Google releases Gemini 4 Argon for coding and cyber defense

Google releases Gemini 4 Argon for coding and cyber defense

Google has publicly launched Gemini 4 Argon, bringing its newest frontier model to paying Gemini app users after an initial rollout restricted to trusted cyber defenders. The release emphasizes complex software engineering work and cybersecurity defense, alongside new tools that let Gemini handle longer-running tasks.

The first model in Google’s Gemini 4 generation now leads the company’s lineup. Google CEO Sundar Pichai announced the launch on X, saying it would be available globally starting today.

Google first introduced Argon on Sept. 30 in a blog post by Google DeepMind SVP Koray Kavukcuoglu. Access initially went to trusted cyber defenders in Google’s Fairwind Program. Google AI Pro and Ultra subscribers can now use the model in the Gemini app.

Coding tools move into the Gemini app

Google says Argon already performs much of its internal engineering work. With the public launch, Google Antigravity’s agentic coding interface is becoming part of the Gemini app on macOS, allowing users to switch between conversations and coding without changing apps.

A separate toggle moves Gemini from chat into tasks mode, where it can continue working across Google apps and other services while users are away from their computers. Both Pro and Ultra subscribers receive access.

Google says the model is designed for extended, complicated assignments, including software engineering, cybersecurity defense, and enterprise work in legal and finance. For longer tasks, it provides more frequent progress updates. In everyday conversations, Google says Argon repeats itself less, uses fewer clichés, and responds with less hedging.

sbb-itb-622dc24

Cyber safeguards face testing

Before the broader release, Google rolled out Argon in phases beginning with cyber defenders and participated in the U.S. government’s voluntary process for pre-release model access.

Google says the model is designed to reject requests that could assist cyberattacks or chemical, biological, radiological, or nuclear attacks. Red teams tested those protections, according to the company.

Google also describes Argon as its most resistant model yet to indirect prompt injection, in which malicious instructions hidden inside material a model reads attempt to redirect its behavior.

On Gray Swan’s benchmark, Google reports that such attacks succeeded against Argon 0.7 percent of the time. The reported rates were 1.0 percent for both Opus 5.5 and Fable 5.1, and 8.5 percent for GPT-6 Astra. Lower scores indicate better performance on that test.

Benchmarks show strengths and limits

Google’s reported coding results are mixed. Argon reached 77.9 percent on DeepSWE v1.1, setting a new high on Google’s main benchmark for long software engineering tasks. But it finished behind GPT-6 Astra on FrontierSWE v2 and behind all three rivals on Terminal-Bench 4.0, where Opus 5.5 led.

The company also reports strong results in business tasks. Argon leads the Vals Index, which weights finance, coding, legal, and tax work according to each sector’s share of U.S. GDP. It scored 51.3 percent on AutomationBench, Zapier’s evaluation of end-to-end business tasks, about nine points ahead of the next model.

Those figures come from Google. Some competitor scores in its comparisons differ from figures Anthropic previously reported. When Anthropic launched Opus 5.5, it also cautioned that "benchmark margins have become a less reliable guide to real-world differences."

The independent Artificial Analysis Intelligence Index gives Argon’s "High" setting a score of 53, tying GPT-6 Astra and placing it one point above GPT-6.1 Sol and Muse Wonder Max. Claude Opus 5.5 and Claude Sonnet 5.5 score about five points higher.

Artificial Analysis combines 10 evaluations spanning math, science, coding, agentic tasks, and reasoning. It says it conducts the tests itself on standardized hardware rather than relying on companies’ reported results. Argon’s score represents a 23-point improvement over Gemini 3.1 Pro Preview.

The launch follows other frontier-model releases, including Anthropic’s Claude Opus 5.5, OpenAI’s GPT-6 Astra, and Meta’s "super intelligence" Muse Wonder, which was released yesterday.

Access and introductory pricing

Gemini app access requires a Google AI Pro or Ultra subscription. Pro costs $19.99 a month. At I/O 2026, Google introduced a $100-a-month Ultra plan and reduced its top Ultra tier from $250 to $200 a month.

Google announced introductory API pricing of:

• $2 per million input tokens
• $10 per million output tokens
• $0.10 per million cached input tokens, a 95 percent discount

Those rates are below the listed prices for Claude Opus 5.5, at $4 and $20 per million input and output tokens, and GPT-6 Astra, at $10 and $50 per million tokens.

The public launch announcement centers on access through the Gemini app. It does not say whether API availability has expanded beyond Google’s earlier rollout plans.

Read the source