Google DeepMind has announced Gemini 4 Argon, a new model designed for long-horizon software engineering, enterprise knowledge work in legal and finance, and cybersecurity defense. The primary technical shift is output length. Argon can generate up to 1M tokens in a single response, an increase from the 64K limit on earlier Gemini versions.
Announcement details
The model targets complex workflows. Google is using a phased approach. It is participating in the U.S. government’s voluntary process for pre-release model access. The company will gather feedback from early testers and iterate on guardrails before a wider release.
Pricing is already public. Argon launches at an introductory $2 per 1M input tokens and $10 per 1M output tokens. Cached input tokens receive a 95% discount, which works out to $0.10 per 1M. After the introductory period, pricing moves to $4 input and $20 output. Logan Kilpatrick confirmed the introductory $2 in and $10 out pricing.
Why the 1M output limit matters
Current frontier APIs cap a single response far lower. Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra each allow 128K output tokens.
Google states that Argon can think deeply and generate hundreds of thousands of tokens in one trajectory. For developers, that means large refactors or long reports without splitting work across turns. The cost is real, though. A full 1M output tokens costs $10 at introductory pricing and $20 after.
Google has not disclosed Argon’s input context window.
Benchmarks: Where Argon leads and where it trails
Google compared Argon against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Argon leads outright on 12 of 18 benchmarks and ties for first on 1.
Where it leads:
- DeepSWE v1.1 (long-horizon software engineering): 77.9%, a new state of the art. Opus 5.5 scores 74.2% and GPT-6 Astra 74.1%.
- Vals Index (economic impact across finance, coding, legal and tax): 68.9%, ranked first.
- AutomationBench (Zapier, end-to-end business execution): 51.3%, ranked first. Opus 5.5 scores 42.5%.
- Harvey Legal Agent Benchmark: 19.6%, against 5.4% for GPT-6 Astra.
- LVBench (long video understanding): 91.7%, a new state of the art.
Where it trails:
- FrontierSWE v2: 55.0%, behind GPT-6 Astra at 65.5%.
- Terminal-Bench 4.0: 57.4%, behind Claude Opus 5.5 at 66.4%.
- OSWorld-2.0 (computer use): 69.2%, behind GPT-6 Astra at 72.6%.
Artificial Analysis reported that Argon equals GPT-6 Astra on its Intelligence Index at 60% of the cost per task, using discounted prices.
Cyber defense: Find, Validate, Patch
Google trained Argon to autonomously find, validate and patch critical software vulnerabilities. Trusted defenders and internal Google teams receive it without cyber guardrails.
On CWE-bench v1, which tests vulnerability remediation, Argon ties for first at 68%. The rival models on that leaderboard run inside their own agent harnesses.
Wiz is already using Argon through its Scan for Good initiative. The model found a critical vulnerability in healthcare software used by hospitals worldwide. Google says previous frontier models had missed it.
Before broad release, Google is strengthening safeguards in 4 areas:
- Misuse defenses for cyber and CBRN risks, including activation monitoring, under its Frontier Safety Framework.
- Indirect prompt injection resistance, where Argon leads Gray Swan’s IPI benchmark.
- Misalignment monitoring of chain-of-thought and actions, with the ability to stop execution.
- Sealed, isolated sandboxes for high-risk training and evaluations.
Argon inside Google
Thousands of Googlers already use Argon. Google shared 4 internal results:
- Argon agents applied memory optimizations across data centers, freeing over 300 TiB, with 500 TiB to 1 PiB projected.
- Agents replaced 32K lines of SIMD code in the libgav1 Rust port. The decoder runs 2.7x faster with identical output.
- Agents are migrating C/C++ codebases to Rust, up to 800K+ lines in the Fuchsia Zircon kernel.
- Argon beat a published quantum algorithm baseline by 40% in minutes.
Comparison: Gemini 4 Argon vs closest competitors
| Feature | Gemini 4 Argon | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Developer | Google DeepMind | Anthropic | Anthropic | OpenAI |
| Availability | Fairwind Program only | Claude API and clouds | Claude API and clouds | OpenAI API |
| Max output per response | 1M tokens | 128K | 128K | 128K |
| Context window | Not disclosed | 1M | 1M | 1.05M |
| Input / output price (per 1M) | $2 / $10 intro, then $4 / $20 | $4 / $20 | $10 / $50 | $10 / $50 |
| Cached input (per 1M) | $0.10 (intro) | $0.20 | $0.25 | $1.00 |
| Open weights | No | No | No | No |
| DeepSWE v1.1 | 77.9% | 74.2% | 67.4% | 74.1% |
| Vals Index | 68.9% | 67.0% | 65.8% | 63.1% |
| FrontierSWE v2 | 55.0% | 62.3% | 56.3% | 65.5% |
| Terminal-Bench 4.0 | 57.4% | 66.4% | 57.9% | 58.2% |
| CWE-bench v1 | 68% (tie) | 67% | 58% | 68% (tie) |
Sources: Google, Anthropic Opus pricing, Anthropic Fable 5.1 docs, OpenAI GPT-6 Astra docs, OpenRouter. Benchmark scores are from Google’s published comparison. GPT-6 Astra prices are its short-context tier.
What it means
Users working in high-volume output scenarios will see a shift in how they handle large tasks. The 1M token limit allows for generating extensive code or reports in a single pass, removing the need for iterative summarisation. However, the cost of generating that volume is substantial at $10 to $20 per million tokens. Access remains restricted to the Fairwind Program, meaning public users must wait for a wider release date.
