Gemini 4 Argon: Benchmarks, Pricing and When You Can Use It

Gemini 4 Argon is Google's first Gemini 4 model, and you can now use it. Google announced it on 30 September 2026 for a small group of cyber defenders, and it's since become generally available. Artificial Analysis scores it 50.4 on its Intelligence Index v5.0.1, at high effort, the only setting it has tested.

It's also already replaced. Google released Gemini 4 Carbon on 28 October at the same $2/$10 introductory price, and Carbon scores 60.0, second only to Claude Fable 5.2 (64.6). For new work, start with Carbon. Argon still has one thing Carbon doesn't: it can write up to 1 million tokens in a single response.

Update (November 12, 2026): Argon is no longer limited to trusted testers. Artificial Analysis also replaced its Intelligence Index v4.3.2 with v5.0.1, so the 52.6 we reported in September can't be compared with today's 50.4, or with any other v5.0.1 score. Every independent number below is new.

Product facts come from Google's announcement. Artificial Analysis data was pulled on 12 November 2026.

Gemini 4 Argon specs and prices

At launch Google described $2/$10 as an introductory price that would rise to $4/$20, without saying when. Carbon launched on the same terms. Budget for the higher price on either model.

Can you use Gemini 4 Argon now?

Yes. At launch, only security teams in Google's Fairwind Program could use Argon, in a version with the cyber guardrails removed so they could find and patch vulnerabilities. Paid API customers and Google AI Ultra subscribers were next in line. Argon is now generally available, and Artificial Analysis measures its output speed at about 90 tokens per second.

The more useful question is whether to use it. Four weeks after Argon's announcement, Google shipped Gemini 4 Carbon at the same price. Carbon scores higher, costs less per task and offers faster effort settings. Argon's remaining edge is output length.

Google's benchmarks: first on knowledge work, mixed on coding

Google compared Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on 18 benchmarks. Argon wins or ties 13 of them. The full table is in Google's announcement, and The New Stack has it in text form.

The wins that matter are the big ones. On AutomationBench, Zapier's test of business workflows, Argon leads Opus 5.5 by almost 9 points. On GraphWalks with inputs between 256K and 1M tokens, it leads Astra by 12. On Harvey's legal agent test it scores nearly three times Fable 5.1, though 19.6% still means it fully completes about one legal task in five. Several other wins, like Vals Index, Chartography and Agent's Last Exam, are within two points.

Coding is where the table turns. Argon has the top DeepSWE score, but it comes last of the four on both FrontierSWE v2 and Terminal-Bench 4.0, 9 points behind Opus 5.5 on the second. If you mostly run coding agents in a terminal, Google's own numbers don't make Argon the obvious pick.

These are vendor numbers, chosen and run by Google. Use them to see where Argon is strong, and use the independent scores below to compare it with other labs.

Artificial Analysis: 50.4, behind Carbon and the top Claude and GPT models

On the Artificial Analysis Intelligence Index v5.0.1, Argon scores 50.4 at high effort. That puts it behind Gemini 4 Carbon (60.0), Claude Opus 5.5 (59.1), GPT-6 Astra (56.9), GPT-6.1 Sol (54.6) and Claude Sonnet 5.5 (54.5), all at their best settings.

High is the only setting Artificial Analysis has tested for Argon, while most rivals' headline scores come from max effort. Comparing everyone at high is closer, with the caveat that each lab's "high" spends a different amount of thinking.

At high effort Argon is fifth, level with GPT-6.1 Sol, and it costs more per task than any model in the table. Carbon scores 9.6 points more at high for less per task. On our LLM leaderboard Argon appears as replaced by Carbon, so our best-model picks no longer choose it.

In September we also showed AA's professional capability indexes for Argon (legal, finance and others). Those were measured on the old index version, so we've removed them.

What Gemini 4 Argon costs per task

At $2/$10, Artificial Analysis spent $9.95 per Intelligence Index task on Argon, thinking included. That's more than Fable 5.2 at high ($9.08), nearly three times Opus 5.5 at high ($3.46) and about 17 times GPT-6.1 Sol at high ($0.59).

If the $4/$20 price applies, the same work would cost about twice as much, roughly $20 per task.

Cached input is the one price advantage. At 95% off, Argon's cache reads cost $0.10 per million tokens at $2, half of Opus 5.5's $0.20. For apps that resend the same long documents or codebase on every call, that gap adds up. The cost calculator runs the numbers for your volume once you know your cache hit rate.

Gemini 4 Argon can write 1 million tokens in one response

Argon can write up to 1 million tokens in one response. Earlier Gemini models stopped at 64K, Claude and GPT-6 models stop at 128K on their standard APIs, and Gemini 4 Carbon lists 65,536.

Google's case for this is long agent runs where the model keeps thinking and writing in one pass instead of stopping and resuming. Its examples are internal: agents porting C and C++ code to Rust across Google, up to the 800,000-line Zircon kernel, and a Rust rewrite of the libgav1 video decoder that Google says runs 2.7 times faster than the earlier port.

A full 1M-token response costs $10 at $2/$10 and $20 at $4/$20. That makes sense for a codebase migration or a long report. It's also a lot of tokens to pay for if a run goes the wrong way, so cap max_tokens until you trust how the model behaves on your work.

Built for cyber defense first

Google trained Argon to find, validate and patch software vulnerabilities, and that's why its first users were security teams. On CWE-bench v1, which tests fixing vulnerabilities, Argon ties GPT-6 Astra at 68%, with Opus 5.5 at 67%. Wiz, which Google owns, says Argon found a serious flaw in healthcare software used by hospitals that earlier models had missed.

