Did Google just dethrone OpenAI and Anthropic?

Summary

The video reviews Google’s newly released Gemini 4 Argon, comparing it with GPT-6 Astra and Claude Opus 5.5. The host presents Argon as a major improvement over earlier Gemini models, but finds its strengths uneven.

Model and training

The video describes Argon as a sparse mixture-of-experts transformer trained from scratch on Google TPUs, rather than as a fine-tune of an older Gemini model. Its stated knowledge cutoff is March 2026. The host calls it an “omni” model that can work with multiple input and output formats.

Benchmarks

Argon is reported to lead on deep software-engineering tasks, the Val index, legal questions, and long-video understanding, scoring 91.7% on LV Bench. Its results are not uniformly the best:

• Claude Opus 5.5 beats it on Terminal Bench 4.0.
• GPT-6 Astra leads on Front Tier SWE by more than 10 points.
• Argon ranks ninth for web development on Code Arena, despite placing first on Arena AI’s general text ranking.

Pricing and access

The video cites introductory pricing through November 30 of $2 per million input tokens and $10 per million output tokens.

The host also criticizes the rollout. Argon initially replaced Gemini 3.1 in Google’s Antigravity environment for AI Ultra subscribers. Pro users received access later and with a lower quota.

Hands-on tests

• Promo video: Argon generated a polished-looking but, in the host’s view, mediocre AI promo. The host preferred Claude’s code-generated motion graphics, finding them more attractive and easier to customize and inspect.
• Interactive 3D horse in Three.js: Claude and GPT-6 produced comparatively convincing results. Argon’s output had a clean interface but a poor horse model and an inappropriate anatomical detail, which the host used as a humorous example of an alignment failure.

Overall assessment

Argon appears strong on several benchmarks and is competitively priced, but the host’s coding and generation tests suggest it does not consistently outperform Claude or GPT-6.

Sponsor feature

The video also promotes Macroscope, an AI tool for reviewing and approving pull requests. It uses language-specific AST walkers to understand code changes, runs custom rules as GitHub checks, and can automatically approve qualifying PRs after checks and review. The host says it auto-approves 40% of its customers’ PRs.

Main speakers and sources

• Fireship narrator/host: Provides the model review, benchmark commentary, and hands-on comparisons.
• Google Gemini 4 Argon: The model being tested; its claimed capabilities and benchmark results are discussed in the video.
• Claude Opus 5.5 and GPT-6 Astra: Competing models used for comparison.
• Macroscope: Sponsor featured in the pull-request review segment.