Tech • AI • Robotics • Game
Gemini 4 Argon Is Here (Fully Tested)
9.4/10TL;DR
Google has launched Gemini 4 Argon, a new flagship AI model aimed at long, complex professional tasks, with early testing suggesting major gains in knowledge work, writing and code analysis, though coding-focused rivals remain strong.
KEY POINTS
Flagship model for complex work
Gemini 4 Argon is positioned as Google’s most powerful model to date, built for extended, multi-step tasks in areas such as coding, finance, legal work and cybersecurity. In benchmark results shared by Google, Argon placed first in 13 of 19 tests and tied for first in one more, with particular strength in knowledge work and long-context use.
Not a full coding takeover
Despite the strong benchmark showing, Argon does not clearly dominate every coding category. For software development, rivals including Opus 5.5 and GBT6 Astra were described as still highly competitive, suggesting the market remains split between models optimized for code generation and those better suited to broader professional reasoning.
Rapid app and game generation
In practical tests inside Anti-Gravity, Google’s coding tool, Argon built a top-down capture-the-flag paintball game in essentially two prompts: one to generate the game and one to add a difficulty slider. It also produced a Star Fox-style space shooter in a single prompt, indicating strong one-shot prototyping ability for interactive projects.
Mind-controlled game prototype
One of the more unusual demonstrations involved connecting a Bluetooth meditation headband to a computer and using measured brainwave states to control a game. Argon generated an application that automatically connected to the device and moved a rocket based on relaxation or concentration levels, producing a playable if slightly exaggerated form of “mind control.”
Prompting matters for visual polish
Tests indicated that Argon’s coding output is capable but may require more explicit prompting than some rivals to achieve polished visuals. Functional results came quickly, but better-looking games needed more direction, especially compared with Opus 5.5, which was seen as stronger at producing attractive outputs with minimal instruction.
Useful at finding bugs in existing code
Argon also showed value in code review. When given code for an app that sent messages from an Apple Watch to a chatbot through webhooks, the model identified several bugs, including one that could potentially have exposed the webhook. While described as a low-risk issue, such exposure could have allowed outside control of the bot and access to sensitive information, making the fix significant.
More natural and grounded responses
In the Gemini app, Argon was described as noticeably more natural in conversation, with fewer stock phrases and more direct replies. It also performed better at staying grounded in user-provided documents, reducing the tendency to invent missing details and improving reliability for document-based question answering.
Greater persistence on long tasks
Another reported improvement is task persistence. Argon appears less likely to stop early or deliver incomplete work, continuing until a job is fully addressed. That trait was especially noticeable in Gemini Spark, where the model handled research tasks and assembled a PowerPoint presentation with stronger overall execution.
Writing emerges as a standout strength
Beyond research and coding, Argon appears particularly strong at writing. It performed well across scriptwriting, technical writing and creative fiction, including a long-form horror story designed to take 20 to 30 minutes to read. The output was described as cohesive and engaging enough to hold attention in a single sitting.
AI video editing shows maturity
Argon was also used to edit a short video and add motion graphics with limited instruction. The model transcribed and structured the script, then assembled the edit with only a few minor mistakes. In this area, it was judged among the best models tested so far, though still behind Opus 5.5.
Rollout and access
Gemini 4 Argon is rolling out globally to Google AI Pro and Ultra subscribers. It is also expected in AI Mode inside Google Search, while developers can access it through the Gemini API and Google AI Studio. Business customers can obtain it through Gemini for Business.
Main weakness may be the tooling
The strongest criticism was aimed not at the model itself but at the surrounding product experience. Anti-Gravity was seen as lagging behind competing interfaces, with the view that Google’s software harness still needs work to match the quality of the underlying model.
CONCLUSION
Gemini 4 Argon appears to mark a significant step forward for Google, especially in long-context reasoning, writing and professional knowledge work. Its success may depend not only on model quality, but on whether Google can improve the tools wrapped around it.
Ask a question