Gemini 4 Argon: Price, Free Access, Benchmarks (2026)

Gemini 4 Argon is Google’s newest frontier AI model. Google announced it on September 30, 2026. It opened to the public on October 13, 2026. This guide covers what it does, what it costs, who can use it, and how it scores on independent tests.

What is Gemini 4 Argon?

Gemini 4 Argon is Google’s flagship Gemini model. Google built it for complex coding, enterprise knowledge work such as legal and finance tasks, and cybersecurity defense.

Quick facts

• Announced: September 30, 2026
• Public launch: October 13, 2026
• Token window: 1,048,576 tokens (about 1 million), shared across input, thinking, and output
• Output per request: 64K tokens when streaming and 32K when not streaming. Long decode, which needs extra setup, lets one answer run up to about 1 million tokens minus your input.
• Launch price: $2 per million input tokens and $10 per million output tokens, until November 30, 2026
• Standard price after that: $4 input and $20 output per million tokens
• API model ID: gemini-4-argon
• Free? Not for most users. See the access section below.

What makes it different?

Argon is built for long and difficult tasks. Google’s launch posts describe an output limit of 1 million tokens, up from 64,000. The API documentation adds the details. Input, thinking, and output tokens share one window of 1,048,576 tokens. A single request returns up to 64K tokens when streaming. With a feature called long decode, one answer can continue across several requests, up to about 1 million tokens minus your input. Long decode needs extra setup and can run for several minutes.

Google also says Argon answers more directly, hedges less, and avoids clichés. It reports progress more often during long tasks.

How much does it cost?

Google set an introductory price for the API:

• $2 per million input tokens
• $10 per million output tokens
• Cached input tokens cost 95% less than normal input tokens

This price ends on November 30, 2026. After that, the price is $4 per million input tokens and $20 per million output tokens. Google’s API documentation says the standard price starts on December 1, 2026.

Google AI Pro and Ultra subscribers use Argon inside their plan. Google has not published a separate per-message price for the Gemini app.

Is Gemini 4 Argon free?

Not for most people. Based on Google’s launch posts:

• The Gemini API needs a paid tier.
• The Gemini app needs Google AI Pro or Ultra. If you use the free Gemini app and choose Argon, it asks you to upgrade to Pro.
• Google Antigravity and AI Studio access is for Google AI Ultra subscribers.
• Google offers 10,000 scientists the highest Antigravity usage tier free for six months. Selection depends on eligibility, capacity, and approval.
• Cyber defenders can apply to the Fairwind Program.

Google has not announced a free tier for general users. Check Google’s pricing page before you subscribe, because plans change.

How do you use Gemini 4 Argon?

Developers: Use the paid tier of the Gemini API in Google AI Studio or Android Studio. The model ID is gemini-4-argon. Ultra subscribers can also try it in Google Antigravity.

Enterprises: Use Gemini for business, the Gemini API, Google Antigravity, or Google Cloud.

Everyday users: Open the Gemini app or AI Mode in Search with a Google AI Pro or Ultra plan.

Researchers: Apply for the free Antigravity tier.

Cyber defenders: Apply to the Fairwind Program. Approved partners get a version without cyber guardrails.

What are the Gemini 4 Argon specs?

These details come from Google’s Gemini API model page.

What new features does the Gemini app get with Argon?

Google’s Gemini app post lists these changes for Google AI Pro and Ultra subscribers:

• Answers are more direct, with less hedging and fewer repeated points.
• Gemini shares its progress more often during long tasks.
• Google Antigravity’s agentic coding interface is now available inside the Gemini app for macOS.
• Gemini connects to more than 40 apps, including Instacart, OpenTable, Adobe, and Shopify.
• You can hand Gemini tasks that it works on even when you are away from your computer.

Argon also works in AI Mode in Search for Pro and Ultra subscribers.

What can Gemini 4 Argon do?

