Today for AI
HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 10/16/2026

Google Launches Gemini 4 Argon: Tops Benchmarks but Lags in Math and Coding

19 reports archived0 independent sourcesupdated 10/19/2026, 11:13:04
Synthesis & Latest Updates

Google has officially launched Gemini 4 Argon, its new flagship model that surpasses Claude Opus 5.5 in the Vals AI Professional Index with a score of 68.90% and leads in embodied reasoning benchmarks. As the first top-tier proprietary model from DeepMind in seven months, Argon focuses on complex multi-step tasks, featuring a 1 million token unified context window and a 24-hour 'long output' mode for autonomous agent collaboration.

Cross-verified reports highlight Argon's superior cost efficiency, with testing costs approximately half those of competitors. Key technical advancements include reduced input token consumption, built-in deviation monitors for safety, and the lowest hallucination rate (15%) among mainstream models in AA-Omniscience tests. While matching GPT-6 Astra in general intelligence indices, some evaluations note it still lags behind rivals in pure mathematical derivation and competitive coding, occasionally exhibiting overconfidence in complex agentic scenarios.

The release signifies Google's strategy to tighten access to high-end models, relegating free users to the Flash tier while positioning Argon as the new premium standard. With introductory pricing set to double after November 30, this move aims to reclaim market leadership through cost-effective performance and accelerate commercial adoption in finance, law, and robotics.

LATEST/Gemini 4 Argon is fully available via API, with introductory pricing ending on November 30 before doubling. Recent studies demonstrate four Argon agents autonomously completing Minecraft survival mode in 27 hours without human intervention, showcasing strong self-organizing capabilities. Developers are leveraging its Real2Sim technology for robotics, though caution is advised regarding its performance in extreme mathematical logic tasks.

Coverage timeline

Total 19 reports · Latest first
  1. InfoQ 中文T2·88 pts

    Google Launches Gemini 4 Argon for Complex Reasoning and Code Refactoring

    • Gemini 4 Argon ranks first on the Vals index, excelling in multi-step complex tasks across finance, coding, and law.
    • Internal Google data shows Argon increased agent code submissions by over 35% while reducing code rollback rates by nearly one-third.
    • Demonstrated powerful long-context handling by helping the Chrome team separate 6 million lines of legacy code from a 40-million-line repository, compressing months of work into weeks.
    • Security-wise, it reduced attack success rates to 0.7% in Gray Swan indirect prompt injection benchmarks and includes an automated deviation monitor to intercept unintended actions.
  2. The DecoderT2·78 pts

    Four Gemini 4 Argon Agents Autonomously Beat Minecraft

    • Four Gemini 4 Argon agents autonomously beat Minecraft for the first time without human input or predefined strategies.
    • The agents demonstrated advanced collaboration skills, including spontaneous task division, threat warning, recovery from teammate failures, and shared error logging.
    • The run took approximately 27 hours with a cost of $1,620, involving extensive code generation and inter-agent communication.
  5. 量子位官网T2·88 pts

    Google Launches Gemini 4 Argon: Matches GPT-6 Performance with Limited-Time API Discount

    • Gemini 4 Argon ties with GPT-6 Astra Max at the top tier (53 points) in the Artificial Analysis Intelligence Index.
    • At equivalent performance levels, Argon High's average cost per task is approximately 39% lower than GPT-6 Astra.
    • Limited-time launch pricing offers a 50% discount: $2/million input tokens and $10/million output tokens until December 1.
    • Experts note significant leaps in creativity and writing, though complex multi-agent tasks still show occasional overconfidence issues.
  6. The DecoderT2·78 pts

    Google Unveils Gemini 4 Argon: A Unified Model for Robot Simulation, Control, and Evaluation

    • Gemini 4 Argon unifies simulation environment creation, control code generation, and video evaluation, reducing toolchain switching costs.
    • The model enhances understanding of spatial relationships and physical laws, allowing robots to perform high-fidelity virtual rehearsals before touching real hardware.
    • Actual deployment still requires significant human intervention and auxiliary AI agents for state verification, indicating a gap from fully autonomous physical AGI.
  7. IT之家 · 智能时代T2·86 pts

    Google Launches Gemini 4 Argon: 1M Token Output, Tops Vals Index, Refactors Chrome Codebase

    • Gemini 4 Argon supports an output limit of up to 1,048,576 tokens, sharing a single 1-million-token context window for input, reasoning, and output.
    • The model ranks first on the Vals Index for economic domains, excelling in multi-step complex tasks across coding, finance, law, and cybersecurity.
    • It has been used by the Chrome team to refactor tens of millions of lines of code, reducing months of work to weeks, and can autonomously fix critical security vulnerabilities.
    • Introductory pricing is available ($2/M input, $10/M output), doubling after November 30.
  8. Rohan Paul (@rohanpaul_ai)T2·88 pts

    Google Opens Gemini 4 Argon to Paid Users: Leads Vals AI Index with 1M Output Window and Superior Cost Efficiency

