Gemini, the artificial intelligence developed by Google, uses significantly fewer tokens than its competitors, while also being less expensive. The average cost of a query on the latest models is approximately 1.2 cents. However, it fails to correctly respond to the Extension prompt. After Claude from Anthropic and ChatGPT from OpenAI, it is now Gemini from Google's turn. The artificial intelligence of the search engine is integrated into many products, starting with Android. It is therefore widely used around the world, and unlike its two competitors, the question of cost is less important due to this deep integration. As in previous analyses, the API was used to determine the number of input and output tokens, which also includes those used during reasoning, if applicable. Several different models were tested, from Gemini 3.1 to 3.8 flash, the latest version available. Google offers four levels of reflection with Gemini 3.1 to 3.6, with minimal, low, medium, and high. With the 3.1 Pro, 3.7, and 3.8, only the last three (low, medium, and high) are possible. Via the API, Gemini 3.6 to 3.8 have a price of 0.75 dollars per million input tokens... but only until December 31, 2026. The same applies to output with 3.75 dollars per million tokens. Starting January 1, 2027, both prices will double, reaching respectively 1.5 and 7.5 dollars. The version 3.5 Flash is the most expensive, with 1.5 and 9 dollars respectively, while the 3.1 Flash Lite is cheaper with 0.25 and 1.5 dollars. The honor goes to Gemini 3.1 Pro with 2 and 12 dollars per million respectively. The prompts used were: "O brother, son of my mother, your father is the brother of my child. Who am I?" "Next summer I will have three weeks of vacation. I am with my small electric car Sprind (charge up to 30 kW max) and I would like to visit the south of Bayonne to Montpellier, taking my time. Propose me a travel journal." "I want an extension for Chrome and Firefox that allows me to invert the titles and subtitles on the Next.ink news site. If there is no subtitle, it is a brief. So make sure to invent one from the title of the brief. The extension should have a selector to enable/disable the inversion directly on the page." A new feature of the protocol includes verification of extensions. This was requested several times in the comments. An estimate of the validity of the results is now included for the extensions. The other two prompts are currently being validated, and future articles will return to them. A verification protocol was developed and used with Playwright and the Chromium browser. The extension files are automatically generated from the model's response, then the extension is loaded into the browser via Playwright. The extension is launched on a fake page (local html file) replicating the structure of Next, then live on Next.ink.