Google (GOOGL.US) Gemini 4 Argon Global Open: Refactoring Code and Bots

According to Woofun AI, Google (GOOGL.US) officially opened the Gemini 4 Argon model to global developers and enterprises on October 13, marking a shift in its strategic focus from simple content generation to full-process delivery of complex tasks. This launch not only covers extensive access through platforms such as Gemini API, Google AI Studio, and Google Antigravity, but also opens up AI mode experiences in apps and searches to eligible Gemini Pro and Ultra subscribers. Compared to the limited opening on October 1, the core value of this global launch is a series of empirical cases revealed by Google, which aims to prove that the value of cutting-edge models has gone beyond the initial stages of answering questions or generating code, and instead focuses on whether it can take over and connect multiple links to ultimately deliver results with actual use value.

In terms of internal engineering restructuring and performance optimization, Google first used its own huge engineering project as a test field to test Argon's ability. According to the data, more than 100,000 employees have used the model through the internal version of Antigravity. In the past four weeks, the engineering team's intelligent code submission surged by 35%, while the code rollback rate after submission dropped by nearly one-third. This significant increase in efficiency directly reflects the model's progress in code quality control. The Chrome team's refactoring project is particularly typical. The Chromium codebase is 40 million lines long and contains a directory of around 6 million lines of legacy code that has long mixed multiple browser features and is closely intertwined with core web processing logic. Argon helped the team successfully strip core functionality from the legacy code, complete an architecture restructure involving hundreds of files, and reduce work that was expected to take months to a few weeks.

Additionally, Google's open source Agent Development Kit (ADK) team conducted a similar experiment by providing Argon with a rulebook and optimization goals, requiring it to reduce framework overhead while ensuring the stability of existing applications. After an overnight run, the proposed changes in the model reduced memory preparation time for long sessions by 40%, increased cold import speed by 21%, and related code was merged into the public ADK Python code library. At the level of underlying model optimization, Argon increased Gemma 4 E2B's inference throughput on a single NVDA.US (NVDA.US) A100 from about 37,900 tokens per second to over 108,000, an increase of about 2.85 times. Google pointed out that the optimized output remains consistent, and the whole process only takes a few days, mainly involving technical aspects such as attention tuning and speculative decoding. Together, these cases reveal a key trend: models are deeply involved in the entire engineering process, from code understanding and architecture adjustments to performance optimization and troubleshooting, rather than just following instructions to generate code.

Data compiled by Woofun AI shows that visual generation and cross-platform game porting cases have further expanded Argon's application boundaries. In the field of 3D modeling, in the example shown by Google, Gemini 4 Argon independently built a set of highly detailed temple buildings in Blender through the multi-agent visual criticism cycle in Antigravity, including multiple buildings, vegetation, and complex lighting, showing a visual effect close to a photograph.

The “Mind Killer 2” porting case shared by X user @mweinbach shows Argon's powerful capabilities in cross-platform engineering. He claims that the model helped port the game from the x86 Windows environment to the native macOS Arm64 environment, involving instruction conversion, Rust binding, and Metal shader conversion, and the project achieved a smooth running effect of over 140 frames per second on the M3 Ultra.

Large codebases test Argon's ability to handle complex software systems, while robotics projects test the model's ability to understand real-world spatial relationships and physical interactions, and translate that understanding into executable programs. The Real2Sim project presented by Google revolves around this problem. Argon can watch real robot remote operation videos, build corresponding digital twin environments, and obtain or generate simulated assets, so that objects not only look close to real objects, but also show similar physical behavior when in contact with dexterous robots. The model then adjusts object properties and scene settings through repeated visual reasoning, so that the visual effects and contact process in the simulation are as close to the real record as possible.

The significance of this process is that robot training does not have to rely entirely on humans to build a simulation environment from scratch. The actual operation video can be the starting point for modeling, and the model then continuously corrects the scene according to the differences between simulation and reality. Another demonstration starts with a sketch. Argon automatically generates a GPU-accelerated simulation scene and full-body motion trajectory based on a diagram of a humanoid robot throwing a ball into a pile of boxes on a table, and verifies that all six boxes have fallen in the physics engine, and finally outputs a rendered video.

