Histoire
octobre 3, 2026
Google affirme que Gemini 4 Argon est une IA de pointe — alors pourquoi est-il toujours sous verrou ?
Google présente Gemini 4 Argon comme un retour très attendu dans l’IA à enjeux élevés, tandis que son déploiement étroitement contrôlé reflète à la fois les risques cybernétiques d’un modèle puissant et les questions persistantes sur la capacité des victoires aux benchmarks à se traduire par des performances fiables dans le monde réel.
Google’s new flagship arrives after a conspicuous gap. The company had indicated that its next major model would appear in June, but the expected Gemini 3.5 Pro never materialized; reports linked the delays to weak performance and internal morale problems. 1
On Wednesday, Google unveiled Gemini 4 Argon, arguing that the extra time has produced a model capable of sustained work in software engineering, finance, legal analysis and cyber defense. Sundar Pichai called it a system with “frontier performance in complex workflows, cyber defense and software engineering,” adding that Google teams were already using it extensively. 
2
The company’s evidence is ambitious. Google says Argon matches OpenAI’s GPT-6 Astra at the top of CWE-bench for finding and patching security flaws, sets a new mark on long-horizon software-engineering tasks, and has helped free more than 300 tebibytes of data-center memory without new hardware. 3 DeepMind chief Demis Hassabis also highlighted a separate Vals Index result placing Argon first at 68.9%. 
4
Yet Google is not treating that performance as a license for a broad release. Argon is initially going to vetted government and cybersecurity partners through the Fairwind Program while the company takes part in the US government’s voluntary pre-release review process. Pichai said the model has “frontier safeguards” and would be made available “as soon as we can and as safely as we can.” 5
Tulsee Doshi, Google’s Gemini product lead, framed the limited rollout as a way to get a strong defensive tool to the people who need it while building confidence: “a model of this caliber and this level of frontier performance is meaningfully important for defenders.” 6
That caution is also an implicit acknowledgment of the unanswered question. Bloomberg reported that some Google employees found Argon underwhelming in internal tests—an account Google disputed. The divide underscores the familiar frontier-AI problem: leaderboard scores can signal progress, but developers will decide whether Google is truly back only when Argon performs outside the lab. 1