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

A user says Antigravity finally reset Gemini and Argon quotas after the limits became ridiculous, and asks Google for more compute since the weekly cap hits within an hour.

Antigravity has reset Gemini and Argon usage quotas, which a user says was badly needed given how fast Gemini 4 Argon consumes usage.

Compound Engineering 3.31.0 adds Antigravity, including Gemini 4 Argon, as a Gemini peer for cross-model review, plus test and comment pruning in /ce-simplify-code and fixes for unattended review results and /lfg stalls.

deepagents now supports Gemini 4 Argon, with 1M input tokens, long decode up to 1M, and agentic video understanding, plus docs and release blog links.

A user who tested Gemini 4 Argon says it burns usage quickly, runs very slowly, and needs heavy direct guidance for good results, so they would rather use 3.8 Flash for now.

Gemini 4 Argon took first place on micro1's LegalReasonBench for long-horizon, multi-step legal reasoning with a score of 50.8%, which the author says leaves plenty of room to improve.

A developer experimenting with court mapping in Gemini 4 Argon found it pretty good overall but still wrong on some positions, and shared a video on how it works.

Android developers with a paid Google AI Pro subscription can log into Antigravity within Android Studio to use Gemini 4 Argon.

A user calls Argon a spiky model that is extremely strong at physics, and thinks only some of its spikes have been found so far.

Google's official account rounds up the week's highlights: the Gemini 4 Argon launch, the new Fitbit Edge and a Research@ event.

A user quotes the Gemini 4 Argon safety report line about warm conversational language implying human-like care and calls it one of the dumbest AI sentences they have seen.

Gemini 4 topped micro1's legal agent benchmark at roughly a third of the cost of other models, though no agent scored above 50%.