Google Releases Gemini 4 Argon for Pro and Ultra Subscribers

Google has launched Gemini 4 Argon worldwide, making the first model of its Gemini 4 generation available to paid subscribers in the Gemini application. Google Chief Executive Sundar Pichai announced the general release on X, expanding access to Google AI Pro and Ultra subscribers following an initial restricted rollout announced on Sept. 30 by Google DeepMind Senior Vice President Koray Kavukcuoglu for cyber defenders in Google's Fairwind Program, as reported by Mashable Tech (https://mashable.com/tech/google-launches-gemini-4-argon).

The release arrives as frontier model developers accelerate deployment schedules following earlier delays with Google's Gemini 3.5 Pro. Google positions Argon for complex software engineering, enterprise knowledge work, and defensive cybersecurity, claiming the model produces less conversational hedging and provides regular status updates during long-running tasks.

Along with the foundation model, Google updated the Gemini client interface. The macOS Gemini application now directly embeds Google Antigravity's agentic coding interface, allowing users to alternate between chat prompts and code editing without leaving the app. Google also added a dedicated tasks toggle that permits Gemini to execute background operations across Google applications and connected third-party platforms while users are away from their computers. Integration support has expanded to more than 40 external services, including Shopify, Adobe, OpenTable, and Instacart.

Benchmark results present a nuanced competitive picture across vendor-reported and independent testing. In Google's own 18-evaluation comparison table, the company reported Argon holding top scores in 14 tests, tying OpenAI's GPT-6 in cybersecurity, and trailing on five metrics. Google highlighted a 51.3 percent mark on Zapier's AutomationBench business task suite and a 77.9 percent score on its internal DeepSWE v1.1 engineering test. However, Argon trailed GPT-6 Astra on FrontierSWE v2 and fell behind Anthropic's Claude Opus 5.5 on Terminal-Bench 4.0.

Independent testing by the Artificial Analysis Intelligence Index—which runs standardized evaluations across math, science, reasoning, agentic tasks, and coding—scored Gemini 4 Argon on its high reasoning setting at 53. That ties GPT-6 Astra and places Argon one point ahead of GPT-6.1 Sol and Meta's Muse Wonder Max, while trailing Claude Opus 5.5 and Claude Sonnet 5.5 by five points. Artificial Analysis noted that the score marks a 23-point increase over Google's previous flagship, Gemini 3.1 Pro Preview. Introductory testing costs on the index stand at $1.99 per task, with standard pricing projected at $3.98 per task compared to $5.98 for Opus 5.5.

Mashable Tech noted discrepancies between Google's published competitor scores and figures previously disclosed by Anthropic. Google listed GPT-6 Astra at 58.2 percent on Terminal-Bench 4.0, compared to Anthropic's reported 57.9 percent, and cited Opus 5.5 at 63.3 percent on Terminal-Bench Science, where Anthropic had recorded 58.7 percent under default effort settings. Anthropic previously noted at launch that narrow benchmark margins offer diminishing reliability for real-world performance differences.

Access to Gemini 4 Argon is bundled into consumer subscriptions rather than metered API billing. Google AI Pro costs $19.99 monthly, while Ultra plans run between $100 and $200 per month following pricing restructuring at I/O 2026. By comparison, Anthropic bills Claude Opus 5.5 via API at $4 per million input tokens and $20 per million output tokens, while OpenAI charges $10 and $50 per million tokens for GPT-6 Astra. Google has not specified whether developer API availability has widened beyond earlier deployment phases.

On safety evaluations, Google stated that Argon underwent red-team testing and pre-release review under the U.S. government's voluntary framework, incorporating safeguards against assisting cyberattacks or chemical, biological, radiological, and nuclear threats. On Gray Swan's Injection Prompt Test measuring vulnerability to hidden malicious instructions, Google reported an attack success rate of 0.7 percent for Argon, compared to 1.0 percent for Opus 5.5 and Fable 5.1, and 8.5 percent for GPT-6 Astra.