According to Zhitong Finance APP, on Wednesday local time, Google (GOOGL.US) officially launched its flagship AI model, Gemini 4 Argon, to developers and businesses worldwide. This is Google’s Gemini
The first model of the fourth generation, currently at the top of its model lineup. Google CEO Sundar Pichai announced on X that the model is now available globally. This move, following months of delayed releases and leadership changes at DeepMind, is seen as Google’s renewed direct challenge to OpenAI and Anthropic.
Google says Gemini 4
Argon excels at complex, multi-step "deep work" in fields such as programming, finance, law, cybersecurity, and science. In everyday chat scenarios, the model is designed to be more direct, less ambiguous, with fewer repetitions and less reliance on clichés; for longer tasks, it provides more frequent progress updates. Google also stated that Argon has already taken on a significant portion of the company’s internal engineering work, and this update will enable Google
Antigravity’s agent-based coding interface is embedded directly into the macOS Gemini app, allowing users to switch between chatting and coding without leaving the application.
It is reported that Koray, Senior Vice President at Google DeepMind,
Kavukcuoglu previously hinted at Argon in a blog post on September 30, when the model was limited to trusted network defenders within Google's Fairwind program. With today's public release, Google
AI Pro and Ultra subscribers can use Argon, AI in Search, in the Gemini app.
Mode is also available for integration; the Antigravity coding tool is restricted to Ultra subscribers only.
Google also announced that it will integrate Fairwind
The program is open to hundreds of new partners. The initiative provides unrestrained versions of Argon to vetted cybersecurity defenders. In September, Google offered this model only to a select group of cybersecurity partners. Additionally, Google will grant free six-month access to the highest usage tier of Antigravity to 10,000 scientists.
In terms of pricing, Google has set an aggressive introductory API price for Argon: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens, bringing the cost to just $0.10 per million. According to the first source, these rates will double after November 30. In comparison, Anthropic’s Claude
Opus 5.5 charges $4 and $20 per million input and output tokens, respectively, while OpenAI's GPT-6 Astra charges $10 and $50.
The subscription price for Google AI Pro is $19.99 per month. Google at I/O
In 2026, a new $100-per-month Ultra plan was added, and the top-tier Ultra monthly fee was reduced from $250 to $200. The Gemini app also introduced several new features: a toggle switch was added to the main chat window to switch between "Chat" and "Tasks" modes; in Tasks mode, Gemini can continue working across Google apps and other services even when the user steps away from their computer, a feature available to Pro and Ultra subscribers.
Gemini can currently connect to over 40 apps, including Instacart, OpenTable, Adobe, and Shopify, with Google positioning it to handle tasks such as ordering groceries, reserving restaurant tables, and designing or building online businesses.
Before the release of Argon, Google's AI division went through a challenging period. In September, the company stated it no longer planned to release Gemini 3.5.
Pro, while Pichai initially stated that the model would launch in June. DeepMind has also undergone restructuring following the retreat of co-founder and CEO Demis Hassabis to a secondary role, with several leaders of the Gemini model leaving the company. The release of Argon comes amid a busy autumn of frequent launches of cutting-edge AI models: Anthropic released Claude.
Opus 5.5, OpenAI released GPT-6 Astra, and Meta's "superintelligence" model Muse Wonder was also released yesterday.
On benchmark performance, Google's own evaluations show that Argon leads in multiple metrics. According to the independent firm Artificial Analysis's Intelligence
In the Index, Argon, with high inference mode enabled, ties with GPT-6 Astra at 53 points, surpassing GPT-6.1 Sol and the newly released Muse Wonder.
Max scores 1 point higher; however, the same index shows it trailing behind Claude Opus 5.5 and Claude Sonnet 5.5 by approximately 5 points.
The Google comparison table includes 18 results, with Argon ranking first in 14, tying with GPT-6 in cybersecurity, and trailing in 5. All data above is sourced from Google, though AI companies often also publish evaluations that favor their own models.
Google stated that Argon's maximum advantage is concentrated in knowledge work. It cited Vals
The Index weights financial, programming, legal, and tax jobs based on their share of U.S. GDP. On AutomationBench, which tests end-to-end business tasks on Zapier, Argon scored 51.3%, leading the next model by approximately 9 percentage points. Coding results were more mixed: Argon performed on DeepSWE, Google’s primary benchmark for long software engineering tasks.
Reached a new high of 77.9% on v1.1, but lagged behind GPT-6 Astra on FrontierSWE v2; on Terminal-Bench
On 4.0, Argon lags behind all three competitors, while Opus 5.5 leads.
Some of Google's competitor data also contradicts Anthropic's report. According to relevant reports, Anthropic compared GPT-6 Astra on Terminal-Bench
On the 4.0 score, it is listed at 57.9%, while Google lists it at 58.2%; on the Terminal-Bench Science benchmark, Anthropic reports Opus
5.5 is 58.7% at default effort, while Google lists it at 63.3%. Anthropic released Opus
At 5.5, it was warned that "the benchmark gap has become less reliable in guiding real-world differences."
Artificial
Analysis independently tested Argon's "High" setting. The index combines ten evaluations—including mathematics, science, programming, agent tasks, and reasoning—into a single composite score, and states that all tests were conducted on standardized hardware without relying on corporate self-reported data. In this index, Gemini
Argon (High) scored 53, significantly higher than the median score of 26 for similar models; although it still lags behind heavyweight models like Claude and ChatGPT, it outperforms Google’s previous top model, Gemini 3.1.
Pro Preview increased by 23 points. At the introductory price, each Intelligence Index task completed on Argon costs approximately $1.99, roughly equivalent to GPT-6
60% of Astra's cost; expected to rise to $3.98 after the discount ends, approximately 1.2 times Astra's cost. In comparison, Opus 5.5 costs $5.98 per task.
In terms of security and protection, before its consumer release, Google rolled out Argon in phases, initially targeting cybersecurity professionals while participating in the U.S. government’s voluntary pre-release model access program. Google states that for general users, the model is designed to reject requests that could assist in cyberattacks or chemical, biological, radiological, or nuclear attacks, and these safeguards have been tested by a red team.
Google also stated that Argon is its most resistant model to indirect prompt injection to date. In Gray Swan's Injection Prompt
In the test benchmark, this test measures how easily a model can be hijacked by malicious instructions hidden within the input content. Google reported a 0.7% success rate for attacks against Argon; Opus 5.5 and Fable
5.1 are all 1.0%, while GPT-6 Astra is 8.5%. In this test, lower values are better.