What Is Gemini 4 Argon? Pricing, Access and Benchmarks

Meet Gemini 4, Google's Most Ambitious Pre-Training Run Yet

Gemini 4 Argon is Google DeepMind's new frontier AI model, officially announced and launched in phases on September 30, 2026. Google initially released it to trusted cyber defenders through the Fairwind Program, with paid API customers and Google AI Ultra subscribers named as the next groups in line. The launch confirmed introductory pricing of $2 per million input tokens and $10 per million output tokens, an industry-leading 1 million-token output limit, and a focus on complex, long-horizon work across software engineering, enterprise knowledge, and cybersecurity.

Argon's architecture and parameter count remain undisclosed, and its initial availability is deliberately narrow. It is a launched model, but not yet a generally available one.

Updated 2026-10-06: Gemini 4 Argon officially launched on September 30, beginning with limited Fairwind access. Google has published pricing, benchmark results, an expanded output limit, and its broader rollout plan.

Key Takeaways

• Google officially announced Gemini 4 Argon on September 30, 2026, after confirming its pre-training run on July 21.
• The model launched first to trusted cyber defenders through Google's Fairwind Program rather than through a broad public API or app release.
• Argon is built for deep reasoning across long-horizon software engineering, enterprise knowledge, and defensive cybersecurity workflows.
• Google expanded the output limit from 64K tokens to 1 million tokens; it has not stated Argon's input context window.
• Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input priced at a 95% discount.
• Google reports 77.9% on DeepSWE v1.1, ahead of the comparison figures for Claude Opus 5.5 and GPT-6 Astra, but Argon trails Opus on Terminal-Bench 4.0 and the Artificial Analysis Intelligence Index.
• Broad Gemini API, Vertex AI, Gemini app, enterprise, and consumer availability still has no confirmed date.
• Gemini 3.5 Pro was skipped, making Argon the next flagship Gemini release rather than a parallel future model.

What Is Gemini 4?

Gemini 4 Argon is Google DeepMind's frontier model for complex workflows, particularly real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. Google officially introduced Argon on September 30, 2026 as a model built to sustain deep reasoning across long-horizon tasks.

The name first entered the public record on July 21, when Google said: "We have already started our most ambitious pre-training run yet, for Gemini 4, and can't wait to share more." Logan Kilpatrick, who leads Google AI Studio, repeated the confirmation on X the same day. Two days later, Sundar Pichai described Gemini 4 as a "significantly larger" frontier model and identified coding and autonomous agents as priorities.

Those statements are now historical development milestones rather than clues about an unreleased system. The resulting model is Argon, the first Gemini release to use a codename instead of the familiar Pro, Flash, and Flash-Lite tier labels. Google has not said what additional Gemini 4 models will be called.

The early interpretation that this would be Google's largest model was directionally about the run's ambition, but Google still has not disclosed a parameter count or detailed architecture. As Andrew Curran read the original announcement on X: "I take this to mean it will be the largest model they have ever trained." That remains an observer's interpretation, not a confirmed size specification.

Gemini 4 at a Glance

Field	Detail
Developer	Google DeepMind
Official model name	Gemini 4 Argon
Type	Frontier AI model
Status	Launched in phases on September 30, 2026
Positioning	Deep reasoning across complex, long-horizon workflows
Modality	Full input-modality specification not disclosed
Input context window	Not stated
Output limit	1 million tokens, up from 64K
Parameters	Not disclosed
Priority workloads	Software engineering, enterprise knowledge work, and defensive cybersecurity
Introductory pricing	$2/M input tokens; $10/M output tokens
Standard pricing	$4/M input tokens; $20/M output tokens
Cached input	95% discount from the input-token price
Availability	Fairwind trusted cyber defenders first; broader rollout pending
API and app access	Paid API and AI Ultra access planned next; no broad date announced
Open weights	Not announced
Official launch date	September 30, 2026

The distinction between launch and general availability matters here. Argon is an official product with published pricing and specifications, but Google is expanding access gradually while gathering feedback and iterating on safeguards.

How Gemini 4 Works and What Makes It Different

The exact technical architecture of Gemini 4 has not been disclosed. Google has, however, confirmed the product characteristics and workloads that distinguish Argon from the Gemini 3 generation.

Most Ambitious Pre-Training Run. This was Google's original description in July. The completed run produced a new frontier model rather than an incremental Gemini 3.x checkpoint. Google still has not published the parameter count, whether Argon uses a Mixture-of-Experts design, or other architecture-level details.

Long-Horizon Reasoning. Google says Argon was built to sustain deep reasoning across complex workflows. Its most visible specification is the 1 million-token output limit, increased from the previous 64K limit. That is an output ceiling, not a confirmed 1 million-token input context window; Google has not published the latter.

