Google Gemini 4 Argon

Overview

Google Gemini 4 Argon is a frontier model released by Google DeepMind on 2026-09-30, the first product in the Gemini 4 series, announced in a post by DeepMind SVP and Google Chief AI Architect Koray Kavukcuoglu. This is Google's first frontier proprietary model in nearly seven months to surpass the Flash tier, and the repeatedly delayed Gemini 3.5 Pro has been confirmed as canceled. Argon is aimed at deep reasoning for long-horizon complex workflows, with a focus on real-world software engineering, enterprise knowledge work such as legal and finance, cybersecurity defense, and creative writing. Its output token limit has increased from 64K in the previous Gemini generation to 1 million, allowing a single response to handle long-output tasks such as large-scale code migration.

On 2026-10-13, Google made Argon available worldwide to developers, businesses, and consumers: developers call it through the API in AI Studio and Google Cloud, Google AI Ultra subscribers use it in Antigravity, and Ultra and Pro subscribers reach it in the Gemini app and AI Mode in Search, while free and Plus users are served the Flash models. Argon replaces the Gemini Pro class as the new frontier series. On 2026-10-30, Google released it as a generally callable tier, forming the second tier of the same generation alongside Carbon, which stays limited to qualified organizations; Argon is directly callable by developers and individual users. Against the backdrop of Gemini 3.5 Pro's failure, DeepMind leadership changes, and talent attrition, Argon is seen as Google's response to the narrative that it has fallen behind, and as a product that embodies its shift from emphasizing frontier capabilities to emphasizing cost-effectiveness.

Hydrogen, released November 12, fills in the execution side of this generation: Argon and Carbon keep the deep-reasoning role, Hydrogen finishes tasks quickly, and Orchestrator can switch between all three inside Antigravity.

Choosing among the three starts with arithmetic. Argon and Carbon both sit at 2 dollars input and 10 dollars output during the introductory period, moving to 4 and 20 on December 1, 2026, while Hydrogen is 1.50 and 7.50 with flat pricing across the full million-token window. The premium buys two different things: Argon and Carbon buy deeper thinking, Hydrogen buys speed. Between Argon and Carbon the split is token spend, with Argon leaner on output and Carbon spending more on exploration and verification. If the task only has to run, Hydrogen is enough, and Argon earns its price on long documents and wide code migrations that have to be reasoned through in one pass.

Key Features

• Million-level output tokens: The output token limit has increased from 64K in the previous Gemini generation to 1M (1 million), allowing a single response to handle long-output tasks such as large-scale code migration. The 2026-10-13 global release carries the one million token output limit.
• Long-horizon software engineering capability: It scored 77.9% on the DeepSWE v1.1 long-horizon software engineering benchmark, higher than GPT-6 Astra's 74.1%, Claude Opus 5.5's 74.2%, and Claude Fable 5.1's 67.4% in officially published comparisons. Cursor reports a CursorBench score of 52.8%.
• Enterprise knowledge work and business execution: It scored 51.3% on Zapier AutomationBench for end-to-end business execution, higher than Astra's 41.4%, Opus 5.5's 42.5%, and Fable 5.1's 31.4%; 19.6% on the Harvey legal agent benchmark, about three times the second place; and 65.4% on Vals Finance Agent v2, surpassing Claude Opus 5.5 and GPT-6 Astra.
• Cybersecurity defense and vulnerability remediation: It scored 68% on CWE-bench v1 vulnerability remediation, tied for first with GPT-6 Astra; it is first made available through the Fairwind program to trusted cyber defenders such as governments, critical infrastructure operators, and core technology platforms. That version does not include cybersecurity guardrails, while the public version retains refusals for cyber and CBRN attacks.
• Long-video understanding and creative writing: It scored 91.7% on LVBench for long-video understanding; the official materials list creative writing as one of its key scenarios.
• Third-party evaluation performance: On independent organization Artificial Analysis's Intelligence Index, Argon scored 53, tied with GPT-6 Astra and entering the top three intelligence tier; it topped the Arena.ai Text Arena leaderboard with 1,525 points; on Arena's Agent Arena, Argon (High) ranked 8th, with a net improvement of 7.92%. On 2026-10-15, ARC Prize published verified scores of 97.5% on ARC-AGI-3 under the provider adapter harness (49.8% under the standard harness), 91.3% on ARC-AGI-2, and 98.5% on ARC-AGI-1.
• Reasoning tiers and platform availability: On 2026-10-09, Google added a Gemini 4 Argon model reference inside Antigravity with low, medium, and high reasoning levels, and the Gemini web app now shows the same three tiers for every model, unifying the reasoning selector across tools. The model is available on OpenRouter, Cursor, and OpenCode, and can be accessed via Google AI Studio and Google Cloud.
• Research applications: On 2026-10-14, Google DeepMind published a cluster of research applications: researchers used it to find new counterexamples that further refute the Grad conjecture, which had stood in plasma physics for close to sixty years; to complete a round of enzyme engineering in a matter of hours; and to speed up work on quantum error correction and algorithm design. Google also announced free access to Gemini 4 Argon for ten thousand scientists.