Google shared real examples from its own teams:

• Chrome: Argon helped refactor a roughly 6-million-line legacy folder inside Chromium’s 40-million-line codebase. Google says this finished months of work in weeks.
• Agent Development Kit: Argon worked overnight on Google’s Python framework. It cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%.
• Data centers: A team of Argon agents freed more than 300 TiB of memory across Google’s data centers once rolled out.
• Video decoding: Argon agents helped build a memory-safe version of libgav1 that runs 2.7x faster than the earlier Rust port, with identical video output.
• Rust migration: Argon agents are moving C and C++ code to Rust, from small libraries up to more than 800,000 lines in the Fuchsia Zircon kernel. Google says these rewrites go through audits and review before production.
• Gemma models: Google reports a 2.85x inference speedup on Gemma 4 E2B using Argon, with no quality drop.

These are Google’s own results. Independent users will keep testing the model over the coming weeks.

How does Gemini 4 Argon perform on benchmarks?

Here are the scores Google reports in its launch posts:

Notice that the CWE bench is a tie for first, not a clear win. Google also reports leading results on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark.

Independent testers have also scored the model. The next section covers those results.

Where does it rank on independent tests?

Vals AI runs the Vals Index. It measures model performance on finance, coding, legal, and tax work. It weights each sector by its share of U.S. GDP.

At launch: Argon ranked first on the Vals Index with 68.9%. Vals AI said this was the first time a Gemini model topped the index.

Now: Vals AI released Vals Index v2.2 on October 19, 2026. It updated several benchmarks. As of October 22, 2026, the top of the index looks like this:

Argon still leads two parts of the index: Legal Research (54.81%) and Harvey’s Legal Agent Benchmark in the Vals version, HLAB Vals 1.1 (27.08%).

One important note from Vals AI: Some tasks for Claude Sonnet 5.5 and Claude Opus 5.5 were served by a fallback model after a provider refusal. If Vals counts those as failures, their scores drop to 69.03% and 64.16%. Vals says Argon had no fallbacks, so it would rank first on that basis.

Argon also ranks second on the Vals Finance Index (64.89%) and second on Finance Agent v2.1 (67.66%).

Science results differ. On the Vals Science Index, GPT-6 Astra ranks first with 65.09%. Argon ranks fifth with 55.36%. This index averages three agentic science benchmarks, so the order differs from the main index.

Artificial Analysis results

Artificial Analysis runs its own Intelligence Index. It scores Argon (high reasoning) at 53. That matches GPT-6 Astra (53) and sits one point above GPT-6.1 Sol (52). On its model page, Artificial Analysis ranks Argon #8 of 214 models. Each model setting counts as a separate entry, so one company can appear several times. Artificial Analysis says Google is back among the top three labs on intelligence.

From Artificial Analysis’s launch analysis:

• Argon ranked first on its AutomationBench test at 78%, 7 points ahead of Claude Sonnet 5.5 (71%).
• On Terminal Bench 4, Argon scored 57%. That is behind Claude Sonnet 5.5 (64%), Claude Opus 5.5 (60%), and GPT-6 Astra (59%).
• Argon had a 15% hallucination rate on AA Omniscience. That is the lowest of any model scoring 45 or higher on its Intelligence Index.
• At the introductory price, it cost $1.99 per Intelligence Index task. Artificial Analysis estimates $3.98 at the standard price.

These launch figures may have moved since. Check the live leaderboard before you quote them.

How does it compare with GPT-6 Astra and Claude?

Here is a side-by-side view using Vals AI’s published figures as of October 22, 2026. Prices are the standard rates Vals lists per million input and output tokens.

How to read this table:

• No model wins every row. Claude Sonnet 5.5 leads the Vals Index. GPT-6 Astra leads the Vals Science Index. Argon has the lowest cost per Vals Index test among these four.
• Argon’s listed price is the standard rate. Its launch price is $2 and $10 until November 30, 2026.
• Vals AI notes that some Claude scores include tasks served by a fallback model after a provider refusal. Its note says Argon would rank first if those tasks counted as failures.
• Benchmarks do not predict every real task. Test the models on your own work.

Has Google finally caught up?

Short answer: Yes, Argon is competitive with the best models. It does not clearly lead every test.

