuits, chain-of-thought faithfulness.
- Infra and cost
Inference, GPUs, price per token, and what changes when cost drops tenfold.
- Multimodal
Image, audio, video and text in the same model — and what it costs.
- Regulation and Brazil
PL 2338, ANPD, PBIA, the labor market and what is happening with AI here.
- Product
Launches, pricing, APIs and product decisions that change what can be built.
Today, one person: João Campista, in São Paulo. I read a couple of hundred things a week and pick five. Triage is AI-assisted; the reading, the judgment and the writing are mine.