Gemini 4 Argon Officially Launches: 1M-Token Output and #1 on Vals Index

Google DeepMind has officially moved Gemini 4 Argon (model identifier gemini-4-argon) to general availability. Following closed evaluations with key infrastructure partners, the model is now available on Google AI Studio, Google Cloud, and the Gemini consumer app for Google AI Ultra and Pro subscribers. As the opening flagship of Google's fourth-generation Gemini ecosystem, Argon is no longer confined to standard conversational text generation—it explicitly targets long-horizon reasoning, enterprise-grade software engineering, and multi-step autonomous workflows.

1M-Token Output Decode: Breaking Through the Context Decay Bottleneck

Over the past two years, frontier models have expanded input context windows to millions of tokens, but sustained generation capability has remained constrained by architectural bottlenecks. Traditional Transformers often show degraded generation quality, looping repetition, and severe latency penalties after outputs exceed 16,000 to 64,000 tokens, forcing developers to build fragile iterative chunking pipelines to generate complex applications or detailed reports.

Gemini 4 Argon solves this fundamental obstacle through a dynamic sparse KV-cache architecture and multi-level speculative decoding on TPU v5p clusters. Argon can maintain mathematical and syntactic coherence across up to 1,000,000 output tokens in a single response stream. In practice, software teams can input system requirements and directly receive a complete, production-ready microservices code repository—including database migrations, Docker configurations, unit test suites, and OpenAPI documentation—without losing architectural consistency.

Gemini 4 Argon Technical Specifications

Parameter	Gemini 4 Argon	Industry Competitor Standard
Model identifier	gemini-4-argon	Proprietary closed endpoints
Input context window	1,000,000 tokens (native multimodal)	200,000 – 1,000,000 tokens
Output decode window	1,000,000 tokens (sustained generation)	8,192 – 64,000 tokens
Base input API price	$2.00 / 1M tokens	$3.00 – $15.00 / 1M tokens
Base output API price	$10.00 / 1M tokens	$15.00 – $60.00 / 1M tokens
Context caching discount	95% ($0.10 / 1M cached tokens)	50% – 75% discount

What is KV-cache? In the Transformer architecture, the model needs to "look back" at all previous tokens to compute attention when generating each new token. To avoid redundant computation, inference systems cache the already-computed Key and Value vectors at each layer—this is the KV-cache. Its VRAM/memory footprint is proportional to the "generated length," so when output extends to hundreds of thousands or even millions of tokens, the cache expands dramatically and becomes the primary bottleneck for long outputs.

Dynamic sparse KV-cache works by: rather than retaining the full cache for every historical token, it dynamically evaluates which historical positions are truly important for the current generation, pruning low-contribution cache entries or storing them at reduced precision, keeping only a sparse but critical subset. This controls memory growth while preserving coherence in long-range dependencies as much as possible.

Speculative Decoding is an acceleration technique: a smaller, faster "draft model" first guesses several tokens in succession, then the main model verifies those guesses in parallel all at once, accepting the correct parts and rejecting the incorrect ones. It can significantly boost throughput without changing the output distribution. Argon's "multi-level speculative decoding" on TPU v5p clusters means using multiple drafters of different scales in layered collaboration to match decoding efficiency in ultra-long output scenarios.

Vals Index Benchmarks: Dominance in Key Industry Evaluations

Unlike increasingly saturated synthetic multiple-choice benchmarks, the independent Vals Index measures real reasoning rigor in core economic domains such as legal statutory analysis, quantitative finance, complex software engineering, and corporate taxation. Gemini 4 Argon ranks first on the leaderboard, surpassing Anthropic Claude 4.5 Opus and OpenAI o1 across all evaluation pillars.

Vals Index and Engineering Benchmark Results

Evaluation Metric	Gemini 4 Argon	Claude 4.5 Opus	OpenAI o1
Vals Index composite score	89.4% (#1)	85.2%	84.1%
SWE-bench Verified (resolved issues)	68.2%	61.4%	58.7%
Financial modeling and SEC 10-K parsing	92.1%	88.6%	86.3%
Legal statutory conflict resolution	88.7%	87.1%	83.9%
Autonomous tool synthesis and invocation	91.5%	84.8%	82.0%

These figures show that Argon's advantage does not come from gaming a single dimension, but from maintaining leadership across multiple professional scenarios including law, finance, software engineering, and autonomous tool invocation. In particular, its autonomous tool synthesis and invocation score exceeds 91%, which directly underpins its agentic development capabilities in Antigravity.

The distinction between Vals Index and SWE-bench is worth explaining. SWE-bench Verified is a software engineering benchmark derived from real GitHub repository issues and pull requests; the model must actually modify code and pass the repository's own tests, so it measures "verifiable code-fixing ability" rather than text similarity. Vals Index, by contrast, leans more toward assessing economic value: its questions mostly come from real reasoning tasks in professional domains such as legal statutes, quantitative finance, and corporate taxation, and often have no single standard answer string—instead, domain experts or strict rubrics judge whether the reasoning process is rigorous, making it considered harder to "game" than traditional synthetic multiple-choice benchmarks (such as MMLU-type tests).

