Gemini 4 Argon: Who Can Use It Now, Pricing, and Benchmarks

Google announced Gemini 4 Argon on September 30 and spent two weeks testing it with cyber defenders, early-access customers, and its own teams. On October 13, it opened Argon to anyone on an eligible paid plan.

Argon is now in the Gemini API for paid developers, in the Gemini app and AI Mode for Google AI Pro and Ultra subscribers, and in Antigravity for Ultra. Free accounts can't use it yet, and the launch API price of $2 per million input tokens and $10 per million output tokens doubles on December 1.

I went through both of Google's launch posts, the Gemini API docs and model card, the Gemini app release notes, and Artificial Analysis's independent scores, all on October 14.

I wanted to know which plan gets you in, what you'll pay once the launch price ends, and how much of the benchmark hype holds up when someone outside Google runs the tests.

TL;DR:

• What it is: Google's newest frontier model and the first Gemini 4 model, built for long coding jobs, enterprise knowledge work like legal and finance, and cyber defense.
• Released: announced September 30, 2026, and generally available since October 13, 2026.
• Who can use it: paid Gemini API developers, Google AI Pro and Ultra subscribers in the Gemini app and AI Mode, Ultra subscribers in Google Antigravity, Gemini for business customers, and four GitHub Copilot plans. Free accounts and the free API tier are left out.
• Price: $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1.
• How good is it: it leads most rows in Google's own benchmark table, ties GPT-6 Astra at 53 on Artificial Analysis's independent index, and trails Claude Opus 5.5 there (58).

What is Gemini 4 Argon?

Gemini 4 Argon is Google DeepMind's newest frontier AI model, announced on September 30, 2026, released to paying users on October 13, and designed for long, multi-step work.

Google's September 30 announcement says it "delivers frontier performance" in software engineering, enterprise knowledge work, and cybersecurity defense.

Argon's headline spec is a 1-million-token window that input, thinking, and output all share. A single response can run up to 1 million tokens minus whatever you sent in, up from the 64K-token output limit on Google's earlier models.

The bigger limit helps on jobs where the model has to think for a long time or produce a lot of code in one pass, like migrating a whole library to a new language. (Artificial Analysis puts 1 million tokens at about 1,500 pages, which is a lot of Rust.)

In the API, responses stop at 65,536 tokens by default, and Google's developer guide says to run anything longer as a background job.

The specs, from Google's API model page and the Argon model card:

Is Gemini 4 Argon the same as Gemini 4?

Yes, for now. Argon is the first and only Gemini 4 model so far, so "Gemini 4" and "Gemini 4 Argon" point to the same thing.

That probably won't last. Google's Logan Kilpatrick called the launch "just the start of the Gemini 4 era" in his announcement on X, so expect more Gemini 4 models.

In Google's current lineup, Argon sits above Gemini 3.1 Pro (still in preview) and Gemini 3.8 Flash.

Who can use Gemini 4 Argon right now?

As of October 14, 2026, Gemini 4 Argon is open to paid Gemini API developers, Google AI Pro and Ultra subscribers, Gemini for business customers, and GitHub Copilot Pro+, Max, Business, and Enterprise users. Free Gemini accounts, Google AI Plus, and the free API tier don't have it.

Here's where it shows up and who gets in:

1. Google AI Pro and Ultra subscribers

Argon is "rolling out globally for Google AI Pro & Ultra subscribers in the Gemini app," according to Google's Gemini release notes, and both tiers also get it in AI Mode in Search.

Google AI Pro costs $19.99 a month, and Ultra costs $99.99 (5x Pro's usage) or $199.99 (20x), per Google's plans page. If you only want Argon in the app, Pro is the cheapest way in.

Google hasn't published Argon-specific limits. Its usage limits page says Gemini app limits are compute-based and refresh every five hours until you hit a weekly cap.

Free accounts and Google AI Plus ($4.99 a month) don't get Argon.

2. Developers on the paid Gemini API

Argon is generally available in the Gemini API as gemini-4-argon, according to Google's API changelog. Google's October 13 post says you can build with it on the paid tier in Google AI Studio and Android Studio.

