What Is Gemini 4 Argon? 1M-Token Output
Daniel Okonkwo
Senior ML Engineer

TLDRGemini 4 Argon launched globally with long decode up to 1M tokens, $2/$10 introductory pricing, and access through Google’s paid API, Gemini app, Search, and Antigravity.
Inside Gemini 4 Argon: The 1M-Token Output Model
Gemini 4 Argon is Google DeepMind’s frontier reasoning model for long-horizon software engineering, enterprise knowledge work, defensive cybersecurity, and science. Google announced it on September 30, 2026, before officially launching it globally on October 13, 2026. It is now available through the paid Gemini API and Google’s consumer and enterprise products. Its defining specification is long decode up to 1,048,576 tokens within a unified input-output budget. Introductory pricing is $2 per 1 million input tokens and $10 per 1 million output tokens, according to the official announcement.
Updated 2026-10-14: Gemini 4 Argon officially launched globally on October 13 for paid API customers, Google AI Pro and Ultra subscribers in the Gemini app and Search, and Ultra subscribers in Antigravity.
Key Takeaways
- Gemini 4 Argon is a proprietary frontier model built by Google DeepMind and officially launched on October 13, 2026.
- The model uses a unified 1,048,576-token input-output budget and supports long decode up to that limit.
- It is available through the paid Gemini API, the Gemini app and Search for Pro and Ultra subscribers, and Antigravity for Ultra subscribers.
- Introductory API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens through November 30, 2026.
- Google reports 77.9% on DeepSWE v1.1 and 68% on CWE-bench v1.
- The official API model ID is
gemini-4-argon, with Low, Medium, and High thinking levels.
What Is Gemini 4 Argon?
Gemini 4 Argon is the first model in the Gemini 4 generation. Google describes it as a frontier model designed to sustain deep reasoning across complex, multi-step workflows rather than answer only short, isolated prompts.
Its target domains are unusually specific: real-world software engineering, legal and financial knowledge work, research, science, and defensive cybersecurity. Google has used Argon internally for coding, infrastructure optimization, technical writing, robotics, and quantum-computing research.
Argon is also the public model name, not merely an internal leak label. Google has not explained why it chose “Argon” or how that name maps to familiar Gemini tiers such as Pro, Flash, or Ultra. Google has indicated that other Gemini 4 versions are in development, but their names, specifications, and release dates remain unconfirmed.
The model completed its broader launch on October 13 after an initial Fairwind rollout to trusted cyber defenders. It is now globally available to paying developers, enterprises, and subscribers rather than being limited to the early-access program. For the broader generation-level background, see Meet Gemini 4, Google’s Most Ambitious Pre-Training Run Yet.
Gemini 4 Argon at a Glance
A third-party Artificial Analysis model profile records Argon’s benchmark, token-use, and pricing results. Google’s developer documentation now confirms the model’s official modalities, token limits, model ID, thinking levels, caching behavior, and API support.
| Specification | Gemini 4 Argon |
|---|---|
| Developer | Google DeepMind |
| Type | Proprietary frontier reasoning model |
| Modality | Text, image, video, audio, and PDF input; text output |
| Context window | Unified 1,048,576-token input-output budget |
| Maximum output | 65,536 tokens by default synchronously; up to 1,048,576 with long decode, minus consumed input tokens |
| Introductory pricing | $2 per 1 million input tokens; $10 per 1 million output tokens through November 30, 2026 |
| Standard pricing | $4 per 1 million input tokens; $20 per 1 million output tokens from December 1, 2026 |
| Cached input | 95% discount; implicit caching only |
| Availability | Paid Gemini API and Gemini platform; Pro and Ultra in the Gemini app and Search; Ultra in Antigravity |
| License | Proprietary |
| Model weights | Not available |
| Public API model ID | gemini-4-argon |
| Thinking levels | Low, Medium, and High |
| Announcement date | September 30, 2026 |
| Global launch date | October 13, 2026 |
How Gemini 4 Argon Works and What Makes It Different
Google has not published Argon’s parameter count, underlying architecture, training-compute figure, routing design, or complete training recipe. Its “frontier” designation describes capability and positioning, not a disclosed parameter scale.
The clearest difference is Million-Token Output. Argon has a unified 1,048,576-token budget that includes input, thinking, function calls, and final output. Its standard synchronous output limit is 65,536 tokens, but long decode can extend generation across continuation requests up to the remaining unified budget.
Argon’s defining specification is therefore long decode up to 1 million tokens, not an independent 1 million-token output allowance on top of a separate input window.
