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

A user points out a double standard: Gemini 3.8 Flash topping DeepSWE was called benchmark saturation, Sol topping it was called SOTA, and Gemini 4 Argon topping it is called saturation again.

A Portuguese-language post says DeepMind released Gemini 4 Argon to a closed test group, scoring 53 on the Artificial Analysis intelligence index to match Astra and lead Sol by one point, with a 15% hallucination rate versus 51% and $1.99 per task at a 50% launch discount.

A hands-on comparison shows Gemini 4 Argon scoring 55.0% on FrontierSWE, nearly triple Gemini 3.8 Flash's 19.6%, and calls it the best Gemini model tested.

An opinion post notes that Google matched Astra/Fable 5.1 performance for the first time, and the community immediately called Argon benchmaxxed; the author asks whether the model is gaming benchmarks or people just aren't used to Google winning.

The poster lists Argon's 1M output window and agentic coding results and asks readers to vote on whether it beats Opus 5.5, GPT-6.1-sol or Grok 4.7.

A video argues that after GPT-6 Astra and Fable 5 many thought Google was out of the AI race, but Gemini 4 Argon beats both on major benchmarks, making it a three-way contest again.

ITmedia reports that Google announced Gemini 4 Agron rather than Pro, initially available only to cyber defense organizations, with an output limit of 1 million tokens.

A short hands-on post says Gemini 4 now beats Astra and Fable on many tasks, without naming specific benchmarks.

The author shares output from a Gemini 4 checkpoint that competed on Arena as gemini-3.8-flash and calls the results very good.

Gemini 4 Argon reaches 53 on the Artificial Analysis Intelligence Index for fifth place, matching GPT-6 Astra, at $4/$20 per million tokens with a 50% discount.

A user ran a prompt on the Gemini 4 Argon checkpoint on Arena asking for as detailed an SVG image of Ana de Armas as possible.

Gemini 4 Argon ranks first on Mercor's APEX-Agents benchmark with 82.2% Pass@1, the first model to pass 80%, and leads in consulting, investment banking and corporate law.