    • Gemini 4 Argon is now available to paid API users, Ultra/Pro subscribers, and Google Antigravity platform users.
    • Scores 68.90% on the Vals AI professional index, leading Claude Sonnet 5.5 (67.04%) and Opus 5.5 (66.97%).
    • Features a 1 million token output context window, far exceeding the 64K limit of Gemini 3.1 Pro Preview.
    • Average test time is under 47 minutes at a cost of $15.68, significantly lower than Claude Opus 5.5's $32.14.
  9. The DecoderT2·88 pts

    Google Launches Gemini 4 Argon Globally as New Frontier Model, Replacing Pro Tier

    • Gemini 4 Argon is now globally available via API and key interfaces like Google Search AI mode.
    • Tier restructuring: Argon replaces Pro as the flagship; free/Plus users are limited to Flash models, while Ultra/Pro subscribers gain access to Argon.
    • Safety metrics mention prompt injection resistance rates for GPT-6 Astra and Claude Opus 5, though test condition parity is not explicitly confirmed.
  10. Artificial Analysis (@ArtificialAnlys)T1·86 pts

    Gemini 4 Argon GA: Matches GPT-6 Intelligence with Lowest Hallucination and Major Agentic Gains

    • Gemini 4 Argon is Google's first public proprietary model above the Flash tier in over 7 months, closing the high-end product gap.
    • It matches GPT-6 Astra's intelligence level while reducing cost per task by ~40%, with cache read discounts increased to 95%.
    • The model shows significant gains in agentic evaluations and achieves the lowest hallucination rate among top-tier models.
  11. 宝玉 (@dotey)T2·86 pts

    Google Launches Gemini 4 Argon: Targets Complex Reasoning Tasks with Introductory API Pricing at $2 Input and $10 Output per Million Tokens

    • Gemini 4 Argon is the first model in the Gemini 4 series, specialized for multi-step complex tasks (coding, finance, legal, security).
    • API uses introductory pricing: $2/million input tokens and $10/million output tokens; prices double to $4/$20 starting December 1.
    • Features a unified 1,048,576-token context window shared across input, thinking, and output; standard single-call output limit is 65,536 tokens.
    • Introduces a 'long-output' mode allowing background tasks to run for up to 24 hours, generating content in segments to bypass single-call limits.
  12. Hacker News AIT2·78 pts

    Google Releases Gemini 4 Argon: Scores 87% on ERQA Benchmark to Lead Embodied AI and Robotics Workflows

    • Gemini 4 Argon achieves 87% accuracy on the ERQA embodied reasoning benchmark, outperforming other frontier models
    • The model demonstrates strong physical spatial understanding and visual-semantic reasoning, enabling zero-shot robot control
    • It covers the entire robotics lifecycle, including MuJoCo simulation practice and Real2Sim twin reconstruction
  15. Google DeepMindT1·86 pts

    Google Releases Gemini 4 Argon: Scores 87% on ERQA Benchmark to Advance Robot Control and Real2Sim Reconstruction

    • Gemini 4 Argon scores 87% on the ERQA benchmark, demonstrating leading capabilities in solving 3D spatial problems.
    • Covers the full robotics lifecycle, including MuJoCo mobile manipulation, zero-shot control, and closed-loop steering of video world models.
    • Supports Real2Sim technology, enabling reconstruction of physically calibrated simulation twins from robot cameras or egocentric human videos.
  16. Google DeepMindT1·88 pts

    Google Launches Gemini 4 Argon: A Frontier Model for Complex Multi-Step Tasks with Global API Access and New Pricing

    • Gemini 4 Argon is globally available via API for developers and enterprises, integrated into Google AI Ultra/Pro subscriptions and Search AI Mode.
    • The model's core strength lies in handling complex, multi-step tasks across domains like coding, finance, legal, cybersecurity, and science, emphasizing deep reasoning and high precision.
    • Clear commercial pricing established: after the introductory period ends on Nov 30, 2026, costs are set at $4 per 1M input tokens and $20 per 1M output tokens.
  17. Ars Technica AIT2·78 pts

    Google Rolls Out Gemini 4 Argon to Paying Users with Enhanced Safety and Embodied Reasoning

    • Access is currently restricted to paying users, with lower-tier plans lacking full functionality.
    • An automated misalignment monitor has been added to halt processing upon detecting unintended workflows.
    • The model shows measurable improvements in embodied reasoning and successfully debugged legacy code in a drone simulator.
  18. The Verge AIT2·86 pts

    Google Launches Gemini 4 Argon: Targets Frontier Reasoning and Reduced Hallucinations to Rival Top Models

    • Gemini 4 Argon signals Google's return to the top tier of frontier models, claiming competitive reasoning capabilities.
    • Key improvements include reduced hedging and redundancy, aiming for higher quality responses in professional scenarios.
    • The AI competitive landscape remains volatile, showing how quickly power balances can shift through rapid iteration.
  19. Google AI BlogT1·85 pts

    Google Launches Gemini 4 Argon: A Frontier Model for Complex Multi-Step Tasks with Global API Access and $4/M Input Pricing

    • Gemini 4 Argon is now globally available via API for developers and enterprises.
    • The model excels at complex, multi-step tasks in domains like coding, finance, and law, outperforming peers on benchmarks.
    • Public API pricing is confirmed: $4/1M input and $20/1M output tokens effective after November 30, 2026.