The "autonomous tool synthesis and invocation" item in the table examines whether the model can, given a goal, write its own tool functions (such as querying APIs, parsing files, performing calculations) and correctly orchestrate the call sequence—a core indicator of Agent capability. Argon scores 91.5% in this category, clearly higher than both competitors, which is also the direct basis for its ability to serve as the primary reasoning engine for autonomous agents in Antigravity—a model with weak tool-calling ability, even if strong in single-turn Q&A, would struggle to handle multi-step autonomous workflows.

Native Google Antigravity Integration: Agent-First Development

One of the core pillars of Argon's release is its deep integration with Google Antigravity. Antigravity is Google's flagship agent-first development environment, offering both a desktop IDE and a command-line interface. Google AI Ultra and enterprise subscribers can run Argon as the primary reasoning engine behind autonomous agent teams.

In Antigravity, Argon has direct shell execution, recursive file modification, and proactive multi-agent delegation capabilities. Faced with complex legacy system migration tasks, Argon can autonomously generate sub-agents to dissect database schemas, write integration tests, execute unit test runners, analyze compiler diagnostics, and iteratively fix issues without human intervention. The million-token output window ensures complete refactoring operations execute within a unified session without losing structural context.

The meaning of "agent-first development environment" and "multi-agent delegation." Traditional IDEs are "humans write code, tools assist"; an agent-first environment puts the AI agent at the center, with humans playing more of a goal-setting and review role. Antigravity offers both a desktop IDE and CLI, meaning the agent can operate a graphical interface as well as run in headless environments.

Multi-agent Delegation means the primary agent can break a large task into subtasks, dynamically generating and directing multiple "sub-agents" to work in parallel or in series—for example, one sub-agent reads the database schema while another writes integration tests, and the primary agent then aggregates results and decides the next step. This approach can break through the context and parallelism limits of a single agent.

Recursive file modification and shell execution give the agent real hands-on capability: the former means the agent can traverse directories and continuously edit multiple interdependent files; the latter means it can directly run build, test, and diagnostic commands and feed command output into the next round of reasoning. It is precisely the combination of "million-token output window + 91.5% tool-calling score + these execution capabilities" that allows complete refactoring to be completed coherently in a single session without artificially splitting tasks and losing structural context.

Aggressive Pricing and Enterprise-Grade Availability

Google has set aggressive enterprise migration pricing for Argon. Input tokens cost $2.00 per million and output tokens $10.00 per million, 30% to 50% lower than comparable frontier reasoning models. In addition, Google offers a 95% discount on cached tokens, reducing prompt re-ingestion costs to just $0.10 per million tokens, making continuous background monitoring, security log analysis, and automated customer workflows economically viable.

Strategic Assessment

Gemini 4 Argon marks a decisive shift in the AI race from conversational novelty to heavy-industry-grade practical tooling. With the #1 Vals Index ranking, an unrivaled million-token output decode window, and native execution capabilities in Google Antigravity, Google DeepMind has delivered an indispensable reasoning engine for modern developers, software architects, and enterprise technology leaders.

Originally published on Tech Boss

The context caching discount is key to understanding this pricing structure. When enterprises repeatedly use the same long prompt (for example, an entire body of regulations, a codebase, or a long-term security policy), they can cache these tokens server-side on the first request, and subsequent requests reuse the cached results directly without recomputation. Argon pushes the price of cached tokens down to $0.10 per million, equivalent to a 95% discount off the standard input price.

This is especially critical for high-frequency, long-context, highly repetitive-prompt scenarios such as "continuous background monitoring, security log analysis, and automated customer service"—their costs are almost entirely determined by the input side, and the caching discount can reduce total overhead by an order of magnitude, making previously uneconomical high-frequency inference workflows operational. By contrast, the industry generally offers only 50%–75% caching discounts, and many vendors' caches have short validity periods or minimum usage thresholds, making the 95% discount a substantive structural advantage for large-scale enterprise deployment, not just a marketing number.

One-line Summary

Google releases Gemini 4 Argon, entering enterprise-grade reasoning with million-token output, #1 on Vals Index, and Antigravity integration.

Paragraph Summary

Google DeepMind has moved Gemini 4 Argon to general availability on AI Studio, Google Cloud, and the Gemini consumer app. The model targets long-horizon reasoning, enterprise-grade software engineering, and multi-step autonomous workflows: through dynamic sparse KV-cache and multi-level speculative decoding, it achieves sustained coherent output of a million tokens in a single response, breaking through the quality decay bottleneck traditional models face after 16K–64K output. It ranks #1 on the independent Vals Index at 89.4% and leads Claude 4.5 Opus and OpenAI o1 across SWE-bench Verified, financial modeling, legal statutory conflict resolution, and autonomous tool invocation. Argon natively integrates with Google Antigravity, supporting shell execution, recursive file modification, and multi-agent delegation. Pricing is aggressive: $2/1M input, $10/1M output, with caching discounts up to 95%.