Use Cases

• Long-horizon software engineering: used for debugging, migration of C/C++ to Rust codebases of up to more than 800,000 lines, and algorithm design.
• Enterprise knowledge work: scenarios such as legal agents, financial agents, and end-to-end business execution.
• Cybersecurity defense: made available through the Fairwind program to trusted cyber defenders, and Wiz has already used Argon in the Scan for Good project to scan public infrastructure.
• Long-video understanding: scored 91.7% on the LVBench long-video understanding benchmark.
• Creative writing: one of the key scenarios listed officially.
• Scientific research: refuting the Grad conjecture in plasma physics, enzyme engineering, quantum error correction, and algorithm design, with free access for ten thousand scientists.

Pros

• The output token limit has increased from 64K to 1 million, allowing a single response to handle long-output tasks such as large-scale code migration.
• Among 18 officially published comparisons, it leads in 12 and is tied for first in 1, covering scenarios such as long-horizon software engineering, end-to-end business execution, legal agents, financial agents, and vulnerability remediation.
• It leads on the Vals Index (weighted by the GDP share of U.S. industries).
• In third-party evaluations, it scored 53 on the Artificial Analysis Intelligence Index, tied with GPT-6 Astra and entering the top three intelligence tier; it topped the Arena.ai Text Arena leaderboard with 1,525 points; and ARC Prize verified 97.5% on ARC-AGI-3 under the provider adapter harness, 91.3% on ARC-AGI-2, and 98.5% on ARC-AGI-1.
• Internal usage results are notable: Google engineers have already used it for debugging, migration of C/C++ to Rust codebases of up to more than 800,000 lines, and algorithm design; Argon agents freed 300 TiB of memory in data centers by discovering optimization opportunities, and one agent rewrote a video decoding component to be 2.7 times faster than the original Rust version.
• Pricing is cost-effective: introductory pricing is $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached input tokens; Artificial Analysis puts its intelligence index level with GPT-6 Astra at roughly 60% of the cost per task.
• Broad availability: since 2026-10-13 it is available worldwide to developers, businesses, and consumers, and it landed on OpenRouter, Cursor, and OpenCode; on 2026-10-30 it was released as a generally callable tier directly usable by developers and individual users.
• Research impact: on 2026-10-14 Google DeepMind published applications including new counterexamples further refuting the Grad conjecture, enzyme engineering completed in hours, and acceleration of quantum error correction and algorithm design, alongside free access for ten thousand scientists.

Pricing

Introductory pricing is $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached input tokens; after the introductory period ends, it returns to $4 per million input tokens and $20 per million output tokens. Artificial Analysis puts its intelligence index level with GPT-6 Astra at roughly 60% of the cost per task. Google is giving 10,000 researchers free access to the highest Antigravity usage tier for six months. For pricing information not provided, refer to the official website.

