A model believed to be Google’s next flagship AI system, Gemini 4 Pro, has shown up on the Arena benchmarking platform under the name “gemini-3.8-flash.” Developer reports about the listing began circulating on September 17. The system carries the internal codename Argon. Google has not announced it.

Leaked results from Arena testing indicate the model tops OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 on coding, reasoning, and agentic work.

Reported Scores on Arena

Developers who ran the model shared these figures:

  • 88% on DeepSWE v1.1 for AI-agent coding tasks
  • 95.3% on Terminal-bench 2.1 for encoding capability
  • 86.8% on OSWorld-2.0 for computer proficiency

Each score is said to beat the corresponding numbers from Astra and Fable.

Testers also noted strong results in SVG and 3D generation. One developer reported building a full creative showcase website in 14 minutes. Others said the model produced complete interactive games from a single prompt.

Token Capacity and Pricing

The model is described as supporting a 10-million-token input limit and a 256,000-token output limit. It also includes persistent cross-session memory.

Reported pricing places it as the lowest-cost option among current frontier models: $2.25 per million input tokens and $11.25 per million output tokens.

Recent History of Google’s Gemini Releases

The possible Gemini 4 Pro checkpoint arrives after a stretch of internal shifts. The Wall Street Journal reported in late August that Gemini 3.5 Pro had been canceled. Internal candidates, according to that report, “weren’t sufficiently better than the Flash series.”

Google instead released four Flash models in succession starting in May. The latest, Gemini 3.8 Flash, arrived on September 2.

On July 21, Google DeepMind stated it had begun “our most ambitious pre-training run yet, for Gemini 4.” Alphabet CEO Sundar Pichai repeated the statement on the company’s Q2 earnings call the next day. Beyond those comments, Google has released no further details on Gemini 4 variants, timelines, or technical specifications.

What Google Has Not Confirmed

Google has not acknowledged the Arena tests or verified any information about Gemini 4 Pro.

Community estimates suggest a possible public release around October 2026, though no official schedule exists.

The production version of Gemini 3.8 Flash, released September 2, carries a 1-million-token context window and costs $0.75 per million input tokens. The appearance of a different model under the same public name on Arena has led some observers to conclude that Google is using the label as cover for early evaluation of Gemini 4 Pro.

At a recent Berkeley event, Google DeepMind chief strategy officer Jasjeet Sekhon remarked that recursive self-improvement has become “a key component of the rationale behind massive AI investments.” That statement has drawn fresh attention amid unverified claims that Gemini 4’s pre-training involved an RSI closed loop.