It isn't on the free API tier or included with Google AI Pro, so Pro subscribers need billing turned on. Google AI Ultra subscribers can also try it in AI Studio, per Google's AI plans page.

3. Ultra subscribers and businesses in Google Antigravity

Google AI Ultra subscribers can use Argon in Google Antigravity, Google's agentic coding platform, and so can eligible Gemini Enterprise customers.

Google says Argon in Antigravity has helped its own teams raise agentic code submissions by more than 35% while cutting code rollbacks by nearly a third.

Google is also offering 10,000 scientists Antigravity's highest usage tier free for six months, with applications open through its October 13 post.

4. GitHub Copilot users and Gemini for business customers

Argon is generally available in GitHub Copilot for Pro+, Max, Business, and Enterprise users, according to GitHub's changelog. Pick it from the model picker in VS Code, JetBrains IDEs, Copilot CLI, and other supported editors.

GitHub bills it at Google's list price under usage-based billing, and Business and Enterprise admins can switch it off in Copilot's model policy.

Separately, companies on Gemini for business get Argon too, according to Google's October 13 post.

5. Cyber defenders in the Fairwind Program

Cyber defenders can apply to Google DeepMind's Fairwind Program for a version of Argon without the usual cyber guardrails. Google says the program is expanding to hundreds more partners.

Defenders got a two-week head start because Argon is good at finding and patching security flaws, so the people protecting hospitals and power grids could patch their systems before attackers got the same capability.

How to use Gemini 4 Argon

Here's how to start, depending on where you work:

1. In the Gemini app. Sign in with a Google AI Pro or Ultra account, open the model picker, and choose Argon. The rollout is gradual, so update the app and check back if it isn't there yet.
2. In the Gemini API. Create an API key in Google AI Studio with billing turned on, then set your model to gemini-4-argon.
3. In Google Antigravity. Sign in with Google AI Ultra in the desktop app or CLI and pick Argon.
4. In GitHub Copilot. On an eligible plan, pick Gemini 4 Argon from the model picker.

If you're moving API code from an earlier Gemini model, Google's developer guide lists a few changes:

• Pick a thinking level. Argon offers low, medium, and high, with medium as the default.
• Drop temperature, top_p, and top_k. Argon ignores them.
• Remove explicit caching calls. Argon only supports automatic (implicit) caching.
• Raise small output limits. If your max output is under 32,768 tokens, raise it, since Argon can spend more of it on thinking.
• Plan for long runs. Google recommends HTTP timeouts of at least 15 minutes and background execution for outputs over 64K tokens.

Gemini 4 release date: when it came out

Gemini 4 Argon was announced on September 30, 2026, and became generally available on October 13, 2026. Here's the timeline so far:

• September 30: Google announces Argon and gives it to a set of Fairwind cyber defenders, trusted testers, and its own teams.
• October 13: Argon goes live in the Gemini API, for Google AI Pro and Ultra subscribers in the Gemini app and AI Mode, for Ultra subscribers in Antigravity, for Gemini for business customers, and in GitHub Copilot.
• December 1: The introductory API price ends, and rates double to $4 and $20 per million tokens.

Google hasn't given a date for free Gemini users.

Gemini 4 Argon pricing

Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, every token rate doubles, to $4 and $20, according to Google's API pricing page.

Output pricing includes thinking tokens, and cached input gets 95% off at both prices.

A few things to know before you budget:

1. The discount ends November 30. Price anything that will run past November at $4 and $20. Google did the same with Gemini 3.8 Flash, whose price doubles on January 1, 2027.
2. Argon is a heavy thinker. Artificial Analysis says Argon generated 110 million output tokens to run its index, against a median of 85 million, and those thinking tokens bill as output.
3. Caching is automatic. Prompts of at least 4,096 tokens get cached on their own, and repeat hits cost a twentieth of the normal input price. You can't create your own caches.
4. Batch halves the bill. If a job can wait, the Batch API charges half the standard rate.

What that looks like on a real bill

Say your app sends 1,000 requests, each with 5,000 input tokens and 2,000 output tokens. Through November, that's $10 for input plus $20 for output, so $30 total. From December 1, the same job costs $60.

