AI models: price in reais, context and WhatsApp cost

1. Prices read from each company's official pricing page (the page and the date of each price are on the model page); open-weights models through the Hugging Face API, queried on Oct 9, 2026. Real at the PTAX selling rate of the Central Bank of Oct 9, 2026: R$ 4.9892 per dollar.
2. The conversation math: 10 agent replies, each reading 8,000 input tokens (instructions, what the company knows and the conversation history) and writing 300 output tokens. This is the basis CertuAI uses to estimate the cost per reply. It does not include cache discounts or the extra tokens of models that “think” before answering, so a reasoning model can come out more expensive. In a CertuAI measurement from Aug 5, 2026, with two real businesses on Gemini, the cost came to between R$ 72 and R$ 77 per 10 thousand messages, below this estimate.
3. Source: LMArena, text leaderboard, public dataset on Hugging Face (CC BY 4.0 license), published Oct 8, 2026, queried Oct 9, 2026. What the number measures: Score from the LMArena text leaderboard, in the style-control version and the overall category: people send the same question to two models without knowing which they are and vote for the better answer, and LMArena turns the votes into a score. The higher it is, the more often the model won. The confidence band is the interval LMArena itself publishes for the score. The rank is the one on the full leaderboard, which has models outside this catalog. How the name was matched: Same company and same name, ignoring periods, spaces and hyphens. It also accepts the name with the reasoning level at the end (high, medium, low, minimal) and the name with a date at the end when LMArena only knows that model by a date.
4. Source: Wikidata (CC0 license), queried Oct 9, 2026. What the number measures: Country (P17), founding date (P571), headquarters (P159) and official website (P856) of the company's item on Wikidata. How the name was matched: Item picked by hand and checked against the official site: the item's site must be on the company's domain.
5. Source: Open Portuguese LLM Leaderboard (open leaderboard of models in Portuguese) (Apache 2.0 license), queried Oct 9, 2026. What the number measures: Average, from 0 to 100, of nine tests in Portuguese: ENEM (Brazil's national high school exam), USP and Unicamp entrance exams (BLUEX), the OAB exam (Brazil's bar exam), sentence understanding (ASSIN2 RTE and STS, FaQuAD NLI), hate speech (HateBR and Portuguese Hate Speech) and sentiment in tweets (TweetSentBR). The leaderboard only evaluates open-weights models, running the weights at the indicated precision, and is archived (it does not take new models). How the name was matched: The evaluated repository on Hugging Face is the same official company repository the catalog uses for the model. The normalized average (NPM) subtracts what a random guess would get right on each test.
6. Source: Hugging Face, public models API, queried Oct 9, 2026. What the number measures: Data from the company's official repository on Hugging Face: the license declared on the model card, the total parameters of the safetensors files (when the repository publishes them), downloads in the last 30 days and all-time, likes, the main task, whether the download requires accepting terms first and the position limit from config.json (max_position_embeddings, a technical ceiling that is not the advertised context window). How the name was matched: Text-generation repositories from each company's official organizations are included: the 6 most downloaded in 30 days and the 6 with the most likes, without quantized versions, Base models, safety or test models, with at least 1,000 downloads in 30 days. A repository with the same name as a model on the same company's pricing page becomes a single model.

The official pricing pages

No API price here: Qwen (Alibaba Cloud's terms prohibit collecting content with bots without written authorization); Z.ai (Z.ai's terms prohibit bots and other automated means of obtaining data); Moonshot AI (Moonshot's usage agreement prohibits extracting data from the services by any means other than the API); Microsoft (Microsoft's terms of use prohibit copying or reproducing information obtained from the services); Cohere (Cohere's terms prohibit copying or storing a relevant part of the site's content); AI21 Labs (AI21's terms prohibit accessing the site's content by an automated process); rekaai (Reka's terms prohibit bots, scrapers and crawlers); writer (Writer's terms prohibit scraping and crawlers); perplexity (Perplexity's terms could not be read (the site blocks bots) and the page itself warns that the Sonar API stops being maintained on Sep 27, 2026). In these cases the price exists on the company's page, but was not read into this one; the open models of those companies appear without a price.

How the math is done: the price of an AI model is charged per token, which is a piece of a word. Input is what the model reads; output is what it writes. The price in reais is the dollar price times the PTAX selling rate of the last business day.

History: started on Sep 25, 2026. So far no price has changed since the first record; the list of drops appears on this page on the day a price changes.

What was left out of the catalog: from the official pricing pages, 32 models: 18 (voice, audio, image or robotics model: the per-token price is not for text conversation); 4 (retired from the company's API); 3 (old text-completion model, not a conversation model); 2 (voice model: the per-token price is not for text conversation); 2 (another company's model resold on this page); 1 (embeddings model, not a conversation model); 1 (marked as discontinued on the page); 1 (safety (moderation) model, marked as discontinued on the page).