# Did Google just dethrone OpenAI and Anthropic?

## Executive summary

The video compares Google's new frontier model, Gemini 4 Argon, against competitors like GPT-6 Astra and Claude Opus 5.5. Argon is described as a powerful, sparse mixture of experts transformer trained from scratch using Google's TPUs and Jack's Python library. While Argon shows state-of-the-art performance in deep software engineering and legal benchmarks, it lags behind competitors in specific areas, such as front-tier SWE and web development coding (Code Arena). The video also highlights the practical application of AI in development workflows, specifically detailing how Macroscope uses custom AST walkers and native GitHub checks to auto-approve PRs, reducing manual review load.

## Key takeaways

- Gemini 4 Argon Architecture: Argon is a new frontier model, trained from scratch over six months using Google's TPUs and Jack's Python library. It utilizes a sparse mixture of experts transformer architecture, distinguishing it from older Gemini fine-tunes.
- Performance Benchmarks: Argon is reported as state-of-the-art on the deep software engineering benchmark and number one on the Val index. However, GPT-6 Astra reportedly wins on front-tier SWE, and Argon dropped to number nine on Code Arena for web dev.
- AI Code Review Automation: Macroscope, a sponsor tool, uses custom AST walkers and native GitHub checks to review code changes. It can auto-approve PRs if they pass these checks and have a small blast radius, currently auto-approving 40% of customer PRs.

## Technical details

- LLM Architecture: Gemini 4 Argon is described as a sparse mixture of experts transformer, trained from scratch on Google's TPUs. Its knowledge cutoff is March 2026.
- Benchmarking: Key benchmarks include the deep software engineering benchmark, the Val index (weighing finance, legal, and coding based on US GDP share), and the LV bench (where Gemini dominates with 91.7%).
- Development Workflow Tools: Macroscope builds custom AST walkers for common programming languages, allowing its agent to get real context on every code change for faster, cheaper, and more precise PR reviews.

## Practical implications

- The increasing capability of LLMs (like Argon) for multi-modal output (image, video, sound) suggests a shift toward integrated, highly complex AI applications.
- For build engineers, AI-powered PR review tools (like Macroscope) are becoming critical for managing the volume of code changes, automating checks beyond simple linting.
- The performance gap between leading models (e.g., GPT-6 Astra vs. Argon) suggests that selecting the right model for specific tasks (e.g., legal vs. coding) remains crucial.

## Topics

Large Language Models, Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5, Build Engineering, CI/CD, Code Review, Transformers, TPUs, Macroscope, Fireship Newsletter

Source: https://www.youtube.com/watch?v=b2tjnxX8IHU