Google's other two cyber results, 85.8% on its internal vulnerability benchmark and 70.9% on Wiz's penetration-testing benchmark (reported by The New Stack), are only compared with Google's own Gemini 3.8 Flash Cyber. There's no way to set them against Claude or GPT-6 yet.

Gemini 4 Argon vs Gemini 4 Carbon, Claude Opus 5.5, Sonnet 5.5 and GPT-6.1 Sol

Gemini 4 Argon vs Gemini 4 Carbon

Carbon is Argon's replacement, at the same introductory $2/$10. It scores 9.6 points more at high effort and costs less per task there ($7.76 against $9.95). It also gives you faster settings that Artificial Analysis hasn't tested on Argon: at medium Carbon scores 58.3 in about 33 seconds, and at low 54.2 in about 9 seconds. Carbon has a Coding Agent Index score of 59. Argon has none.

Argon's only remaining edge is output length. If you need more than about 65,000 tokens in one response, Argon can do it and Carbon can't. Otherwise, move to Carbon.

Gemini 4 Argon vs Claude Opus 5.5

Opus 5.5 scores 59.1 at max and 51.8 at high, above Argon's 50.4, and costs $3.46 per task at high against Argon's $9.95. It also leads the Coding Agent Index at 62. On Google's own launch table, Opus 5.5 won Terminal-Bench 4.0 and PostTrainBench, and Argon won most of the knowledge-work and long-context tests.

For coding agents, stay on Opus 5.5. For long-document work, compare Opus 5.5 with Carbon rather than Argon. Our Claude Opus 5.5 guide has its cost at every effort level.

Gemini 4 Argon vs Claude Sonnet 5.5

Anthropic released Sonnet 5.5 on 28 September at $2/$10, the same as Argon's price. It scores 54.5 at max effort and 59 on the Coding Agent Index. At extra high it scores 49.9 in about 47 seconds for $4.09 per task, within a point of Argon's high score for less than half the cost. At high it answers in about 19 seconds for $1.64, with a lower score of 41.5.

Gemini 4 Argon vs GPT-6.1 Sol

GPT-6.1 Sol replaced GPT-6 Sol on 29 September at the same $2/$10. It scores 54.6 at max and 50.3 at high, level with Argon, for $0.59 per task instead of $9.95. At medium effort it answers in about 18 seconds and scores 46.3 for $0.36. For production work where cost per call matters, Sol is the better buy.

What to use instead of Argon

If you already run Argon, test Carbon next to it on a sample of real tasks. Keep Argon only for jobs that need one very long response, and price both at $4/$20 so the end of the intro period doesn't surprise you. Artificial Analysis still hasn't published a Coding Agent score for Argon.

FAQ

Is Gemini 4 Argon available?

Yes. Gemini 4 Argon is now generally available. At launch on 30 September 2026, Google limited it to security teams in its Fairwind Program. Google has since released Gemini 4 Carbon (28 October), which replaces Argon at the same $2/$10 introductory price and scores higher on Artificial Analysis.

How much does Gemini 4 Argon cost?

Artificial Analysis lists Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens. Google announced that as an introductory price that rises to $4/$20, with cached input 95% off. AA spent $9.95 per benchmark task on Argon at $2/$10, more than Gemini 4 Carbon's $7.76 at the same price.

Gemini 4 Argon or Gemini 4 Carbon?

Carbon, unless you need very long single responses. Carbon scores 60.0 on the Artificial Analysis Intelligence Index v5.0.1 against Argon's 50.4, costs less per task at high effort, and has faster settings. Argon can write up to 1M tokens in one response; Carbon lists 65,536.

Is Gemini 4 Argon better than Claude Opus 5.5?

No. On the Artificial Analysis Intelligence Index v5.0.1, Opus 5.5 scores 59.1 at max effort and 51.8 at high, against Argon's 50.4 at high, and costs $3.46 per task at high against Argon's $9.95. Argon led on Google's own knowledge-work and long-context benchmarks at launch. Opus 5.5 leads the Coding Agent Index at 62.

What is the Gemini 4 Argon context window and output limit?

Gemini 4 Argon takes a 1M-token context window and can write up to 1 million output tokens in one response, up from 64K on earlier Gemini models. Claude and GPT-6 models are limited to 128K output tokens on their standard APIs, and Gemini 4 Carbon lists 65,536. Argon accepts text and images and returns text.

What is the Google Fairwind Program?

Fairwind is Google's early-access program for trusted cyber defenders. Its members got Gemini 4 Argon before anyone else, in a version without cyber guardrails, so they could use it to find and fix vulnerabilities. Wiz uses it for its Scan for Good project, which looks for exposures in critical public infrastructure for free.

Is Gemini 4 Argon the same as Gemini 3.5 Pro?

Google hasn't released a Gemini 3.5 Pro. It talked about a new Pro model at I/O in May 2026 but shipped a run of Flash models instead, and The New Stack described Argon as that model's replacement. Gemini 4 Carbon and Argon are now Google's top models, scoring 60.0 and 50.4 on Artificial Analysis.

Intelligence Index v5.0.1 scores, costs per task, response times and speeds from Artificial Analysis, pulled 12 November 2026. Argon's release, launch access, pricing, output limit and benchmark results from Google's announcement; benchmark table and cyber results as reported by The New Stack. Rival prices and limits from Anthropic's model overview, OpenAI's GPT-6.1 Sol page and Artificial Analysis. Published September 30, 2026. Updated November 12, 2026.