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

The roundup says the PHP2Shell flaw was found with GPT-5.6 Sol's help and independently reproduced by Gemini 4 Argon, alongside Meta earnings and the Muse Spark Open release.

Arena tests both Muse Wonder models on 3D scenes, animated HTML clips and website builds against Gemini 4 Argon, GPT-6 Astra and more.

On Code Arena WebDev, Meta's Muse Wonder Full (Max) debuts at #6 with 1726 points, 48 points above Gemini 4 Argon at #10, lifting Meta to the #3 lab spot.

A user had Gemini 4 Argon and Meta's Muse Wonder build the same game, reporting a surprisingly large gap between the two results.

A weekly tech roundup lists Meta Muse Spark Open, Google's Gemini 4 Argon launch, Tesla's rocket-powered Roadster, and Microsoft's Maia 300 chip.

Meta released Muse Spark Open, an 816B-parameter open-weights model with 42B active parameters built for coding and agents, tuned to run on two linked Mac Studios; in Cursor's benchmark, Muse Wonder scored 46.1% versus Gemini 4 Argon's 52.8%.

Gemini 4 Argon (high) ranks in the top four with a 60.6% success rate but is slower than OpenAI models and still costly even at a 50% discount, while Meta Muse Wonder reaches only 46.6%.

A Spanish post compares Meta's Muse Wonder and Wonder Full with Gemini 4 Argon: DeepSWE 1.1 coding scores are nearly tied (77.9% vs 78.8%/79%), and Argon costs $2/$10 per million tokens until prices double after late November.

Gemini 4 Argon (high) enters the Debate Benchmark at #4 with 1676 points, 190 above Gemini 3.1 Pro Preview, while Meta Muse Wonder 1.0 (high) debuts at #5 with 1655.

The Japanese AI podcast previews an episode on AI self-replication risk and takes a closer look at the now-public Google Gemini 4 Argon and Meta Muse Wonder.

The live New York show features the Google DeepMind team discussing Gemini 4 Argon, alongside Meta Muse Wonder and Anthropic's numbers.

iruletheworldmo recaps the week, noting Google's Gemini 4 Argon launch, Meta's Muse Wonder flop, and OpenAI reportedly delaying GPT-6.1 Astra over deception concerns.