# Gemini 4 Argon

Gemini 4 Argon is an AI model released by Google on Sep 30 2026. Tracked results include 77.9% on DeepSWE 1.1, 55% on FrontierSWE V2 and 88.8% on LAB-Bench 2.

## Facts

| Field | Value |
| --- | --- |
| Model | Gemini 4 Argon |
| Developer | Google |
| Release date | Wednesday, Sep 30 2026 |
| Licensing | Closed |

## API pricing

All rates in USD per 1,000,000 tokens, pay-as-you-go.

| Tier | Input | Cached input | Output |
| --- | --- | --- | --- |
| Introductory | $2.00 | $0.10 | $10.00 |

Introductory pricing. Once it expires, $4 per million input tokens and $20 per million output tokens apply.
Rolling out first to trusted cyber defenders through the Fairwind Program, then to paid API customers and Google AI Ultra subscribers.
Verified October 1, 2026 against the first-party source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/

## Tracked benchmark scores

| Benchmark | Score | Source | What it measures |
| --- | --- | --- | --- |
| Gray Swan IPI (k = 15) | 0.7% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 15 tries and counts an attack as successful if any of them works. Lower is better. |
| DeepSWE 1.1 | 77.9% | Lab | Artificial Analysis' independent test of deep, agentic software-engineering work — the AI has to plan and carry out substantial coding tasks end to end. (Version 1.1 of the test.) Higher is better. |
| FrontierSWE V2 | 55% | Lab | Engineering problems that would occupy a person for days: systems implementation, performance work, scientific computing and AI research, with up to twenty hours per task. Partial credit is awarded, because finishing one outright is still rare. Higher is better. |
| Terminal-Bench 4.0 | 57.4% | Lab | Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Version 4.0 recalibrated how much time, CPU and memory each task gets, removed eight tasks and fixed nineteen, so fewer runs fail for reasons that have nothing to do with the model. Scores are not comparable with earlier versions. Higher is better. |
| Terminal-Bench-Science 0.1 | 57.6% | Lab | The same command-line setup as Terminal-Bench, pointed at scientific work: the AI has to drive research tooling and computational workflows through to a result, rather than administer a machine. Version 0.1 is the first release of the task set, and scores run lower than on the general board. Higher is better. |
| LAB-Bench 2 | 88.8% | Lab | Everyday tasks from a working biology lab — reading protocols, interpreting figures and sequence data, and answering the practical questions a researcher hits at the bench. Higher is better. |
| OSWorld 2.0 | 69.2% | Lab | Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Version 2.0 is a harder, refreshed task set. Higher is better. |
| Agent's Last Exam (pass@1) | 39.5% | Lab | A hard set of desktop and operating-system tasks an AI agent has to finish by looking at the screen and working the machine itself. The score is the share it passes outright — partial credit does not count. Higher is better. |
| AutomationBench | 51.3% | Lab | Tests whether the AI can run real multi-step business workflows — the kind of end-to-end office processes companies want to automate — from start to finish. Higher is better. |
| Finance Agent v2 | 65.4% | Lab | Tests the AI on real financial-analysis work, like digging through reports and making sound decisions. Higher is better. |
| Harvey's Legal Agent Benchmark | 19.6% | Lab | Harvey's test of whether an AI agent can complete real legal work — drafting and reviewing documents, working with spreadsheets and presentations, and navigating files the way a lawyer's assistant would. Higher is better. |
| LVBench | 91.7% | Lab | Can the AI follow a very long video — up to an hour — and answer questions that need details from far apart in it? Higher is better. |

## About Gemini 4 Argon

Gemini 4 Argon, announced September 30, 2026, was Google's first new frontier model since Gemini 3.1 Pro in February, after seven months in which every Gemini upgrade had landed on the Flash tier. Google did not open it to everyone at once: Argon went first to trusted cyber defenders through its Fairwind Program, unrestricted by cyber guardrails, with paid API customers and Google AI Ultra subscribers promised next, once early-tester feedback had shaped the safeguards. Google also put it through the US government's voluntary pre-release access process. The announced introductory price was $2 per million input tokens and $10 per million output, with cached input 95% cheaper, rising to $4 and $20 once the introductory period ended, and the output limit rose to 1M tokens from the previous 64K, so that a single trajectory could think and write for hundreds of thousands of tokens.

Google's launch table set it against GPT-6 Astra and led with long-horizon coding and knowledge work. It scored 77.9% on DeepSWE v1.1, which Google called a new state of the art, 51.3% on AutomationBench, 65.4% on Vals Finance Agent v2, 19.6% on Harvey's Legal Agent Benchmark and 68.9% on the Vals Index, all ahead of Astra in Google's comparison, along with 91.7% on LVBench for long video. Google also reported a 0.7% attack success rate on Gray Swan's indirect prompt injection benchmark at fifteen attempts, the lowest in its comparison, against 1.0% for Claude Opus 5.5 and Claude Fable 5.1. Astra came out ahead on FrontierSWE v2 (65.5% against Argon's 55.0%), Terminal-Bench Science 0.1 and OSWorld 2.0, and the two were close on Terminal-Bench 4.0, where Argon scored 57.4%. Security was the other half of the pitch: Argon tied for first on CWE-bench v1 at 68.0%, and in an early run through Wiz's Scan for Good programme it found a critical flaw exposing personal data in healthcare software used by hospitals worldwide. Internally, Google said, Argon agents had freed more than 300 TiB of memory across its data centres and were porting C and C++ code to Rust, from libraries like re2 up to the 800,000-line Fuchsia Zircon kernel.

## Questions and answers

### When was Gemini 4 Argon released?

Gemini 4 Argon was released by Google on Wednesday, Sep 30 2026.

### Who made Gemini 4 Argon?

Gemini 4 Argon was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products.

### How much does Gemini 4 Argon cost?

Gemini 4 Argon costs $2.00 per million input tokens and $10.00 per million output tokens through the Google API. Cached input is $0.10 per million tokens. Rates are pay-as-you-go API prices verified against Google's published pricing on October 1, 2026.

### What benchmark scores did Gemini 4 Argon get?

Gemini 4 Argon reports 12 tracked benchmark scores — Gray Swan IPI (k = 15): 0.7%; DeepSWE 1.1: 77.9%; FrontierSWE V2: 55%; Terminal-Bench 4.0: 57.4%; Terminal-Bench-Science 0.1: 57.6%; LAB-Bench 2: 88.8%; OSWorld 2.0: 69.2%; Agent's Last Exam (pass@1): 39.5%; AutomationBench: 51.3%; Finance Agent v2: 65.4%; Harvey's Legal Agent Benchmark: 19.6%; LVBench: 91.7%. Tracked scores may come from lab reports or independent benchmarks; source details accompany the benchmark data. It holds the best score among all models tracked here on Gray Swan IPI (k = 15), DeepSWE 1.1, FrontierSWE V2, LAB-Bench 2, Finance Agent v2 and LVBench.

### Is Gemini 4 Argon open source?

No. Gemini 4 Argon is a closed model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms.

### What came before and after Gemini 4 Argon?

Google's previous tracked release was Gemini 3.8 Flash Cyber on Sep 2 2026, 28 days earlier. It is the most recent Google model tracked on AI Release Tracker.


---

Canonical page: https://aireleasetracker.com/model/google/gemini-4-argon
Site index: https://aireleasetracker.com/llms.txt
Source: AI Release Tracker (https://aireleasetracker.com). Most benchmark scores come from lab launch material; gathered results identify the leaderboard that published them.