Summary

Google Gemini 4 Argon is a frontier model released by Google DeepMind on 2026-09-30, the first in the Gemini 4 series, with its output token limit increased from 64K to 1 million. It focuses on long-horizon software engineering, enterprise knowledge work such as legal and finance, cybersecurity defense, and creative writing. Among 18 officially published comparisons, it leads in 12 and is tied for first in 1, and in third-party evaluations it entered the top three intelligence tier and topped Text Arena. It was first made available through the Fairwind program to trusted cyber defenders, and on 2026-10-13 it went global to developers, businesses, and consumers, landing on OpenRouter, Cursor, and OpenCode; on 2026-10-30 it was released as a generally callable tier alongside Carbon. On 2026-10-15, ARC Prize published verified scores of 97.5% on ARC-AGI-3 under the provider adapter harness, 91.3% on ARC-AGI-2, and 98.5% on ARC-AGI-1.

Version History

• Formally released as the generally callable tier (2026-10-30): Google released Gemini 4 Argon as a frontier model, forming the second tier of the same generation alongside Carbon, which stays limited to qualified organisations. Argon is directly callable by developers and individual users.
• ARC Prize publishes verified Gemini 4 Argon scores (2026-10-15): Scores 97.5% on ARC-AGI-3 under the provider adapter harness (49.8% under the standard harness), making Google the second frontier lab past 95% on the benchmark after OpenAI; also 91.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1.
• Research applications unveiled: free access for ten thousand scientists, with results in plasma physics and enzyme engineering (2026-10-14): On October 14, 2026, Google DeepMind published a cluster of research applications for Gemini 4 Argon. Researchers used it to find new counterexamples that further refute the Grad conjecture, which had stood in plasma physics for close to sixty years; to complete a round of enzyme engineering in a matter of hours; and to speed up work on quantum error correction and algorithm design. Google also announced free access to Gemini 4 Argon for ten thousand scientists.
• Gemini 4 Argon goes global with a one million token output limit (2026-10-13): On October 13, 2026, Google made Gemini 4 Argon available worldwide to developers, businesses and consumers. Developers call it through the API in AI Studio and Google Cloud, Google AI Ultra subscribers use it in Antigravity, and Ultra and Pro subscribers reach it in the Gemini app and AI Mode in Search, while free and Plus users are served the Flash models. Argon replaces the Gemini Pro class as the new frontier series, priced at an introductory 2 dollars per million input tokens and 10 dollars per million output tokens, with cached inputs 95 percent cheaper. Artificial Analysis puts its intelligence index level with GPT-6 Astra at roughly 60 percent of the cost per task. The model also landed on OpenRouter, Cursor and OpenCode, with Cursor reporting a CursorBench score of 52.8 percent. Google is giving 10,000 researchers free access to the highest Antigravity usage tier for six months.
• Google DeepMind's Gemini 4 Argon is now available on OpenRouter. (2026-10-13): Gemini 4 Argon is now available on OpenRouter and can be accessed via Google AI Studio and Google Cloud. The model scored 51.3% on AutomationBench and 91.7% on LVBench, and tied for first place on CWE-bench v1 with 68%; introductory pricing is $2/M for input and $10/M for output. Trial entry: https://openrouter.
• Gemini 4 Argon appears in Antigravity with unified reasoning tiers (2026-10-09): Google added a Gemini 4 Argon model reference inside Antigravity with low, medium and high reasoning levels. The Gemini web app now shows the same three tiers for every model, unifying the reasoning selector across tools.
• Gemini 4 Argon release (2026-09-30): Google DeepMind released Argon, the first frontier model of the Gemini 4 series, raising the output token limit from 64K to 1M for long-horizon software engineering, enterprise knowledge work in law and finance, and cyber defense. Google reports leads in 12 of 18 comparisons: 77.9% on DeepSWE v1.1, 51.3% on Zapier AutomationBench, 91.7% on LVBench, and a tie for first on CWE-bench v1 with GPT-6 Astra. Introductory pricing is $2 per million input tokens and $10 per million output tokens with 95% off cached input, later rising to $4 and $20. Access begins with trusted cyber defenders via the Fairwind Program, then paid API customers and Google AI Ultra subscribers