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

After using Gemini 3.8 Flash heavily for several days, a user says reasoning, consistency and output quality all dropped, and guesses it was downgraded ahead of Gemini 4 Pro.

A rumor says Gemini 4 Pro benchmarks beat both Claude and Astra; the blogger says he'd welcome it but isn't rushing to conclusions.

The author rounds up models and products that could land this week, including OpenAI DevDay, GPT-6.1 Astra, Claude Sonnet 5.5, Gemini 4 Pro and Nano Banana 2.1.

A screenshot said to be a Gemini 4 Pro benchmark is circulating; the poster says the numbers are hard to believe and calls on Google to confirm whether it's real.

A Japanese-speaking user says recent Gemini 4 Pro talk looks promising, praises 3.8 Flash for natural Japanese text, and hopes Gemini makes a comeback.

The author says that if the leak is true, Gemini 4 Pro beats Astra and Opus 5.5 by a ridiculous margin, and wonders what Google has been working on.

An English account rounds up leaked figures for Gemini 4 Pro: DeepSWE 88.7, Terminal-Bench 95.3, OSWorld 86.8, GDPval-AA 2064, 2M context and $2.25/$11.25 pricing, calling it a clear jump over Opus 5.5 and GPT-6 Astra but unconfirmed.

A purely speculative timeline lists Gemini 4 Pro in October 2026, Gemini 4 Flash in November, and Gemini 5 in September 2027, claimed to surpass Opus 5.5 and GPT-6 Astra.

A post claims Gemini 4 Pro, a 69T-parameter model insiders call ASI, went only to a few companies and was shared as open weights for frontier labs to distill, so it may never go public.

A circulating Gemini 4 Pro table lists DeepSWE 88.7, Terminal-Bench 95.3, OSWorld 86.8, GDPval 2064, a 2M context and $2.25/$11.25 pricing, beating Astra and Opus across the board.

MiaAI_lab says it is hard to believe the claim that Gemini 4 Pro outperforms Opus 5.5, but would be happy to be proven wrong.

The author mocks an unverified Gemini 4 Pro benchmark sheet that surfaced the night before OpenAI DevDay, claiming it beats Opus 5.5 and Astra with a 2M context at a fraction of the price.