Coding and Autonomous Workflows. Coding remained central from the original training announcement through launch. Google says thousands of employees are already using Argon for specialized coding, deeper research, and writing. Argon agents are also working on C/C++-to-Rust migrations ranging from tens of thousands of lines to more than 800,000 lines for the Fuchsia Zircon kernel.

In one published internal example, Argon agents replaced about 32,000 lines of SIMD code in a Rust port of Google's libgav1 video decoder. The resulting decoder ran 2.7 times faster than the prior Rust port while producing identical video output. Google says those large-scale rewrites remain subject to automated and manual auditing, emulation testing, and review.

Scientific and Infrastructure Work. Google reports that Argon improved the spacetime resources of a quantum-computing subroutine by 40% in minutes. Another group of agents analyzed fleet-wide profiling telemetry and identified memory optimizations expected to free more than 300 TiB after rollout, with estimated total savings between 500 TiB and 1 PiB.

Emily, an ML observer, framed the pre-launch ambition this way on X: "all the research in the last two years on world model, multi modalities, long horizon agents and new TPUs will come together for the first time." The world-model and architecture implications remain speculative, but long-horizon agentic work is now part of Google's confirmed positioning.

What You Can Do With Gemini 4

For most developers and consumers, direct use is still gated. Gemini 4 Argon launched first to trusted cyber defenders through the Fairwind Program, while Google continues safety testing and guardrail iteration before expanding access.

Selected Fairwind participants can apply the model to defensive cybersecurity research. Google says Wiz used Argon through its Scan for Good initiative and found a critical vulnerability in healthcare software that earlier frontier models had missed.

Inside Google, Argon is already being used for:

• Specialized coding and large-scale codebase migrations
• Multi-round optimization and profiling work
• Enterprise research and knowledge workflows
• Long-form writing and analysis
• Quantum algorithmic optimization
• Data-center memory-efficiency work
• Defensive cybersecurity testing and vulnerability discovery

The 1 million-token output limit also gives the model room to reason and generate across unusually long workflows. Google has not published an equivalent input context-window figure, so it should not be described as a confirmed 1 million-token context model.

Until Argon's broader rollout reaches developers, Gemini 3.1 Pro remains an existing Google reasoning option. Our analysis of the delayed Gemini 3.5 Pro covers the intermediate flagship Google ultimately skipped before moving to Gemini 4.

How Gemini 4 Compares

A direct comparison is now possible, although the benchmark picture is mixed and Google's launch table had not been independently reproduced at release.

Metric	Gemini 4 Argon	Claude Opus 5.5	GPT-6 Astra
Launch date	Sep. 30, 2026	Sep. 22, 2026	Sep. 3, 2026
Input/output price per 1M tokens	$2 / $10 introductory	$4 / $20	$10 / $50
Output limit	1M	128K	128K
DeepSWE v1.1, vendor-reported	77.9%	74.2%	74.1%
Vals Index, vendor-reported	68.9%	67.0%	63.1%
Terminal-Bench 4.0, vendor-reported	57.4%	66.4%	58.2%
Artificial Analysis Intelligence Index	52.6	57.6	52.7

Argon's strongest launch result is DeepSWE v1.1, where Google's 77.9% figure leads both listed competitors. The comparison is less favorable on Terminal-Bench 4.0, where Opus 5.5 leads at 66.4%, and on the Artificial Analysis Intelligence Index, where Argon's 52.6 trails Opus 5.5's 57.6 and narrowly trails Astra's 52.7.

That makes the launch more competitive than the pre-release uncertainty suggested, but not a clean sweep. Reports before launch also described disagreement inside Google over whether strong benchmark results translated fully into practical coding performance. Google disputed that account and said there was broad internal consensus that Argon was at the frontier.

Our first look at GPT-6 covers one of Argon's closest competitors. Argon's introductory pricing undercuts the Astra pricing listed in the comparison, while performance varies by benchmark.

Availability: How to Access Gemini 4

Gemini 4 Argon has launched, but general access is not yet available. The first official rollout is limited to trusted cyber defenders through Google's Fairwind Program. Google is also participating in the U.S. government's voluntary process for pre-release model access while it expands availability.

Google says paid API customers and Google AI Ultra subscribers are next, followed by broader access for developers, enterprises, and consumers as soon as possible. It has not published a date for any of those stages.

On launch day, Argon had no published public API model ID and was absent from OpenRouter, Vertex AI, Gemini CLI, Cursor, and GitHub Copilot. Subsequent signals point to a gradual testing rollout and early-access infrastructure, but the bundle does not confirm broad production availability on any of those platforms as of October 6. A general Gemini app rollout is likewise not confirmed.