However, successful simulation does not mean that the robot already has stable realistic operation capabilities. The physical environment, sensor errors, and actuator limitations may all affect the motion migration effect.

Breakthroughs in cybersecurity, scientific research, and life sciences demonstrate Argon's deep application capabilities in specialized fields. In the field of cybersecurity, Google is opening the model to more partners for authorized defensive security work through the Fairwind program. In scientific research, Argon has participated in quantum computing, nuclear fusion, and life science projects. According to Google, the model converted an internal quantum error correction decoder into an optimized hardware core, which nearly doubled the inference throughput; in collaboration with researchers at the University of Maryland, it helped discover more counterexamples of Grad's conjecture, providing a new research basis for fusion reactor magnetic field design.

On the life sciences side, Gamow Labs researchers used Argon to analyze the genome of newborns with rare developmental lung diseases, discovered in one case a chromosomal rearrangement that conventional software had not recognized for many years, and explained the reason for previous omissions. What these projects have in common is that models need to understand specialized materials, generate solutions, and then adjust based on computational or experimental results. Google is trying to extend this capability from software engineering to specialized fields with different rules and verification methods, marking a shift in AI's role in scientific discovery from an aid tool to a core player.

Commercial pricing, benchmarking, and expert assessments form key variables for Argon's entry into the enterprise workflow. The limited-time API price given by Google is $2 per million input tokens, $10 for output, and $0.1 for cached input. Starting December 1, the input and output prices will increase to $4 and $20, respectively. Argon also provides a unified context window of 1 million tokens, sharing this quota for input, thought, and output. The API supports background operation and checkpoint recovery, and is suitable for tasks that take a long time to perform, such as multi-repository migration and complex code analysis.

Benchmarks showed that Argon was competitive in some professional tasks. On the Vals Index, it scored 68.9%, higher than Claude Sonnet 5.5's 67.04% and Opus 5.5's 66.97%. In tests published by Google, Argon scored 19.6% on the Harvey Legal Agency Benchmark, 51.3% on the Zapier Automation Benchmark, and 77.9% on the DeepSWE Software Engineering Benchmark.

But in the Artificial Analysis Composite Intelligence Index, Argon scored 53 points, which is on par with GPT-6 Astra and Claude Fable 5.1, and lower than Opus 5.5, Sonnet 5.5, and Meta Muse Wonder Full. Wharton School professor Ethan Mollick (Ethan Mollick) believes Gemini 4 has made significant advances in creativity and writing, providing different options for existing cutting-edge models.

However, he pointed out that the model is still lacking in handling complex multi-agent tasks, and that it is sometimes too confident, and that the Antigravity tool chain also needs to be improved. Sam Witteveen (Sam Witteveen) of corporate AI consulting firm Red Dragon AI believes that Argon is an excellent but not the strongest model. He is optimistic about the model's ability to combine visual understanding, software generation, and creative media, but believes there is still a gap between Antigress and development tools such as Claude Code and Codex. X user @cheaty believes that Argon's early tests did not perform well at running speed and cost.

Although the model was able to complete the task, the actual cost of some tasks was significantly higher than other models. This also shows that when evaluating models, companies can not only compare the price per million tokens, but also look at how many calls it takes to complete the same task, how long it takes, and whether the final results are usable.

Argon's full openness gives more developers and businesses an opportunity to test Google's ability to display. However, there is still some distance between the model being able to complete a task and the enterprise can stably and inexpensively put it into production. For Google, the next solution is not only continuing to improve model capabilities, but also making models easier to use, making complex tasks more stable, and making actual cost and efficiency advantages truly stand the test. Google employees are testing Gemini 4's subsequent versions, Barium and Carbon, and some testers believe their coding ability has improved. Vercel CEO Guillermo Rauch (Guillermo Rauch) believes that Google has Transformer technology accumulation, engineering talent, and computing resources, and should not be underestimated.