That ceiling supports what Google frames as a Long-Horizon Workflow: a task that may involve repository exploration, debugging, testing, migration, documentation, and repeated correction. More output capacity does not guarantee correctness, however. It can also raise latency and cost if a task consumes the full allowance.
At introductory rates, generating 1 million output tokens would incur $10 in output charges, apart from any input-token charges.
The second differentiator is High Reasoning, one of three official thinking levels alongside Low and Medium. Artificial Analysis recorded 53 points for Argon High on its Intelligence Index and a 15% hallucination rate on AA-Omniscience. That result indicates greater willingness to avoid unsupported answers, although the same evaluation placed Argon’s answer accuracy below GPT-6 Astra.
The third differentiator is the Cached Input Discount. Reused prompt prefixes receive a 95% discount from the normal input-token price through implicit caching. Explicit context caching is not supported. This matters for agent systems that repeatedly send the same repository context, policy library, or enterprise knowledge base.
What You Can Do With Gemini 4 Argon
Google’s strongest use cases involve sustained technical work rather than ordinary chat:
- Repository-scale engineering: Argon scored 77.9% on DeepSWE v1.1, a benchmark for long, multi-step software-engineering tasks.
- Code migration: Google says Argon agents are migrating C and C++ codebases to Rust, including projects exceeding 800,000 lines of code.
- Compiler-guided optimization: An Argon workflow replaced 32,000 lines of SIMD code in the libgav1 video decoder. Google says the resulting safe Rust implementation ran 2.7 times faster than the previous Rust port.
- Infrastructure optimization: Argon agents identified changes that freed more than 300 TiB of memory across Google data centers. Google estimates eventual savings between 500 TiB and 1 PiB.
- Quantum research: Google has reported both a 40% improvement over one published quantum-algorithm baseline and a formally verified subroutine with more than 100 times lower spacetime cost than another published baseline.
- Defensive cybersecurity: The model can discover, validate, and help patch vulnerabilities, including generating proof-of-concept evidence for authorized defenders.
- Enterprise analysis: Legal, finance, tax, research, and business-automation evaluations form a major part of Argon’s benchmark positioning.
- Robotics and simulation: Google demonstrated Argon mapping physics from a video, writing a simulation, and teaching an Apptronik Apollo humanoid to sink seven consecutive cornhole throws.
These deployments and demonstrations are vendor-reported. Google says critical code migrations still undergo automated testing, emulation, manual audits, and human review before production deployment.

Source: @kimmonismus
How Gemini 4 Argon Compares
Argon’s benchmark profile is competitive but not a clean sweep. It leads several coding, long-context, and knowledge-work tests while trailing on some terminal, expert, and agent evaluations. The comparison also mixes vendor-reported results with third-party measurements.
| Metric | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| Input price per 1 million tokens | $2 introductory; $4 standard | $10 | $4 |
| Output price per 1 million tokens | $10 introductory; $20 standard | $50 | $20 |
| Maximum output | Up to 1,048,576 tokens with long decode | 128,000 tokens | 128,000 tokens |
| DeepSWE v1.1 | 77.9% | 74.1% | 74.2% |
| Artificial Analysis Intelligence Index | 53 points, High | 53 points, Max | 58 points, Max with fallback |
| Terminal-Bench 4.0 | About 57% | About 59% | About 64% to 66% |
Arena placed Gemini 4 Argon High at number 1 in Text Arena with 1,525 points and number 8 in Code Arena: WebDev with 1,679 points. After the public launch, Argon also appeared at number 5 in Arena’s overall ranking and in the top five for Agent Arena. The Arena announcement described a blended cost of $8 per 1 million tokens.

Source: @arena
Those rankings depend on the evaluator, date, and task slice. Before the broader launch, an Arena CEO commentary excerpt said Argon fell from number 1 overall to number 7 with expert judges and number 10 on agentic tasks. A separate post-launch repository test also placed it below Astra, Fable 5.1, and Opus 5.5 on a set of previously missed bugs.
Argon is a benchmark leader in several categories, but the evidence does not support calling it the best model for every workload.
For additional context on its main OpenAI comparison point, see the earlier analysis of the GPT-6 Astra release.
Availability: How to Access Gemini 4 Argon
Gemini 4 Argon officially launched globally on October 13, 2026, following its initial Fairwind-First Rollout. Access began with vetted cyber defenders participating in Google’s Fairwind Program while Google took part in the U.S. government’s voluntary pre-release model-access process.
Developers and enterprises can now use gemini-4-argon through the paid Gemini API in AI Studio and the Gemini platform on Google Cloud. Google AI Pro and Ultra subscribers can access Argon in the Gemini app and AI Mode in Search. Google Antigravity access is limited to Ultra subscribers.