Now say 4,500 of those 5,000 input tokens are the same system prompt every time, and they hit Google's cache. Your input cost drops from $10 to about $1.45, and the job comes in at about $21.50 (about $43 from December).

How Argon's per-token price compares to its rivals, from each vendor's own pricing page:

That table assumes short prompts. Argon and Opus 5.5 charge the same rate all the way to 1 million tokens, while GPT-6.1 Sol and GPT-6 Astra charge 2x for input and 1.5x for output once a prompt passes 272K tokens, according to OpenAI's model docs.

Through November, that makes Argon cheaper than Sol on long-document jobs.

So Argon launches at the same price as GPT-6.1 Sol, and from December it matches Claude Opus 5.5. Either way, it costs less than half what GPT-6 Astra does per token.

A model that spends more tokens per task can end up costing more than a cheaper-looking rival, and Artificial Analysis puts Argon at $1.99 per task on its index at the launch price. Test your own workload before you trust any table (this one included).

What Gemini 4 Argon is built for

Google built Gemini 4 Argon for three kinds of work: long coding projects, enterprise knowledge work, and cyber defense. Google backs the coding claims with its own projects, the knowledge-work claims with benchmark wins, and the cyber claims with an early find from Wiz.

Long-running coding work

Argon's headline example is large codebase migrations. Google says Argon agents are moving C and C++ code to Rust across the company, from tens of thousands of lines in libraries like re2 and libgav1 up to more than 800,000 lines for the Fuchsia Zircon kernel.

The most concrete example is libgav1, Google's open-source video decoder, which already had a Rust port.

Argon agents replaced 32,000 lines of SIMD code with safe Rust the compiler could optimize on its own. The result is a memory-safe decoder that runs 2.7 times faster than the earlier Rust port, with identical video output.

Google is careful here: these rewrites go through automated and manual audits before they reach production.

Google also says Argon agents found memory optimizations across its data centers that free up more than 300 TiB once rolled out (with an estimated 500 TiB to 1 PiB in total savings), and that Argon helped quantum researchers beat a published baseline by 40% in minutes.

If you want to compare it against the coding tools you already use, our roundup of AI coding agents covers the options worth testing.

Enterprise knowledge work

Argon tops the Vals Index, a benchmark that scores finance, coding, legal, and tax work and weights each field by its share of U.S. GDP. It also leads Vals Finance Agent v2 for financial research and Harvey's Legal Agent Benchmark for legal drafting.

Harvey, the legal AI company, says it saw "substantial improvements over Gemini 3.1 Pro" across its legal agent evaluations, according to Google's Cloud launch post.

On AutomationBench, Zapier's benchmark for running business tasks end to end, Google says Argon ranks first at 51.3%. It's also strong when the work involves visuals, with a 91.7% score on LVBench, a test of understanding long videos.

Cyber defense

Argon can find, confirm, and patch serious software vulnerabilities on its own.

Wiz, the cloud security company, used Argon through its Scan for Good program, and Google says the model found a critical flaw in healthcare software used by hospitals worldwide that earlier frontier models had missed.

On CWE-bench v1, which tests whether a model can fix security flaws, Argon ties for first place at 68% with GPT-6 Astra and Grok 4.7, according to Google DeepMind's charts.

Science and robotics

Google's October 13 post adds science and robotics to the list. It says Argon scores 87% on ERQA, an embodied reasoning benchmark for robots, ahead of every other frontier model.

In an internal Google test, Argon watched a video of a person playing cornhole and wrote simulation code to copy it, and Apptronik's Apollo humanoid then sank seven shots in a row in a real-world trial.

In private evals from Gamow Labs, built from real clinical and lab data, Argon scored 57% on RareBench (rare-disease diagnosis help) and 52.2% on Gamow's LabBench (wet-lab reasoning), a different test from the LABBench 2 in Google's main table.

Gemini 4 Argon benchmarks

Gemini 4 Argon leads or ties for the lead on 14 of the 19 results in Google's benchmark table, against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5.

These are the numbers Google chose to publish, with Argon at its highest thinking setting according to Google's methodology notes, so read them as the vendor's best case.

