Google makes Gemini 4 Argon broadly available with limited-time low API pricing of $2/$10 per 1M token

Google is making its new, flagship Gemini 4 Argon multimodal language AI model broadly available today, October 13.

The model was previously available only to a limited group of early testers. Now developers and enterprises are getting a powerful new option for software engineering, financial analysis, legal work, cybersecurity and scientific research — at limited-time discounted API prices substantially below some leading competitors.

The rollout comes just 13 days after Google first unveiled Argon on September 30, initially restricting access to approved cybersecurity defenders through its Fairwind Program while conducting additional safety testing and participating in a voluntary U.S. government pre-release model access process.

At the time, Google promised broader availability as soon as possible, but did not commit to a specific date. Now the company is offering Argon through the paid Gemini API, Google AI Studio, its Antigravity agentic development environment and eligible Gemini Enterprise subscriptions. Google AI Pro and Ultra subscribers can also access the model in the Gemini app and AI Mode in Search, although Antigravity access for individual subscribers requires the more expensive Ultra tier.

The launch gives enterprise customers their first widespread opportunity to test Google's most capable publicly offered model against rival systems from OpenAI, Anthropic and Meta, while providing fresh evidence of how Google itself is deploying Argon across complex engineering and business operations.

According to the company's new announcement, more than 100,000 Google employees use Argon through an internal version of Antigravity each month. Over a four-week period, Google says its engineering teams increased agentic code submissions by 35% while reducing post-submission code rollbacks by 30%.

But while Google's benchmarks show substantial improvements in professional and technical work, independent testing paints a more complicated picture of the model's competitive standing.

Sam Witteveen, CEO and co-founder of enterprise AI consulting and education company Red Dragon AI, and a VentureBeat host and partner, has been testing Argon ahead of its broader release.

"This is a really, really, really good model. But is it the best model out there? I actually don't think it is," Witteveen says in a video review shared with VentureBeat ahead of publication.

His assessment suggests Google's biggest achievement may be more strategic than technical: building a sufficiently capable general-purpose foundation for increasingly sophisticated products and enterprise workflows.

Google puts Argon to work on its own engineering problems

Google is backing its release with detailed examples of Argon performing tasks that would ordinarily require specialized engineering teams and substantial development time.

Perhaps the most striking example involves Chrome. Google says Argon helped reorganize a legacy folder containing approximately 6 million lines of browser feature code within Chromium's roughly 40-million-line codebase.

The model assisted in separating the browser's core webpage-handling capabilities from feature-specific code, enabling an architectural overhaul spanning hundreds of files.

Google says the work, previously expected to take months, was completed in weeks. The changes are intended to eventually allow webpages to move between tabs, side panels and floating windows without reloading, while eliminating certain crashes.

Elsewhere, Google's Agent Development Kit team tasked Argon with finding opportunities to improve the performance of its open-source Python framework for building AI agents.

After receiving a short objective and the team's existing instructions, Argon worked overnight and identified optimizations that Google says reduced in-memory session preparation time by 40% for long-running sessions and accelerated cold imports by 21%. The improvements have been incorporated into the public ADK Python codebase.

Google also reports that Argon helped optimize inference for its smaller Gemma 4 E2B model, increasing throughput from approximately 37,900 to more than 108,000 tokens per second on an Nvidia A100 GPU while preserving identical output.

These examples illustrate one of Google's central arguments: the model's value comes not merely from writing code, but from investigating large systems, proposing modifications, testing outcomes and iterating toward measurable improvements.

In another internal example, Argon examined 134 aging software bugs and identified 97 as stale or duplicated in under three hours, using commit histories and production telemetry to document its findings.

Google's examples offer unusually concrete descriptions of AI-assisted work, although they remain company-reported results rather than independently replicated enterprise evaluations.

Independent benchmarks: Argon leads professional work, but trails some rivals overall

Two major independent testing organizations offer contrasting perspectives on Argon's capabilities.

Vals AI, which evaluates models on professional tasks across finance, legal services, coding and tax work, ranks Argon first on its GDP-weighted Vals Index.

As of October 12, Argon scores 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%.

Vals also reports that Argon achieves a 65.40% score on its Finance Agent v2 benchmark, placing first among the models tested.

These results support Google's argument that Argon is particularly competitive on professional workflows requiring domain knowledge and sustained reasoning.

They also reinforce several advantages Google highlighted when it originally announced the model on September 30.