The case that Google has caught up

• Argon topped the Vals Index at launch, and it still ranks in the top three after the October update.
• Artificial Analysis scores it level with GPT-6 Astra and says Google is back among the top three labs.
• It leads the two legal components of the Vals Index.
• Argon completed the Vals Index at a lower cost per test than the two models ahead of it, according to the Vals leaderboard.
• Google’s own teams already use it for large engineering work.

The case for caution

• The #1 ranking did not hold after Vals updated its index. Rankings depend on which benchmarks a test includes.
• Argon ranks fifth on the Vals Science Index, behind GPT-6 Astra and others.
• Argon ranks #8 of 214 models on the Artificial Analysis Intelligence Index, and it trails three rivals on Terminal Bench 4.
• Most other scores in this guide come from Google’s launch posts. Independent testing takes time.
• CWE bench is a tie for first, not a win.

The takeaway: Google has closed much of the gap with OpenAI and Anthropic. Argon sits in the top three on the Vals Index and is level with GPT-6 Astra on the Artificial Analysis index, though it ranks #8 of 214 models there because each model setting counts separately. It is not the clear winner everywhere. The best model depends on your task, so test your own work before you switch tools.

What are the limitations of Gemini 4 Argon?

Argon is strong, but it has trade-offs:

• Speed: Artificial Analysis measures about 82 output tokens per second on Google’s API. It also measures a time to first token of about 45 seconds on high reasoning.
• Token use: Artificial Analysis says Argon used about 62K output tokens per task in its launch test, compared with about 27K for GPT-6 Astra. Lower prices offset this at launch, but heavy token use raises your bill.
• Long agentic tasks cost more: Vals AI reports that Argon’s cost per test rises on long agentic work. At launch, it listed $193.78 per test on the CUA bench.
• Weaker spots: On the Vals AI model page, Argon scores 4.83% on the CUA bench (8th of 9 models) and ranks 17th of 109 on MedScribe.
• Long decode needs setup: Changing the model ID alone does not unlock the full 1 million token output. Long decode needs the SDK option or your own continuation loop, and it can run for several minutes.
• Plan limits: The Gemini app needs a paid plan, and the API needs a paid tier.

Is Gemini 4 Argon safe?

Google says Argon has its most complete safety and security stack so far. Key points from Google:

• External red teams could unlock less than 1% of the biology subtasks Google tests.
• Violations on an adversarial stress test dropped nearly twofold compared to earlier Gemini models.
• Shortcut attempts on coding benchmarks fell to 2%.
• A monitor reads the model’s reasoning and actions and stops unsafe workarounds in real time.

Google released a version without cyber guardrails only to trusted defenders and its own teams. Public users get the version with guardrails.

Who should use Gemini 4 Argon?

Good fit:

• Developers working on large codebases or migrations
• Legal and finance teams that work with long documents
• Enterprises building agents and automations

Maybe wait:

• Casual users who only need quick answers
• Anyone who needs a free option
• Teams that need more independent test results first

Frequently asked questions

When was Gemini Argon released?

 Google announced it on September 30, 2026. Public access began on October 13, 2026.

How much does it cost?

 The launch price is $2 per million input tokens and $10 per million output tokens until November 30, 2026. After that, it is $4 and $20.

Is it free?

 Google has not announced a free tier for general users. Some scientists can apply for free Antigravity access.

What is the Gemini 4 Argon output limit?

Input, thinking, and output tokens share a window of 1,048,576 tokens. One request returns up to 64K tokens when streaming. With long decode, one answer can run up to about 1 million tokens minus your input.

Is it free in the Gemini app?

 No. The free Gemini app asks you to upgrade to Google AI Pro to use Argon.

Is Argon better than GPT-6 Astra?

 It depends on the test. On the Vals Index, Argon scores higher than GPT-6 Astra. On the Vals Science Index, GPT-6 Astra scores higher.

Where can I use it? 

Use it in the Gemini app, AI Mode in Search, AI Studio, Android Studio, Google Antigravity, Gemini for business, and Google Cloud, depending on your plan.

What is the Fairwind Program?

It is Google’s program that gives trusted cyber defenders access to Argon without cyber guardrails.