Google has announced Gemini 4 Argon, its new frontier AI model for complex, long-running work across software engineering, finance, legal research and cybersecurity. The launch immediately raises practical questions: How much will Gemini 4 Argon cost? When can ordinary users access it? What does the much-discussed 1 million token limit actually mean? And do Google's benchmark results show a clear lead over OpenAI and Anthropic?
Updated October 1, 2026. This guide separates confirmed launch facts from marketing claims and explains the benchmark numbers, pricing, access plan, security features and real-world use cases with links to the original sources.
Gemini 4 Argon at a glance
| Feature | Confirmed detail |
|---|---|
| Model | Gemini 4 Argon |
| Initial access | Trusted cyber defenders through Google's Fairwind Program and trusted testers |
| Next users | Paid API customers and Google AI Ultra subscribers, according to Google; no exact date announced |
| Intro input price | $2 per 1 million input tokens |
| Intro output price | $10 per 1 million output tokens |
| Cached input | 95% off the introductory input-token price |
| Post-intro price | $4 input / $20 output per 1 million tokens |
| Maximum announced output | 1 million output tokens, up from 64K in the previous limit Google cites |
What is Google Gemini 4 Argon?
Google describes Argon as the frontier model at the center of the Gemini 4 generation. Instead of optimizing primarily for short chat responses, the company is emphasizing what it calls long-horizon workflows: tasks that may require sustained reasoning, many steps, tool use and large amounts of generated work before completion.
The official launch highlights four broad areas: real-world software engineering, enterprise knowledge work, multimodal reasoning and defensive cybersecurity. Google says thousands of its own employees have already used Argon internally for coding, research and writing.
Independent reporting adds an important piece of context. Reuters reported that Argon is larger than Google's previous top-tier Pro models and that Google is positioning it against frontier systems from OpenAI and Anthropic. Reuters also noted that Google did not announce a specific date for general public availability.
Is Gemini 4 Argon available now?
Not to the general public. At launch, Google is rolling Argon out first to a limited set of trusted cyber defenders through the Fairwind Program. The company says it is also participating in the U.S. government's voluntary process for pre-release model access.
Google says the broader rollout will begin with paid API customers and Google AI Ultra subscribers, followed by wider developer, enterprise and consumer access. However, as of October 1, 2026, Google has not published a firm calendar date for that next stage.
This distinction matters for searches such as “Gemini 4 Argon release date” or “how to use Gemini 4.” The model has been announced and is in limited deployment, but that is not the same as a normal public launch inside the Gemini app.
Gemini 4 Argon pricing
Google's launch pricing is unusually specific. During the introductory period, Argon is set to cost $2 per million input tokens and $10 per million output tokens. Cached input receives a 95% discount, which implies $0.10 per million cached input tokens at the introductory input rate.
After the introductory period, Google says the standard rate will become $4 per million input tokens and $20 per million output tokens. Google has not attached an end date to the introductory period in the launch announcement.
For a simple cost intuition, 100,000 uncached input tokens would cost about $0.20 at the introductory rate, while 10,000 output tokens would cost about $0.10. Actual application costs will depend on prompt size, output length, caching, tool use and how often a workflow calls the model.
Does Gemini 4 Argon really have a 1 million token context window?
This is one of the easiest launch details to misstate. Google's announcement specifically says Argon has an industry-leading 1 million token output limit, increased from 64K. That means the model can potentially continue a single generated trajectory for an exceptionally long time.
An output limit is not automatically the same thing as an input context window. Unless Google separately specifies the input-context capacity for a particular API or product configuration, describing the launch simply as a “1M context window” can blur two different technical limits.
Why would such a large output allowance matter? Google argues that complex coding migrations, research jobs and multi-step agent workflows may need hundreds of thousands of tokens of intermediate reasoning and generated work before reaching a finished result. Most everyday chat users will never need anything close to that maximum.
Gemini 4 Argon benchmarks: where it leads and where it does not
Google published a broad benchmark set, and the results are more useful when viewed selectively rather than reduced to one “best AI” headline. The table below uses figures shown on Google's Gemini model page.
| Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| Vals Index — knowledge work | 68.9% | 63.1% | 67.0% |
| AutomationBench | 51.3% | 41.4% | 42.5% |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.6% |
| Harvey Legal Agent Benchmark | 19.6% | 5.4% | 3.8% |
| DeepSWE v1.1 — agentic coding | 77.9% | 74.1% | 74.2% |
| FrontierSWE v2 | 55.0% | 65.5% | 62.3% |
| Terminal-bench 4.0 | 57.4% | 58.2% | 66.4% |
| PostTrainBench | 45.3% | 44.3% | 49.3% |
The pattern is important: Argon posts strong results in several knowledge-work and long-horizon coding evaluations, but it does not lead every coding or science benchmark Google publishes. Benchmark scores also do not guarantee the same result for every real-world workload, and differences in tools, prompting, evaluation design and model configuration can matter.
