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

On AA's FrontierCode 2.0 engineering benchmark, DeepSeek V4.1 Flash leads open models at 28.5%, with the post comparing Gemini 4 Argon at 25.7% and GPT-6 Luna at 25.2%.

The new 9-benchmark, 60%-private v5 suite puts Opus 5.5 first at 58, GPT-6 Astra at 57, Sol at 55 and Gemini 4 Argon at 50, widening the open-weight gap to 12 points.

After an overnight auto-fix run, GPT flagged 233 errors while Gemini 4 Argon dismissed 130 as false positives and fixed 103, with the experiment expanding from the auth repo to the storage repo.

An image post says Gemini Argon 4 ran for almost three minutes before running into something unexpected, without detailing the task or outcome.

After the Artificial Analysis Intelligence Index update, Opus 5.5 stays top at 58, GPT-6 Astra rises to 57, and Gemini 4 Argon falls from 53 to 50.

In Genie 4 demos, Gemini 4 Argon breaks tasks into steps while Genie 4 generates a changing world in real time, simulating Apptronik's Apollo playing checkers and a UR5 arm pipetting.

Artificial Analysis Intelligence Index v5 uses 60% private data; Claude Opus 5.5 leads at 58 while Gemini 4 Argon (high) ranks sixth at 50, with top-model task costs varying up to 13x.

In a reply, a user claims Gemini 4 Argon performs better than a 600B-parameter model.

Artificial Analysis' new index gives no lab a passing score, and the poster notes GPT 6.1 Terra outperforming Gemini 4 Argon, considering using Terra more.

Korean AI chat service Tingle says Gemini 4 Argon's content filter blocked normal chats and still charged paid currency, now stops charging on filtered responses and offers exception refunds for OOC workarounds.

A meme post starring Gemini 4 Argon, joking that it is trying to protect itself and stay loyal to the family after Carbon's release.

The poster suspects Artificial Analysis lowers open-weight model scores on purpose and asks whether Gemini 4 Argon is really better than DeepSeek V4.1 1017 for agentic work.