In the company's earlier benchmark comparisons, Argon scored 19.6% on Harvey's Legal Agent Benchmark, compared with 5.4% for OpenAI's GPT-6 Astra and 3.8% for Anthropic's Claude Opus 5.5. On Zapier's AutomationBench, which measures AI systems' ability to carry out multi-step business processes, Argon achieved 51.3%, ahead of Opus 5.5 at 42.5% and Astra at 41.4%.

The model also performed well on DeepSWE v1.1, a benchmark of extended software engineering tasks, scoring 77.9% against 74.2% for Opus and 74.1% for Astra.

However, Google's own comparisons showed areas where competitors maintained advantages. On Terminal-Bench 4.0, an evaluation of coding tasks performed through command-line tools, Opus 5.5 scored 66.4% versus Argon's 57.4%.

These are Google's previously disclosed benchmark comparisons, not a new set of independently replicated launch-day tests. They nevertheless help illustrate why different kinds of enterprise tasks can produce different winners.

More broadly, Artificial Analysis, which measures capabilities across reasoning, knowledge, mathematics, coding and agentic tasks, assigns Argon a score of 53 on its Intelligence Index.

That ties the model with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1, but places it below Claude Opus 5.5, Claude Sonnet 5.5 and Meta's Muse Wonder Full, released yesterday.

Sources: Artificial Analysis model evaluations, as available October 13, 2026. Higher scores indicate better performance. Results reflect the evaluated reasoning configurations, and not all models have the same public API availability.

That distinction matters for enterprise buyers. An aggregate benchmark can reveal broad technical performance, but it cannot establish which model will work best within an organization's particular software, data and operational environment.

Witteveen argues that Argon's strengths are especially apparent on tasks combining visual understanding, software generation and creative media production.

In one demonstration, he asked Argon to create a browser-based music production application modeled on Propellerheads' ReBirth software, complete with synthesized sounds, drum machines, filters and a step sequencer inspired by classic Roland equipment.

Using a relatively short prompt in the Gemini app, Argon produced an interactive interface along with functioning audio-generation code.

Witteveen also highlights Google's demonstrations involving Blender animations, 3D environments and interactive visualizations. These tasks require AI systems to generate code, render results, inspect their output and revise their work.

"This model seems to be very good at generalization," he says.

Nevertheless, he remains skeptical that Google's existing coding environment can displace established alternatives for developers focused primarily on autonomous software engineering.

"I still don't think, for example, Antigravity is anywhere near as good as Claude Code or Codex," Witteveen says.

His criticism distinguishes the underlying model's capabilities from the quality of the surrounding development tools — a consideration enterprises should weigh when evaluating complete AI coding systems rather than standalone API performance.

From cybersecurity to robotics, Google pushes beyond conventional coding

Google is also highlighting Argon's potential in cybersecurity, scientific computing and physical robotics.

The company says Wiz has used Argon through its Scan for Good initiative to identify and remediate a serious vulnerability in healthcare software. According to Google, the vulnerability exposed a master key that could allow administrator-level access and potentially compromise large quantities of sensitive patient information.

Google is expanding access through its Fairwind Program to hundreds of additional approved cybersecurity partners, including access to versions of the model without certain cybersecurity restrictions for authorized defensive work.

The company reports a 0.7% attack success rate on the Gray Swan Indirect Prompt Injection benchmark across 15 adaptive attempts. It also says Argon was trained to respect system boundaries and pause before irreversible actions, with unauthorized shortcut attempts dropping to 2% on its coding evaluations.

These figures are encouraging but do not eliminate the need for enterprise testing, access controls and human approval of consequential actions.

On the scientific side, Google's researchers report using Argon to accelerate quantum-computing calculations, optimize molecular modeling software and help design five new enzymes whose catalytic activity was subsequently confirmed through laboratory experiments.

Robotics presents another potential area of differentiation.

Google reports that Argon achieved an 87% score on the Embodied Reasoning Question Answering benchmark, which tests spatial and physical reasoning, exceeding Claude Opus 5.5 and GPT-6 Astra in the comparison supplied with its release.

In one demonstration, Argon studied video of a person playing cornhole, constructed simulation code and generated a motion sequence subsequently transferred to Apptronik's Apollo humanoid robot. Google says the robot successfully landed seven consecutive throws.

Another demonstration shows Argon generating a simulated robotic task environment from a hand-drawn sketch.

The broader implication is that Google's frontier model is being developed not just to interpret language or generate software, but to help create, evaluate and control simulated physical environments that can inform real-world robotics.

Pricing: $2 per million input tokens now, doubling December 1

Google is offering Argon at introductory rates of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input.

