Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

A user says Gemini 4 Argon opens to Antigravity IDE users within 24 hours and shares early hands-on impressions after already getting access.

Google's graph mining teams used Gemini 4 Argon with Antigravity's Teamwork agent to compress months of RDMA networking and TPU kernel work into hours, with FarArray up to 180x faster.

Mercor's APEX-Agents leaderboard shows GPT-6.1 Terra at 80.9% Pass@1 in management consulting, still below Gemini 4 Argon's 88.4%, at $0.19 per task.

A tester finds Gemini 4 Argon ranks third on the DeepRSI benchmark, beating Grok 4.7, Kimi K3 and its Flash predecessor but trailing Fable and Astra.

A Google staffer says Gemini 4 Argon is slowly rolling out to Antigravity IDE users, with the update expected within 24 hours.

A brief reply to another post that simply names Gemini Argon, with no reasoning or use case given.

The poster shares a screenshot they present as proof that Google is using Gemini 4 Argon internally.

In the Code2Video-AA benchmark, Opus 5.5 tops the list, the GPT-6 family follows closely and Gemini 4 Argon has caught up, prompting a difficulty increase.

The new benchmark turns 114 marketing briefs into code-rendered videos; Claude Opus 5.5 leads by 93 points, GPT-6 Astra ties for second, and Gemini 4 Argon serves as the judge model.

9to5Google reports Google acknowledges Gemini 4 Argon ranks well behind Claude and GPT-6 on Android coding benchmarks.

In Antigravity, a single text prompt has Gemini 4 Argon model and rig assets in Blender, batch import them into Unity, and generate editor tools for wave physics, shaders and lighting.

A Base ecosystem daily recap notes Capminal integrated Gemini 4 Argon into OpenCAP with a 1M-token context window, among other trading and ecosystem updates.