Gemini 4 Argon Released: What's New, How to Use It, Pricing, and How It Compares with GPT and Claude
Google released Gemini 4 Argon on October 13, 2026. One million output tokens, a benchmark table against GPT-6 Astra and Claude Opus 5.5, robotics use, how to access it, and API pricing, based on Google's official announcements.
Google released its next-generation model, Gemini 4 Argon, worldwide on October 13, 2026. The general launch followed a September 30 release to security specialists. On Hacker News it became the day's hottest story within hours, with more than 860 points and over 680 comments. This story sticks to facts checked directly against Google's official blog announcement and the Google DeepMind model page.
Gemini 4 Argon at a glance
| Item | Details |
|---|---|
| General availability | October 13, 2026 (worldwide) |
| First release | September 30, 2026, to cyber security partners (Fairwind program) |
| Output token limit | 1 million tokens (previously 64,000) |
| Consumers | Google AI Pro and Ultra subscribers, in the Gemini app and AI Mode in Search |
| Developers | Gemini API paid tier, model ID gemini-4-argon |
| API launch price | $2 input / $10 output per million tokens (until Nov 30) |
| Strengths | Long coding sessions, legal and financial knowledge work, cyber defence, scientific research |
What changed: five key differences
1. One million output tokens: it finishes in one go
The headline change is the output limit. Google says it raised Argon's output token limit from 64,000 to one million tokens. That means long mathematical proofs, entire multi-file apps or large legacy code migrations can be produced in one pass without being cut off. The API also introduces a combined one-million-token limit for input and output, plus a new "continuation" mechanism.
2. Long-running engineering work
Google shared specific examples of using Argon internally.
- Chrome refactoring: it helped split a roughly 6-million-line legacy folder out of the 40-million-line Chromium codebase, finishing "months of work in weeks".
- Python codebase optimisation: working autonomously overnight on the open-source ADK (Agent Development Kit), it cut long-session setup time by 40% and cold import time by 21%.
- Gemma inference speed: on an A100 GPU, it raised Gemma 4 E2B throughput from about 37,900 to more than 108,000 tokens per second, roughly 2.85 times.
- Inside Google, agent code submissions in Antigravity rose by more than 35% and rollbacks fell by about a third.
3. Specialist work such as law and finance
Google says Argon ranks first on the Vals Index, which measures performance on economically valuable tasks. Among early-access companies, Box cited "14% higher accuracy than Gemini 3.1 Pro", and financial analysis start-up Kepler reported "more than 99% agreement with analyst-built baselines on financial models from SEC filings". These are each company's own evaluations.
4. A different answering style
According to the Gemini app announcement, Argon "hedges less, answers more directly, and cuts unnecessary repetition and clichés". On long tasks it shares progress more often. Hacker News reactions were mixed: some said the difference was big for writing and general questions, while others noted that, going by the official figures, coding is a relative weak spot.
5. Cyber defence and safeguards
Argon tied for first on CWE-bench, a vulnerability-fixing benchmark, at 68%, and Google calls it "our most prompt-injection-resistant model yet". According to the announcement, it lowered the success rate of 15 adaptive attacks on the Gray Swan indirect prompt injection benchmark to 0.7%. Security professionals can apply to the Fairwind program to use a version without security guardrails.
Competitor comparison (Google's official benchmarks)
The figures below are selected rows from the table on the Google DeepMind model page. Keep in mind that Google measured and published them itself.
| Area | Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Knowledge work | Vals Index | 68.9% | 63.1% | 65.8% | 67.0% |
| Finance | Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% |
| Legal | Harvey Legal Agent | 19.6% | 5.4% | 6.7% | 3.8% |
| Coding | DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% |
| Coding | FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% |
| Coding | Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% |
| Science | Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% |
| Long context | GraphWalks 256k~1M | 84.2% | 71.8% | 65.0% | 66.8% |
| Computer use | OSWorld-2.0 | 69.2% | 72.6% | Not published | Not published |
| Security | CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% |
In short, Argon clearly leads in knowledge work, finance, law and long context, while GPT-6 Astra or Claude Opus 5.5 score higher on some coding tests (FrontierSWE, Terminal-bench) and science terminal tasks. If coding is your main use, compare models task by task rather than sticking to one.