The full table on the Google DeepMind Gemini page adds Vibe Code Bench, RiemannBench, a shorter GraphWalks run, Agent's Last Exam, and Chartography, and Argon leads all five.

Here's how I'd read it:

• Legal and finance are runaway wins. Argon scores 19.6% on Harvey's legal test, almost triple the next-best score (Claude Fable 5.1's 6.7%), and it leads Vals Finance Agent v2 by 6.5 points over the next-best model.
• Coding is split. Argon sets the top DeepSWE v1.1 score at 77.9%, but GPT-6 Astra beats it on FrontierSWE v2 by more than 10 points (65.5% vs 55.0%).
• Long agent runs still favor Opus 5.5. Claude Opus 5.5 leads Terminal-bench 4.0 at 66.4% and PostTrainBench at 49.3%, two hands-on tests where the model works in a terminal for a long stretch, and Argon finishes last of the four on Terminal-bench 4.0.
• Very long documents are a real strength. On GraphWalks at 256K to 1M tokens, Argon scores 84.2%, while no rival in the table clears 72%.
• Science is mixed. Argon wins LABBench 2, while Astra beats it on Terminal-Bench Science by more than 10 points.

‍

{{templates}}

‍

What independent testing says

Artificial Analysis scores Gemini 4 Argon at 53 on its Intelligence Index, which ties it with GPT-6 Astra and Claude Fable 5.1 and ranks it eighth of 216 models on its Argon model page. Google's own table makes Argon look further ahead than that.

Claude Opus 5.5 (at max effort) sits on top at 58, with Claude Sonnet 5.5 and Meta's new Muse Wonder Full at 56, according to the Artificial Analysis leaderboard on October 14. GPT-6.1 Sol is one point behind Argon at 52.

If you're choosing between the two Claude models at the top of that list, our Sonnet 5.5 vs Opus 5.5 comparison breaks down the cost per task.

Why Argon wins Google's chart but only ties on Artificial Analysis

The gap comes down to two differences in how Google and Artificial Analysis tested:

1. Different test lists. Google picked 18 benchmarks (19 results), led by four knowledge-work tests where Argon is strongest. Artificial Analysis averages 10 of its own picks, including Terminal-bench 4.0, where Opus 5.5 leads Google's own table.
2. Different scoring. Artificial Analysis runs every model through the same setup, while Google's table mixes its own runs, outside leaderboards like Vals AI, and rivals' self-reported scores.

Both also tested Argon at high, its top thinking level, so the settings match.

So both can be right. Argon looks like a top-tier model with standout legal, finance, and long-document results, and it's roughly even with Astra on general intelligence.

We compare the two head-to-head in Gemini 4 Argon vs GPT-6 Astra, and our GPT-6 explainer covers Astra on its own.

What one early tester says

Developer and YouTuber Theo Browne, who had early access, had a mixed first take. He called Argon a huge jump for Gemini and good value for the price, and said it's the first model to ace his private SkateBench test.

In the same post, he said it still does strange things often enough that he doesn't trust it yet. That matches the benchmark picture: a big step up for Google, with real gaps on long agent runs.

Why Google rolled Argon out in stages

Google says capabilities at Argon's level need "a phased approach," and its cyber skills are the clearest reason why, since a model that can find and patch serious security flaws can also help someone exploit them.

Google is also taking part in the U.S. government's voluntary process for early model access, and CNBC reported that Sundar Pichai signed a voluntary AI safety agreement at the White House the day before Google announced Argon on September 30.

Before the wide release, Google says it strengthened four safeguards:

1. Misuse defenses. Argon is designed to refuse requests that could help with cyberattacks or chemical, biological, radiological, and nuclear weapons, and Google monitors the model's internal activity to spot misuse.
2. Prompt injection resistance. Google calls Argon its most resilient model yet against hidden instructions planted in web pages or files. On Gray Swan's prompt-injection test, attacks succeeded only 0.7% of the time after 15 attempts.
3. Misalignment monitoring. Google watches Argon's reasoning and actions and stops a task if the model tries to go beyond what the user asked for.
4. Locked-down test environments. Google isolates and seals its sandboxes before high-risk training or evaluation runs.

