Gemini 4 Argon Review: Features, Benchmarks & Pricing

Gemini 4 Argon is Google's new frontier model and the first of its Gemini 4 generation. After a limited preview that began on September 30, 2026, Google made it broadly available on October 13, 2026: developers can build with it through the paid Gemini API, and Google AI Pro and Ultra subscribers can use it in the Gemini app and AI Mode in Search. It targets long, complex work in software engineering, finance, legal research, science, and cybersecurity defense.

This Gemini 4 Argon review covers what changed at launch, what Google's benchmark results show, how much it costs, where you can access it, and where its limits sit. Argon is not integrated into Chat Smith yet, but you can already use Gemini 3.5 Flash and other Gemini models there, alongside models from OpenAI, Anthropic, xAI, and DeepSeek.

Gemini 4 Argon is Google DeepMind's most capable model to date, built for deep reasoning across long, multi-step workflows such as code migrations, financial analysis, and frontier research. Google says it outperforms other frontier models on benchmarks across several domains and ranks first on the Vals Index, which measures performance in finance, coding, legal, and tax work.

Its headline spec is length. Through the Gemini API, Argon supports an industry-leading 1-million-token unified input-output limit, and Google has added a new continuation protocol to help developers handle very long outputs. That gives the model room to work through hard problems and return large results, such as a full codebase refactor, within a single task.

Google also points to changes everyday users will notice. Argon follows style instructions better than any previous Gemini model, answers more directly with less hedging, avoids repeating itself, and skips clichés. Those improvements matter as much for emails and reports as they do for code.

Argon sits above the Flash models that make up most of Google's current lineup, such as the one covered in our Gemini 3.8 Flash review. If you want to see how Google's models compare with Anthropic's more broadly, our Gemini vs Claude comparison is a good starting point.

Five areas stand out in Google's launch announcement.

Argon set a new state of the art on DeepSWE v1.1 with 77.9%, a benchmark of real-world, long-horizon software engineering. Inside Google, teams using Argon in Google Antigravity, the company's agent-first development platform, increased agentic code submissions by more than 35% while cutting code rollbacks by nearly a third.

The examples are large in scale. Argon helped the Chrome team separate core webpage handling from a roughly 6-million-line legacy folder inside Chromium's 40-million-line codebase, finishing months of refactoring in weeks. Working overnight on Google's Agent Development Kit, it cut in-memory session prep time by 40% in long sessions, and in a few days it made Google's open Gemma models run 2.85 times faster with no drop in quality.

Argon leads the Vals Index and performs strongly on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark. On Zapier's AutomationBench, which tests end-to-end business workflows, it ranked first with 51.3%. For planning tasks such as drafting with an AI business plan generator, these are the gains most professionals would notice.

Document-heavy work is another strength. At Isomorphic Labs, Argon synthesizes complex PDFs and patents to pull out the key information, turning a process that usually takes weeks into hours. It is a far more demanding version of what an AI PDF summarizer does for everyday reading.

Argon is strong when work involves visuals: analyzing professional charts, acting on a series of documents, or finding details in long videos. On LVBench, a long-video understanding benchmark, it scores a state-of-the-art 91.7%, the kind of capability behind tools that summarize YouTube videos in seconds.

Spatial reasoning improved as well. Argon scores 87% on the Embodied Reasoning (ERQA) benchmark, which Google says is ahead of all other frontier models. In one internal test, it watched a raw video of a person playing cornhole, then wrote the simulation code that let a humanoid robot sink seven shots in a row.

Researchers are already using Argon on hard scientific problems. It helped find a broader class of counterexamples to Grad's conjecture in fusion research, nearly doubled the throughput of a quantum error-correction decoder, and topped Gamow Labs' RareBench (57%) and LabBench (52.2%) evaluations for rare-disease diagnosis and wet-lab reasoning. Google is also offering 10,000 scientists its highest Antigravity usage tier free for six months, subject to eligibility and selection.

Google trained Argon to find, validate, and patch critical software vulnerabilities. It ties for first place on CWE-bench v1 with 68%, and security firm Wiz used it to close gaps that exposed sensitive government data and hundreds of thousands of private patient records in healthcare software. At launch, Google began expanding its Fairwind Program to hundreds of additional cyber defense partners, who get a version of Argon without cyber guardrails.

The table below collects the key figures from Google's announcements. Most benchmark results come from Google or its partners, so treat them as a starting point until more independent tests are published.

Taken together, the numbers point to a model built for depth: long coding sessions, multi-step business tasks, scientific analysis, and security work. Google also reports leading results on the Vals Index, Vals Finance Agent v2, and Harvey's Legal Agent Benchmark.

For developers, Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% off. The introductory period ends on November 30, 2026, after which the price rises to $4 and $20. These rates apply to the API, not to Gemini app subscriptions.

Where you can use Argon depends on who you are. Google is rolling it out globally through these channels:

Gemini 4 Argon is not integrated into Chat Smith yet. You can use Gemini 3.5 Flash and other Gemini models today: free accounts include limited use, and Chat Smith Pro raises usage limits for $9.99 per month or $39.99 per year.

The first limit is the cost of entry. There is no free consumer access: the Gemini app requires a Google AI Pro or Ultra plan, the API requires the paid tier, and Antigravity access is tied to Ultra. Developers should also budget for the price doubling once the introductory period ends on November 30.

The benchmark results also come mostly from Google. Earlier press coverage noted that Argon still trailed rivals on some of the coding benchmarks Google included, so it is worth comparing it with models like GPT-6.1 Sol and Claude Opus 5 on your own tasks before switching.

Safety controls shape the experience too. Argon is designed to refuse requests that could help with cyberattacks or dangerous biological materials, and Google runs a real-time monitor that reads the agent's reasoning and actions to block unintended shortcuts; on coding tests, shortcut attempts fell to 2%. Finally, the 1M-token limit is an API capability that needs the right setup, so it does not mean every chat reply will be that long.

Gemini 4 Argon is not in Chat Smith yet, but several Gemini models already are, with no Google AI subscription needed. Pick one from the model list and start chatting on Android, iOS, or in the browser.

For everyday work, Gemini 3.5 Flash is the fastest way in: quick answers, writing, research, and summaries. For tasks that need more careful reasoning, such as long documents or complex analysis, Gemini 2.5 Pro gives you a deeper option. For images, Google's Nano Banana Pro handles image generation and editing in the same app.

Because Chat Smith also includes models from OpenAI, Anthropic, xAI, and DeepSeek, you can send the same prompt to a Gemini model and its rivals and keep the answer that holds up. For research-heavy questions, the built-in Deep Research feature runs advanced search and reasoning across sources, and Chat Smith Pro adds higher usage limits and more premium models for $9.99 per month or $39.99 per year.

If you already pay for Google AI Pro or Ultra, or build on the paid Gemini API, Gemini 4 Argon is worth trying now. It brings state-of-the-art results in long-horizon coding, business automation, long-video understanding, and defensive security, a 1M-token limit, and an aggressive introductory price through November 30. Its main drawbacks are the paywall and how much of the evidence still comes from Google itself.

If you don't need Argon's depth, Google's current Gemini models cover most everyday work. You can try them now in Chat Smith, an AI chatbot that puts Gemini next to the latest models from OpenAI, Anthropic, xAI, and DeepSeek.