The practical access picture is therefore:

1. Available now: Selected trusted cyber defenders through Fairwind, plus Google's internal users.
2. Named as next: Paid API customers and Google AI Ultra subscribers.
3. Planned later: Broader developers, enterprises, and consumers.
4. Still missing: A public stable API model ID, broad Vertex AI listing, general Gemini app access, and a firm rollout date.

This is a phased launch, not an unreleased model and not a general-availability release.

What We Don't Know Yet

The launch resolved Gemini 4's name, date, initial access program, output limit, pricing, target workloads, and first benchmark table. The remaining open questions are narrower:

• Exact parameter count and underlying architecture
• Argon's input context-window size
• Its complete input-modality specification
• The date for paid API, AI Ultra, Gemini app, Vertex AI, enterprise, and consumer access
• The stable public API model ID
• Whether the broader Gemini 4 family will use Pro, Flash, and Flash-Lite labels or additional codenames
• How long introductory pricing will remain in effect
• Whether Google will publish weights
• Independent reproduction of Google's launch benchmarks
• The score behind Google's claim that Argon leads Gray Swan's indirect prompt-injection benchmark

Gemini 3.5 Pro's fate and Gemini 4's release timing are no longer open questions: Google skipped the delayed 3.5 Pro flagship, and Argon launched on September 30, 2026.

Frequently Asked Questions

What is Gemini 4?
Gemini 4 Argon is Google DeepMind's frontier AI model for complex, long-horizon workflows across software engineering, enterprise knowledge work, and defensive cybersecurity. Google officially announced it and began a phased launch on September 30, 2026.

When was Gemini 4 released?
Google officially announced Gemini 4 Argon on September 30, 2026, and began rolling it out that day to trusted cyber defenders through the Fairwind Program. Google has not given a date for broad public availability.

Is Gemini 4 available now?
Gemini 4 Argon is available only through a limited phased rollout, initially to trusted cyber defenders in the Fairwind Program. Broad Gemini API, Vertex AI, Gemini app, and consumer access are not yet confirmed; Google says paid API customers and Google AI Ultra subscribers are next.

How is Gemini 4 different from Gemini 3.5 Pro?
Gemini 3.5 Pro was skipped after its planned rollout stalled, while Gemini 4 Argon comes from Google's new, significantly larger pre-training run. Argon is positioned for long-horizon reasoning, coding, enterprise workflows, and cybersecurity rather than as an incremental 3.5-series update.

How much does Gemini 4 cost?
Google announced introductory Gemini 4 Argon pricing of $2 per million input tokens and $10 per million output tokens, with cached input priced at a 95% discount. The listed standard rates are $4 per million input tokens and $20 per million output tokens, but Google has not said how long introductory pricing lasts.

Will Gemini 4 be open source?
Google has not announced an open-weights release for Gemini 4 Argon. The confirmed distribution plan uses a controlled Fairwind rollout followed by hosted access for paid API customers and Google AI Ultra subscribers; whether weights will ever be released remains unanswered.

How does Gemini 4 compare to GPT-6?
On vendor-reported results, Gemini 4 Argon scores 77.9% on DeepSWE v1.1 versus 74.1% for GPT-6 Astra and 68.9% versus 63.1% on the Vals Index. Argon trails slightly on Terminal-Bench 4.0, 57.4% versus 58.2%, and on the Artificial Analysis Intelligence Index, 52.6 versus 52.7. These results were not independently reproduced at launch.

What to Watch Next

Three signals will move this page forward. First, watch for a stable Gemini 4 Argon model ID in the Gemini API or Vertex AI catalogs and confirmed access for paid API and Google AI Ultra customers. That will mark the shift from a security-first phased launch to usable developer availability.

Second, watch for independent benchmark reproduction. Google's DeepSWE result is strong, but Argon's lower Terminal-Bench and Artificial Analysis scores make real-world coding evaluations especially important.

Third, watch the rest of the Gemini 4 lineup. Google has not said whether Argon will stand alone or be followed by Pro, Flash, and Flash-Lite variants. Public Gemini app access, input-context specifications, and a full model card would resolve most of the remaining product questions.

Update — 2026-09-13

Several rumor accounts now claim Gemini 4's training has progressed beyond the stage described at publication. One post says pre-training is complete and post-training has started, with a first checkpoint already circulating internally; another frames the alleged early completion as the result of discoveries made during training (0x0SojalSec, 0x0Logicrw). These remain unverified claims, not evidence of a public model or confirmed training milestone.

The same speculation has pushed expected timing toward late September or October, but Google has announced neither a launch date nor Gemini 4 benchmarks, specifications, or access. For now, the confirmed public record remains unchanged; the new reports are best treated as signals of possible progress rather than a status update from Google.

While Argon's broader rollout continues, on kie.ai you can try Gemini 3.1 Pro, Gemini 3.6 Flash, and Claude Opus 4.8.