A model developers suspect is Google’s next flagship AI system appeared on the Arena benchmarking platform under the name “gemini-3.8-flash.”. The model carries the internal codename Argon and has not been officially announced by Google, but leaked benchmark results shared by developers suggest it outperformed OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 in coding, reasoning and agentic tasks..
Developers who shared the benchmark data said the model scored 88% on DeepSWE v1.1 for AI-agent coding tasks, 95.3% on Terminal-bench 2.1 for encoding capability, and 86.8% on OSWorld-2.0 for computer proficiency.. The same developers said the model has a 10 million-token input limit, a 256,000-token output limit and persistent cross-session memory..
Developers who tested the model also reported strong performance in SVG and 3D generation. One developer said it produced a full creative showcase website in 14 minutes, while others said it generated complete interactive games from a single prompt.. Reported pricing for the model was $2.25 per million input tokens and $11.25 per million output tokens, which would make it cheaper than other leading frontier models named in the leaked comparisons..
The suspected Gemini 4 Pro test follows a period of changes in Google’s model rollout. The Wall Street Journal reported in late August that Gemini 3.5 Pro had been scrapped because internal candidates “weren’t sufficiently better than the Flash series.”. Google instead released four successive Flash models since May, culminating in Gemini 3.8 Flash on Sept.
2.. Google DeepMind said on July 21 that it had begun “our most ambitious pre-training run yet, for Gemini 4,” and Alphabet CEO Sundar Pichai repeated that statement on the company’s second-quarter earnings call the next day.. Google has not acknowledged the Arena testing or confirmed details on Gemini 4 Pro, including its specifications, release timing or whether the Arena listing was part of internal evaluation..
Community estimates have pointed to a possible public release around October 2026, though no official timeline has been announced.. The production version of Gemini 3.8 Flash, released on Sept. 2, has a 1 million-token context window and pricing of $0.75 per million input tokens, a gap that has fueled speculation that Google used the “gemini-3.8-flash” label on Arena as a disguise for early Gemini 4 Pro testing..
Jasjeet Sekhon, Google DeepMind’s chief strategy officer, said at a Berkeley event that recursive self-improvement had become “a key component of the rationale behind massive AI investments,” amid unverified claims that Gemini 4 pre-training benefited from an RSI closed loop.. Featured image credit.
Tags: FeaturedgeminiGoogle