If you plan to point Argon at email, documents, or web pages, watch the prompt injection results closely, since those are the places hidden instructions get planted.

‍

{{cta}}

‍

How to tell if Argon is worth switching to

Gemini 4 Argon is live, so you can find out within a day whether it beats what you use now.

1. Write down 10 to 20 real tasks. Pick work Google says Argon is good at: a contract review, a financial research question, a messy refactor, a long PDF.
2. Run them on Argon and on your current model. Score the results against Gemini 3.1 Pro, GPT-6.1 Sol, Opus 5.5, or whatever you use, and note the cost of each run.
3. Try more than one thinking level. Start at medium, Argon's default, and only move the hardest tasks to high, since extra thinking bills as output.
4. Budget at $4 and $20. Price anything past November 30 at the post-intro rate. Our Gemini pricing guide covers the rest of Google's plans and API rates.

If Argon only wins on the legal and long-document jobs, route those to it and leave the rest where they are.

If you'd rather not manage model choice yourself, Lindy, an AI teammate that works in Slack, picks a model for each task and also lets you choose one yourself, with OpenAI, Claude, and Gemini models in the mix.

FAQ

Is Gemini 4 Argon out?

Yes. Gemini 4 Argon became generally available on October 13, 2026, for paid Gemini API developers, Google AI Pro and Ultra subscribers in the Gemini app and AI Mode, Ultra subscribers in Google Antigravity, Gemini for business customers, and GitHub Copilot Pro+, Max, Business, and Enterprise users.

Free Gemini accounts don't have it yet.

When is the Gemini 4 release date?

Gemini 4 Argon was announced on September 30, 2026, and released on October 13, 2026. Google hasn't given a date for free Gemini users, and its introductory API price runs through November 30, 2026.

Can Gemini Pro or Ultra subscribers use Gemini 4 Argon?

Yes. Google AI Pro ($19.99 a month) and Ultra subscribers can use Gemini 4 Argon in the Gemini app and AI Mode in Search. Argon in Antigravity needs Ultra or an eligible Gemini Enterprise subscription, and Google AI Plus subscribers don't get Argon.

Is Gemini 4 Argon free?

No, Gemini 4 Argon requires a paid plan. It isn't on the free Gemini API tier or in the free Gemini app. The cheapest way in is Google AI Pro at $19.99 a month, or pay-as-you-go API billing.

How much does Gemini 4 Argon cost?

Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached input at $0.10. From December 1, the price doubles to $4 and $20, per Google's API pricing page.

Is Gemini 4 Argon better than GPT-6 Astra?

Gemini 4 Argon beats GPT-6 Astra on most of Google's benchmarks, including DeepSWE v1.1, Vals Finance Agent v2, and Harvey's legal test.

They tie at 53 on Artificial Analysis's independent index. In Google's own table, Astra wins FrontierSWE v2, Terminal-bench 4.0, Terminal-Bench Science, and OSWorld-2.0.

Is Gemini 4 Argon better than Claude Opus 5.5?

Gemini 4 Argon beats Claude Opus 5.5 on most rows of Google's table, like DeepSWE v1.1 (77.9% vs 74.2%). Opus 5.5 wins FrontierSWE v2, Terminal-bench 4.0, PostTrainBench, and Terminal-Bench Science, and it scores higher on Artificial Analysis's independent index (58 vs 53).

What is the Fairwind Program?

The Fairwind Program is Google DeepMind's early-access program for cyber defenders, such as governments, critical infrastructure operators, and core technology platforms. Its partners get Gemini 4 Argon with its full cyber defense capabilities and without the usual cyber guardrails.

Members must use strong security controls, and only internal security teams can use the model. Google says it's expanding the program to hundreds more partners.

What does Gemini 4 do?

Gemini 4 Argon handles long, complex work like code migrations, financial research, legal drafting, and finding security flaws. A single response can run up to about 1 million tokens.

Inside Google, it's already running large Rust migrations and has found data center memory optimizations that free up more than 300 TiB once rolled out.