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

The author explains Google's Fishfood/Teamfood/Dogfood naming tiers and says server configs show Gemini 4 Carbon actively testing at Teamfood stage, alongside a newly exposed Gemini 4 Argon Max Mode variant.

StatsWire checked Google's python-genai SDK git history and found the gemini-4-flash-preview line added June 10 and deleted August 10, absent from main for 2.5 months with no new info today.

A user running Arena tests saw output they thought was Haiku 5.5 but traced it to low-tier Gemini 3.8 Flash, speculating Gemini 4 Flash may already be in Arena.

A user asks which will launch first: Gemini 4 Flash, codenamed Hydrogen, or Fable 5.5.

A user reports Gemini 4 Argon blocked nearly all invalid GPT code reviews, catching 13 erroneous items.

Testing Gemini 4 Argon as it rolls out to all Antigravity users, the author found it refused all 100 black-box hacking labs but built a full 3D cathedral from a single image.

A 48-hour packet-capture test of Gemini 4 Argon finds Pro still at $15/week and Ultra at $50/5h, plus a mid-conversation cache reset that spikes single requests to $0.2–0.3.

The poster calls Gemini 3.8 solid but hopes Flash improves coding and cuts hallucinations, citing its speed and writing strength.

A leak claims Gemini 4 Flash is in internal testing under the codename Hydrogen, and could pressure OpenAI if it nears Argon at a lower price.

A Japanese user compares Opus 5.5, GPT-6 Astra and Gemini 3.1 Pro on one task, finds Gemini worst, and buys the Gemini 4 Argon add-on.

A post claims a gemini-3.8-flash-router appeared in Antigravity ACP two days ago and was speculated to be Gemini 4 Flash, now reportedly named Gemini 4 Hydrogen.

A Chinese user asks what happens to the US model race if Gemini 4 Flash, codenamed Hydrogen, really beats GPT 6.1 Sol.