Gemini 4 Argon全面开放 体验糟糕不推荐开Ultra会员

Recently, a new wave of large language models has hit the market, with the latest being Gemini 4 Argon from Google, which finally opened to the public on October 14 after a brief, limited release on October 1. The earlier launch was only for a select security program, available to government agencies, critical infrastructure teams, and security firms, so the general public had to wait two weeks to access it.

Looking back at the past month, major tech players have been launching models nonstop: GPT-6 Astra from OpenAI on September 4, Grok 4.7 on September 22, Claude Opus 5.5, GPT-6 Sol, and Luna on September 23, GPT-6.1 Sol during OpenAI's DevDay on September 30, Meta's Muse Wonder on October 13, and finally Gemini 4 Argon on October 1 and October 14 respectively.

Currently, the top tier of models is led by Anthropic, with OpenAI close behind, while Meta and Google are playing catch-up, and Grok is still lagging far behind in performance.

To test Gemini 4 Argon, I upgraded to Google's Ultra plan for $100, even though I haven't paid for Gemini in a long time. After a few hours of testing, my take is that this model is unlikely to challenge the positions of Claude Opus 5.5 and GPT-6 Astra as top-tier models.

Google's Turbulent Road to Gemini 4 Argon

Before diving into the model itself, it's important to understand the context that led to Gemini 4 Argon, as it represents a make-or-break moment for Google. Back last November, Gemini 3 Pro was considered one of the best models worldwide, but then Google entered a chaotic period. In February, Gemini 3.1 Pro was released, followed by a promise at May's I/O event to launch Gemini 3.5 Pro within a month. The launch kept getting delayed, with Google citing insufficient code capabilities and pressure from competitors like Anthropic and OpenAI. Eventually, the project was canceled internally.

On August 5, Google's AI division underwent major changes: Demis Hassabis stepped down as CEO of Google DeepMind, becoming chairman and Alphabet's chief scientist, replaced by Koray. That same day, Jeff Dean, a 27-year Google veteran known as the "姐夫," announced his departure to start a new venture. Jeff Dean joined Google in 1999 when it was still a small company, co-creating foundational technologies like MapReduce, BigTable, and Spanner with Sanjay Ghemawat. He co-founded Google Brain in 2011, led the creation of TensorFlow, and was a key technical lead for Gemini. Google even created a special 11th-level position for him and Sanjay. A popular meme, "Jeff Dean Facts," has circulated for years, with jokes about his programming skills being so advanced he optimizes physics and doesn't receive compiler warnings.

With Jeff Dean and Sanjay gone, industry observers noted that Gemini was struggling while Google Cloud was thriving. Google had allocated many computing chips to competitors like Anthropic, so their own models were falling behind, but their cloud business selling computing power was growing. For months, Google relied on its Flash series models, while competitors like OpenAI released weaker models and Anthropic released more niche offerings.

Against this backdrop, Gemini 4 Argon is not just another model release for Google—it's a last-ditch effort to regain ground in the competitive AI landscape.

Basic Specs and Pricing

Let's start with the basic details. Gemini 4 Argon is now available: the $20 Pro plan works with the Gemini app, but to use it in Google's development tool comparable to Codex and Claude's desktop app, you need the $100 Ultra plan. This pricing structure has drawn criticism, especially since the translation of the model's performance tiers is unclear.

Gemini models are multi-modal, supporting text, images, audio, and video. The knowledge cutoff is March 2026, which is older than GPT-6 Astra and Claude Opus 5.5. There's a key detail about context windows: when announced on September 30, Google said the output limit was raised from 64k tokens to 1 million, but now it's clarified that the total input and output combined is 1 million tokens. The model uses a continuation protocol: each request can output up to ~64k tokens, and when it reaches that limit, it returns a token to continue writing, allowing developers to extend the output up to 1 million tokens. In comparison, GPT-6 Astra has a native output limit of 128k tokens per request, higher than Argon's.

Pricing is as follows: during the promotional period until November 30, it's $2 per input token and $10 per output token, the same as GPT-6.1 Sol and Claude Sonnet 5.5. After that, it will rise to $4 per input and $20 per output, matching Claude Opus 5.5. For agent tasks, cache costs are important: Argon costs $0.1 per unit during promotion, $0.2 after December 1, while GPT-6.1 Sol is $0.1, Claude Sonnet 5.5 is $0.1, Claude Opus 5.5 is $0.2, and GPT-6 Astra is $1.

Third-Party Performance Tests

Independent tests provide a more objective view. In the AA benchmark, Argon scored 53, tying with GPT-6 Astra. The AA test heavily weights programming and agent tasks, where Claude Opus 5.5 is clearly first, followed by Claude Sonnet 5.5. A notable detail: Argon's average cost per task in AA tests was $1.99, while GPT-6.1 Sol Max was only $0.72, even though their API prices are the same. This is because Argon generates more output per task than Sol, though it's less "hyperactive" than Claude models which produce high-quality work but more extensive output.

Another test from Vals, a firm that evaluates professional skills in fields like finance, law, medicine, and programming, gave Argon an overall score of 68.9, ranking first. The scores for other models: Claude Sonnet 5.5 at 67.04, Opus 5.5 at 66.97, Fable 5.1 at 65.83. Argon also has lower costs, at $15.68 per test run, compared to Opus 5.5's $32.14. Argon excels in finance, law, tax, medical, and programming tasks, but has a major weakness: in the CUA-bench test for computer control, it scored only 4.83%, second to last among 8 models, with Gemini 3.8 Flash at the bottom. GPT-6 Astra was first at 19.17%, Opus 5.5 at 14%. So Argon's overall capabilities are: knowledge work > research > coding > agent tasks.

Usage Experience: Full of Pain Points

Beyond specs, the actual usage experience with Gemini 4 Argon is quite poor. The Gemini web version has long been criticized for bad tool calling and reasoning, and the same applies to its desktop tool. One major issue is language support: competing tools like Codex added Chinese support on launch, and Claude added Chinese support quickly, but this tool, launched almost a year ago, still has no language settings. I checked with competing tools and confirmed there's no option to switch languages.

Another problem is extreme quota consumption. I used my Ultra plan for less than 10 tasks: 5 frontend tests, 3 content creation tasks, and 1 reverse engineering task, and used up 27% of my weekly quota, even though none of these were parallel agent tasks. In comparison, using other top models on a $100 plan doesn't have such high consumption. The frontend output was raw HTML, with no explanation or guidance, requiring manual saving and running to see the result. The quality of the frontend work was far worse than competing models, with much lower detail. I couldn't complete backend tests due to running out of quota, even though I had prepared a set of tasks from my own projects with real bugs that needed fixing.

In content writing, Argon ranked first in an English-only blind test, but in Chinese, the writing is awkward, with leftover phrases from older Gemini models, strange metaphors, and poorly crafted titles that feel unnatural.

Overall, after testing, I don't recommend upgrading to the Ultra plan. If you want to try Gemini 4 Argon, start with the $20 Pro plan. As for my own subscriptions, I use multiple AI services for different needs, but ChatGPT remains the most balanced option. Gemini 4 Argon, based on current performance and experience, won't be added to my regular subscriptions.