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

Business Insider reportedly says Google employees are testing an internal Gemini 4 checkpoint called Carbon on Jetski, with one calling it feel like Opus 5.5 for coding.

A user filmed a Lego bin, had Gemini analyze it frame by frame to catalog bricks, then Argon generated interactive 3D models and instruction sheets.

The Information reports Google is developing a model codenamed Gemini 4 Carbon that appears to improve on last week's Gemini 4 Argon release amid competition with Anthropic and OpenAI.

After hand keypoints worked, the author gave Gemini 4 Argon a numbered body diagram and asked it to locate those points in an image, sharing the prompt, results and mistakes.

Researchers used Gemini 4 Argon agents inside Antigravity to extend Landreman's breakthrough with new symmetry-breaking exact solutions, giving fusion scientists fresh benchmarks for reactor simulation codes.

The author says they want to like Gemini 4 Argon, but the app often crashes or drops prompts and frequently times out before the first token.

A long post maps the nine-week launch timeline of Grok Bot, GPT-6 Astra, Opus 5.5, OpenAI Dots and Gemini 4 Argon, compares two 24/7 agent fleets' limits, and proposes Jev as an exit gate.

A provocative post lines up Gemini Argon 4, GPT-6 Astra, Grok 4.7 and Muse Wonder (claimed to be Fable 5.5) for comparison.

Model-agnostic harness Tasklet now auto-routes tasks across Anthropic, OpenAI, Google and xAI, with cross-model subagent delegation that can hand work to Gemini 4 Argon.

A video compares DeepSeek V4.1 Flash, GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon in real use, saying Argon loses focus and wanders mid-task.

On Artificial Analysis' new AA-PersonalityBench testing humor and playful language across 36 models, Gemini 4 Argon ranks first at 1573 Elo, ahead of Claude, GPT, Grok and DeepSeek.

Argon ranks third on long-horizon real-world SWE tasks behind Opus 5.5 and Fable 5.1 at one-third the cost, tops greenfield implementation at 63%, leads non-GPT frontend fidelity, and lags in identity provisioning.