The robotics version: training robots from video alone
On the same day, Google published robotics examples separately on its developer blog. It says Argon scored 87% on the ERQA benchmark, which measures 3D spatial problem-solving, ahead of every other frontier model.
The standout demo is a "Video-to-Sim-to-Real" pipeline. Argon watched a video of a person playing cornhole, calculated spatial coordinates and full-body joint movements, and wrote the simulation code itself. Transferred to Apptronik's Apollo humanoid, it succeeded seven times in a row in real-world trials. Google says this cut simulation modelling from weeks to hours.
How to use Gemini 4
Consumers: Google AI Pro or higher
Google says Argon is available to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. Prices in South Korea are below (sources are listed in our ChatGPT, Gemini and Claude pricing comparison).
| Plan | Monthly price in Korea | Gemini 4 Argon |
|---|---|---|
| Free | ₩0 | Not available |
| Google AI Plus | ₩7,500 | Not available |
| Google AI Pro | ₩29,000 | Available |
| Google AI Ultra | From ₩119,000 | Available (includes Antigravity) |
Getting started is simple.
- Sign in to the Gemini app (or gemini.google.com) with an AI Pro or higher account.
- Choose Gemini 4 Argon from the model menu in the input box.
- For long jobs, switch from chat to "tasks" mode, and Gemini keeps working while you are away.
Pro and Ultra subscribers can also select Argon from the model menu in AI Mode in Google Search.
Developers: the API and Antigravity
- Gemini API: available now on the paid tier with the model ID
gemini-4-argon. AI Studio, Android Studio and the enterprise Gemini platform are supported. - Google Antigravity: Google's agentic coding tool, where Ultra subscribers can use Argon. Antigravity is now also available inside the Gemini app for macOS.
- Support for researchers: Google is taking applications to give 10,000 scientists the top Antigravity tier free for six months (eligibility and selection criteria apply).
API pricing
| Item | Launch price (until Nov 30, 2026) | Standard price after |
|---|---|---|
| Input (per million tokens) | $2 | $4 |
| Output (per million tokens) | $10 | $20 |
| Cached input | 95% off | Check Google's notice |
The launch price is half the standard rate, so if you want to test large workloads, doing it before the end of November pays off.
Bottom line: who should try it now
- Office workers with lots of documents, research and reports: the knowledge-work benchmark gains and the new answering style are significant, so AI Pro subscribers should try making it their default model.
- Developers: one million output tokens and long autonomous runs are the strengths. Competitors lead on some coding metrics, though, so compare on your own codebase.
- Security and research: check the Fairwind program and the free researcher program.
- Free and AI Plus users: Argon isn't available yet. Google only said it will "release the rest of the Gemini 4 family" and gave no dates.
Frequently asked questions
Can I use Gemini 4 Argon for free?
According to Google, Argon is available in the Gemini app and AI Mode in Search only to Google AI Pro and Ultra subscribers. Developers can use it on the Gemini API paid tier as gemini-4-argon.
How much does the Gemini 4 Argon API cost?
The launch price is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. After November 30, 2026, it rises to $4 for input and $20 for output.
What is the biggest change from the previous version?
The output limit grew from 64,000 to one million tokens, and long coding sessions, complex legal and financial work and cyber defence improved substantially. Answers are also more direct, with less filler.
Is it available outside the US?
Google said Argon launched worldwide on October 13. In South Korea, Google AI Pro (₩29,000 a month) or higher lets you choose the model in the Gemini app.
What does the name Argon mean?
Google introduced Argon only as "the first model in the Gemini 4 family" and did not explain the name.
Sources: Google blog: Introducing Gemini 4 Argon · Google DeepMind Gemini 4 Argon page · Gemini app announcement · Gemini 4 Argon for Robotics · Hacker News discussion
Read next: ChatGPT vs Gemini vs Claude pricing · ATT opt-in rates and the IDFA · What is an MMP? AppsFlyer vs Adjust vs Singular