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

A post assigns models to tasks: Claude for stock picking, ChatGPT 6.1 Sol for investment analysis, and Gemini Argon for research.

AkitaOnRails publishes LLM benchmark v4 comparing Gemini 4 Argon, GPT-Sol 6.1, Terra, Kimi K3.1 Pro and Claude Sonnet 5.5, concluding frontier models have become commoditized.

AkitaOnRails publishes LLM Benchmark v4 pitting Gemini 4 Argon against GPT Sol 6.1 Terra, Kimi K3.1 Pro and Claude Sonnet 5.5, concluding frontier models are now good enough for daily coding.

A user reproduces Argon failing on large single-shot file writes because of a ~60s idle timeout, retrying 9 times, and shares a workaround of splitting edits into 100-line batches.

A developer fed Gemini 4 Argon a single spider turret concept image; it produced a procedural 3D model with rigging, mechanical motion and rapid fire, plus a showcase site.

A builder describes a three-tier AI setup: Claude Opus 5.5 as manager, three project leads, and 40+ workers including Gemini 4 and Gemini 3.8 Flash.

Artificial Analysis released Intelligence Index v5, where Gemini 4 Argon (high) scores 50, behind Claude Opus 5.5 at 58 and GPT-6 Astra at 57, as the closed-open gap widens to 12 points.

A user uses Gemini 4 Argon for 3D model development, generating an 'impossible butterfly' and calling its 3D modeling output impressive.

A Google Q3 earnings preview covers revenue, capex and GCP, noting Gemini 4 Argon processed nearly 25 trillion tokens in its first 72 hours, about 3x Gemini 3.0's launch.

Hands-on tests show Gemini 4 Argon costs only ~1.57x 3.8 Flash at default thinking levels and is cheaper on short tasks, though December price hikes widen the gap.

The post complains that Google AI Plus subscribers are locked out of accessing Gemini 4 Argon.

The author fed two days of CGM glucose screenshots to Gemini 4 Argon, GPT-6.1 Terra, Opus 5.5 and two others; readings matched but reasoning, speed and cost diverged.