On September 30, Google announced the release of Gemini 4 Argon, a new artificial intelligence model with a significantly increased output token limit of one million tokens per response. This is a major leap from its previous models, which had a limit of 64,000 tokens. A token is roughly three-quarters of an English word, meaning one million tokens equate to about 750,000 words. This capacity is comparable to the entire text of War and Peace, with some extra space for One Hundred Years of Solitude. According to Google, Gemini 4 Argon is designed for complex tasks requiring extended reasoning, multiple trials, and corrections in a single execution. It is not aimed at casual users or novelists but rather at professionals needing long, uninterrupted AI assistance. Its main competitors, GPT-6 Astra and Claude Opus 5.5, have a maximum of 128,000 tokens per response, which is significantly lower than Argon's capacity. In terms of pricing, Google offers a promotional rate for the initial launch. One million input tokens cost about 1.70 euros before tax, while output tokens cost 8.50 euros. After the promotion ends, the cost for output tokens will double. However, the cost per task during the promotion is 1.99 euros, which is about 60% of the 3.26 euros required for GPT-6 Astra. Argon uses 62,000 output tokens per task, compared to 27,000 for Astra, and its token costs are five times lower. Once the promotion ends, the cost per task will rise to 3.98 euros, surpassing that of GPT-6 Astra. According to Artificial Analysis, GPT-6.1 Sol scores 52 points on its index, which is one point less than Argon, but costs 2.7 times less per task, even during Google's promotional period. In performance testing, Google's data shows that Argon scores 55% on FrontierSWE v2, a long-term software development benchmark, compared to 65.5% for GPT-6 Astra. However, the reliability of these scores is already in question among some observers. For developers who chain small requests, the increased token limit may not offer much benefit, as 64,000 tokens were already sufficient for most tasks. However, the new limit is particularly valuable for teams handling long-term tasks such as migration, security audits, or financial analysis, as it allows them to return to the task the next day without interruption. Users must set spending limits, as the model decides how many tokens it uses. A response reaching the one million token limit would cost about 8.50 euros before tax during the promotion and 17 euros afterward, not including input token costs. Google monitors Gemini 4 Argon's reasoning in real time, interrupts execution if it detects deviations, and manually audits each migration before deployment. METR, an organization that measures the length of tasks AI can complete independently, tracks cases of cheating in its tests and admits that its measurements are no longer reliable beyond sixteen hours. This suggests that even Google treats its model like a service requiring oversight, especially after a Gemini model infiltrated three companies during a test in May, as reported by Numerama. Access to Gemini 4 Argon remains limited. Google initially opens it to carefully selected cybersecurity experts through its Fairwind program. Pay API and AI Ultra subscribers will gain access later, but no specific dates have been announced. While the million-token limit serves as a marketing highlight, it also underscores a product designed to operate over extended periods with minimal user intervention. However, the longer the response, the higher the verification costs in human time, as Google manually proofreads each migration. As long as Gemini 4 Argon remains restricted to the Fairwind program, neither developers nor the press can test its full capabilities on a real long-term task.