What Google says Argon is already doing internally
Google provided several unusually concrete internal examples. These are company-reported results rather than independent audits, but they help explain the kind of workloads Argon is designed for.
- Quantum computing: Google says Argon improved the spacetime resources of a quantum-computing subroutine by 40% over a published baseline in one example.
- Data-center memory: a group of Argon agents reportedly analyzed fleet-wide profiling telemetry and helped identify optimizations that freed more than 300 TiB of memory after rollout, with Google estimating total potential savings of 500 TiB to 1 PiB.
- Large code migrations: Google says Argon agents are assisting C/C++ to Rust migrations ranging from smaller libraries to more than 800,000 lines in the Fuchsia OS Zircon kernel.
- Video-decoder optimization: Google reports that Argon agents replaced 32,000 lines of SIMD code in a Rust port of libgav1 and produced a version 2.7 times faster than that Rust port while preserving identical video output.
Why cybersecurity is central to the Gemini 4 Argon launch
Cybersecurity is not a side feature in this release. Google says Argon can autonomously find, validate and patch critical software vulnerabilities. Trusted defenders and Google's internal security teams are receiving a version without the cyber guardrails applied to normal users so that defensive specialists can use the full capability under controlled access.
Google says Wiz is already using Argon through its Scan for Good initiative, which searches for serious exposures in critical public infrastructure. In one case described by Google, Argon identified a critical vulnerability affecting healthcare software used by hospitals around the world and exposing sensitive personal information — a flaw that previous frontier models had missed.
On CWE-bench v1, which evaluates vulnerability remediation, Google reports that Argon tied for first with a score of 68%. On an internal Google vulnerability benchmark, the model reportedly found exposures across complex codebases written in 20 programming languages.
Why Google is restricting access at first
The same cyber capability that can help defenders can also create misuse risk. Google says it is strengthening safeguards in four areas before broad release:
- Misuse prevention: defenses against harmful cyber and chemical, biological, radiological and nuclear requests while trying to preserve legitimate research.
- Prompt-injection resistance: automated red teaming and adversarial training against malicious instructions hidden in external content.
- Misalignment monitoring: systems that monitor model reasoning and actions and can stop execution when behavior appears to move beyond the user's intent.
- Hardened environments: more isolated sandboxes for risky training and evaluation.
Google connects these measures to its Frontier Safety Framework. The company says Argon is its most resilient model yet against indirect prompt injection, although those claims will become easier to judge once independent researchers receive broader access.
Gemini 4 Argon vs OpenAI and Anthropic: what can we conclude?
The launch data supports a narrower conclusion than many comparison headlines suggest. Argon leads several of the benchmarks Google publishes, especially in enterprise knowledge work, automation and DeepSWE v1.1. But OpenAI's Astra and Anthropic's Opus 5.5 outperform it on some other coding and science evaluations in the same Google-published table.
Reuters similarly described Google's position as comparable to frontier competitors on important coding and cybersecurity benchmarks rather than presenting Argon as universally superior. Until independent users can test the production model across stable APIs, any broad “winner” claim should be treated cautiously.
When will Gemini 4 Argon be released to everyone?
There is no confirmed public-release date yet. Google's stated sequence is clear — trusted cyber defenders first, then paid API customers and Google AI Ultra subscribers, with wider developer, enterprise and consumer access afterward — but the company says expansion will depend on early feedback and further guardrail work.
That makes “Gemini 4 Argon release date” one of the most important details to watch. A specific date, supported API regions, rate limits and final product availability were not included in the initial announcement.
FAQ
What is Gemini 4 Argon?
Gemini 4 Argon is Google's new frontier AI model in the Gemini 4 family, designed for long, complex workflows in coding, enterprise knowledge work, multimodal analysis and defensive cybersecurity.
Can I use Gemini 4 Argon today?
Most users cannot. Initial access is limited to trusted cyber defenders and testers. Google says paid API customers and Google AI Ultra subscribers are next, but it has not announced an exact date.
How much does Gemini 4 Argon cost?
Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens. After the introductory period, Google says those rates will become $4 and $20 respectively.
Does Gemini 4 Argon have a 1M context window?
Google's launch announcement explicitly identifies a 1 million output-token limit. That should not automatically be described as a 1 million input-context window unless Google separately confirms the input-context specification for the product or API being used.
Is Gemini 4 Argon better than GPT-6 Astra or Claude Opus 5.5?
No single launch benchmark proves an overall winner. Argon leads several Google-published enterprise and coding evaluations, while Astra and Opus 5.5 lead some others. Real-world results will depend on the task and production configuration.
Why is Gemini 4 Argon focused on cybersecurity?
Google trained Argon to find, validate and patch vulnerabilities autonomously. That capability can strengthen defensive security, but it also raises misuse concerns, which is one reason access is being expanded gradually.