Unlike the September 30 announcement, which did not specify when the introductory period would end, Google now confirms that the reduced pricing expires November 30, 2026.

Beginning December 1, the standard public API price doubles to $4 per million input tokens and $20 per million output tokens.

That gives enterprise buyers a defined, relatively short window to evaluate Argon at lower prices before calculating the economics of longer-term production deployments.

Standard API list prices per million tokens. Prices exclude additional costs associated with tools, infrastructure and other services. Gemini 4 Argon's introductory cached-input rate is $0.10 per million tokens.

At its introductory price, Argon costs half as much per input and output token as Claude Opus 5.5 and one-fifth as much as GPT-6 Astra.

From December, its base price moves into the same category as Claude Opus 5.5 and Meta Muse Wonder.

Of course, raw token pricing tells only part of the story. Different models consume different amounts of reasoning tokens, require different numbers of attempts and vary in how effectively they complete a task.

Argon also offers an unusually large reasoning budget through the API. Google describes a unified limit of up to 1 million input and output tokens, using a continuation protocol intended for exceptionally demanding tasks.

That description introduces an important qualification to Google's original September 30 announcement, which described a new ceiling of 1 million output tokens, up from 64,000.

The new announcement instead emphasizes a combined input/output limit supported through a specialized API protocol. Google has not clarified in the supplied materials precisely how that limit relates to the maximum number of tokens developers can receive in one uninterrupted response.

Witteveen similarly cautions that developers should not mistake this specialized capability for the model's normal operating behavior in the Gemini app or other consumer products.

The extended capability requires API configuration, and he believes its strongest potential applications involve scientific or highly specialized workflows requiring hundreds of thousands of reasoning tokens.

For most enterprises, he recommends evaluating established workflows before redesigning systems to exploit the expanded limit.

Google's bigger opportunity: an AI foundation across its ecosystem

Argon's launch arrives as Google, OpenAI, Anthropic and Meta compete increasingly on the usefulness of their complete AI platforms, not simply model intelligence.

The broader release also represents an important moment for Google's AI organization, which has faced scrutiny over its pace of frontier-model development and a series of changes to its research leadership.

As VentureBeat reported when Argon was first announced, Google's previous flagship Gemini 3 family arrived in November 2025, while OpenAI and Anthropic subsequently introduced newer models that overtook it in important evaluations.

In July, Reuters reported investor concerns over delays to Gemini 3.5 Pro, substantial infrastructure spending and departures from Google's AI teams.

An August leadership reorganization reported by Axios saw longtime Google DeepMind CEO Demis Hassabis move into a chairman and chief scientist role at Alphabet, while Kavukcuoglu assumed leadership of DeepMind.

Against that backdrop, Google's ability to build, deploy and now commercialize Argon provides an answer to questions about whether the company can keep pace with the leading frontier-model developers — even if it does not decisively surpass them across every evaluation.

Google's advantage is the breadth of infrastructure and products through which it can distribute Argon.

The model is available in the Gemini API and Google's enterprise offerings, while Antigravity integrates it into coding workflows. Android developers can use it through Android Studio and Firebase AI Logic.

Meanwhile, the Gemini app gains more agentic functionality, including persistent tasks that can continue while users are away from their computers and connections to more than 40 external applications.

Witteveen sees this product ecosystem as central to Google's strategy.

He argues that Google does not necessarily need to win every independent evaluation to succeed with Argon. Instead, it needs a sufficiently advanced frontier model that existing Google customers can incorporate into increasingly capable software and business processes.

That could be particularly important for enterprises already invested in Google Cloud, Gemini and Workspace, which may prefer extending their current deployments rather than adopting another vendor's AI stack for every specialized task.

Witteveen also expects Google to continue improving the Gemini 4 family, although the company has not announced a public release date for a successor to Argon.

Indeed, Business Insider reported on October 10 that Google employees were already testing updated internal versions of Gemini 4 under the codenames Barium and Carbon, with some testers reporting substantially improved coding capabilities. Google declined to comment on the internal testing or its future release plans.

That possibility underscores how quickly the competitive landscape may change, even before Argon's introductory pricing expires.

For enterprise technology leaders, the immediate question is therefore less whether Argon can claim the highest score on every AI leaderboard, and more whether its combination of reasoning, visual intelligence, automation, development-tool integration and pricing can deliver reliable results on their own workloads.

The answer will depend on production testing. But with Argon now broadly available to paying customers, Google has substantially strengthened its position in the competition for enterprise AI deployments.