deepmind.google détecté sur le web

Google ouvre Gemini 4 Argon mondialement à un prix d'appel 2$/10$

Par Publié Mis à jour
7 médias qui couvrent ce sujet
Google Generative AI Coding Tools Robotics ai-business

TL;DR

  • Google's introductory $2/$10 per million token pricing doubles to $4/$20 on December 1, forcing enterprise teams to commit to Argon before year-end.
  • Bloomberg's internal sources say Argon underperforms rivals on front-end coding tasks despite leading DeepSWE and CWE-Bench benchmarks.
  • Four named launch clients, Deloitte, LexisNexis, Ambience Healthcare, and Isomorphic Labs, target consulting, legal, healthcare, and scientific research verticals.

Google rend Gemini 4 Argon disponible mondialement ce 13 octobre via l'API dans AI Studio et sur la plateforme Gemini, pour les abonnés Google AI Ultra dans Antigravity, et pour les abonnés Ultra et Pro dans l'app Gemini et AI Mode dans Search. L'annonce, signée Koray Kavukcuoglu, SVP de Google DeepMind et Chief AI Architect, s'accompagne d'un tarif d'appel: "The model launches today at an introductory price of $2 per million input tokens and $10 per million output tokens (with cached input tokens at 95% off)."

Sur les benchmarks revendiqués par Google, Argon affiche 77,9% sur DeepSWE v1.1 pour le codage agentique et se classe, selon le blog, "#1 on the Vals Index, which measures model performance across critical economic sectors like finance, coding, legal, and tax work." La fiche produit publiée par Google DeepMind place aussi le modèle à 68% sur CWE-bench v1 (remédiation de vulnérabilités) et met en avant "an industry-leading 1M unified input-output token limit via the Gemini API."

Google avance par ailleurs un gain mesuré en production: "Argon in Antigravity has helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third." Côté démonstration physique, le modèle décroche 87% sur Embodied Reasoning (ERQA) et pilote un pipeline Video-to-Sim-to-Real: "When transferred onto Apptronik's Apollo humanoid, the robot quickly sank seven consecutive shots in a real-world trial." La vignette robotique s'inscrit dans un flot soutenu d'alertes IA physique que nous suivons depuis trois mois.

Les scores DeepSWE, CWE-bench et Vals Index relayés aujourd'hui sont ceux publiés par Google, sans test tiers indépendant à l'appui. VentureBeat rapporte que la grille d'appel expire le 30 novembre 2026, après quoi le tarif public standard passera à 4$ par million de tokens en entrée et 20$ en sortie.

Ce qu'en disent les autres médias

Couverture consolidée 24h après publication

  1. Google Blog Lire →

    Official GA post announces pricing tiers, Fairwind-to-public expansion, four named enterprise clients, and positions Argon across software engineering, knowledge work, and cybersecurity.

    Argon delivers frontier performance to help in complex workflows across real-world software engineering, enterprise knowledge work, and cybersecurity defense.
  2. Bloomberg Lire →

    First tier-1 outlet to surface internal employee skepticism about real-world coding, particularly front-end tasks, despite benchmark leads; Google disputed the characterization.

    Some Google employees with direct access say Argon does well on benchmarks but less well when they put it to work, especially on certain coding jobs.
  3. CNBC Lire →

    Frames the GA through financial-sector adoption and the broader shift toward personal AI agents, adding a Wall Street demand lens absent from technical coverage.

  4. TechCrunch Lire →

    Provides competitive benchmarking via Vals against OpenAI Astra and Anthropic Fable; notes the staged Fairwind rollout strategy before the Oct 13 general availability date.

    autonomously find, validate, and patch critical software vulnerabilities.
  5. VentureBeat Lire →

    Foregrounds the December 1 pricing cliff and quotes a practitioner skeptic; frames Google's distribution stack across Workspace and Antigravity as the real competitive moat.

    This is a really, really, really good model. But is it the best model out there? I actually don't think it is. -- Sam Witteveen, Red Dragon AI CEO
  6. GeneLab Lire →

    Quantified practitioner critique: 5 benchmark losses cluster in terminal operations, and the post-December price jump compresses Argon's cost advantage over Claude Code from 59% to 18%.

    The price gap with Claude Code will shrink from a 59% discount to an 18% discount after the December 1st pricing change.

Article original publié par deepmind.google

Lire l'article original →

Titre original : Google déploie Gemini 4 Argon mondialement, 77,9% sur DeepSWE et sortie 1M tokens