The official model supports global, U.S., and EU regions in Agent Studio, along with Low, Medium, and High thinking levels. Google has not published a free API tier for Argon, and public rate limits and production service-level terms still depend on the account and platform.
Gemini 4 Argon is not hosted on kie.ai. Teams comparing frontier models can evaluate the GPT-6 Astra model page as a separate available option.
Users should still avoid unofficial “early access” login pages. Community researchers identified impersonation accounts advertising fake access before the official launch, and legitimate access now comes through Google or established integrations.
What We Don’t Know Yet
Several specifications and operational details remain open:
- Google has not disclosed the parameter count, model architecture, training-compute figure, training recipe, or knowledge cutoff.
- Public rate limits, sustained throughput, latency under long decode, and production service-level terms remain incompletely documented.
- The relationship between Argon and future Gemini 4 Pro, Flash, or other variants remains unconfirmed.
- Independent production testing is still developing, and post-launch benchmark results vary substantially by harness, evaluator, and task.
- Google has not released model weights or self-hosting rights.
- The practical reliability and cost of trajectories approaching the full 1,048,576-token budget remain uncertain.
These gaps matter because a million-token unified allowance can create substantial latency and token consumption. Price per token alone does not establish cost per completed task.
Frequently Asked Questions
What is Gemini 4 Argon?
Gemini 4 Argon is Google DeepMind’s proprietary frontier reasoning model for long-horizon software engineering, enterprise knowledge work, defensive cybersecurity, science, and autonomous multi-step agents. It is the first model in the Gemini 4 family, and its headline capability is long decode up to 1,048,576 tokens within a unified input-output budget.
Is Gemini 4 Argon publicly available?
Yes. Gemini 4 Argon launched globally on October 13, 2026. It is available to developers and enterprises through the paid Gemini API in AI Studio and the Gemini platform, to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and to Ultra subscribers in Google Antigravity.
How much does Gemini 4 Argon cost?
Gemini 4 Argon costs $2 per 1 million input tokens and $10 per 1 million output tokens during its introductory period through November 30, 2026. Standard pricing of $4 per 1 million input tokens and $20 per 1 million output tokens begins December 1, while cached input receives a 95% discount.
Does Gemini 4 Argon have a 1M-token context window?
Yes, but it uses a unified 1,048,576-token budget shared by input and output. The default synchronous output limit is 65,536 tokens, while long decode can extend generation up to 1,048,576 tokens minus the input, thinking, and function-call tokens already consumed.
Is Gemini 4 Argon open source?
No. Gemini 4 Argon is a proprietary hosted model. Google has published the API model ID gemini-4-argon, but it has not released model weights or self-hosting rights.
Gemini 4 Argon vs GPT-6 Astra: which is better?
Neither model is universally better across every evaluation. Gemini 4 Argon leads the reported DeepSWE v1.1 result at 77.9% versus GPT-6 Astra’s 74.1%, while both score 53 points on the Artificial Analysis Intelligence Index. Other tests place Astra ahead on some terminal and bug-fixing tasks.
What can Gemini 4 Argon do?
Gemini 4 Argon can handle long-running coding, code migration, research, legal and financial knowledge work, defensive cybersecurity, scientific problem-solving, video understanding, and autonomous multi-step workflows. Google also uses Argon agents for data-center optimization, quantum research, robotics, and large C/C++-to-Rust migrations.
What to watch next: Google’s rate-limit and service documentation should clarify production throughput and the practical behavior of very long decodes. The other critical signals are independent repository-scale evaluations, real-world cost per completed task, and the specifications of future Gemini 4 variants.
Update — 2026-10-11
New evidence places Gemini 4 Argon inside Google’s operational tooling. Community researchers reported the model appearing in Google Cloud and as the default model in the Antigravity CLI. A separate Vertex AI sighting showed the model ID gemini-4-argon-cyber-permissive. These appearances indicate active preparation for broader access, but they do not by themselves establish a public API endpoint or general availability.
The strongest real-world usage signal came on October 9, when a merged GitHub pull request disclosed “Generated-by: Gemini 4 Argon.” That is evidence of Argon contributing to production-adjacent development work, while access for ordinary developers and consumers remains unconfirmed.
Try related models on Kie.ai: Gemini 4 Argon, Claude Opus 5.5 and Claude Sonnet 5.5.
About Daniel Okonkwo
Daniel writes about inference systems, model architecture, and what new releases actually change for builders.
View all posts by Daniel Okonkwo