Google May Be Secretly Testing Gemini 4 Pro That Outperforms GPT-6 Astra and Claude Fable 5.1

Google DeepMind's Unannounced Model Surfaces on Benchmarking Platform

A next-generation AI model that developers suspect is Google DeepMind's Gemini 4 Pro appeared on the Arena benchmarking platform disguised as "gemini-3.8-flash." The model, carrying the internal codename Argon, has not been officially announced by Google, but leaked benchmark results suggest it outperformed leading competitors including OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 across multiple evaluation metrics1

1

. Developers who accessed the model reported exceptional performance in AI-agent coding, reasoning tasks, and computer proficiency assessments. 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 proficiency1

1

. These results position the suspected Gemini 4 Pro ahead of established models in the competitive AI landscape, signaling Google's push to reclaim leadership in frontier AI development.

Advanced Technical Specifications and Pricing Strategy

The model's technical specifications reveal significant capabilities that distinguish it from current offerings. Developers reported that Gemini 4 Pro features a 10 million-token input limit and a 256,000-token output limit, substantially exceeding the 1 million-token context window of the production Gemini 3.8 Flash released on September 21

1

. The model also includes persistent cross-session memory, allowing it to maintain context across multiple interactions. Reported pricing stands at $2.25 per million input tokens and $11.25 per million output tokens, making it more cost-effective than competing frontier models while offering superior performance1

1

. This pricing strategy, combined with enhanced capabilities, positions Google to capture market share from competitors. The substantial gap between the production Gemini 3.8 Flash pricing of $0.75 per million input tokens and the suspected Gemini 4 Pro pricing has fueled speculation that Google used the "gemini-3.8-flash" label on the Arena benchmarking platform as a disguise for early testing.

Breakthrough Performance in Content Creation and Interactive Applications

Developers testing the model reported strong performance in SVG and 3D generation, expanding its utility beyond traditional language tasks. One developer documented that the model produced a full creative showcase website in just 14 minutes, while others generated complete interactive games from a single prompt1

1

. These capabilities in content creation and interactive applications demonstrate the model's versatility for digital design, game development, and professional training scenarios. Early testing on platforms highlighted Gemini 4 Pro's ability to create intricate vector graphics with animations and interactive elements, valuable for web design and user interface development2

2

. The model's proficiency in generating detailed 3D designs makes it relevant for architects, game developers, and product designers requiring precision in complex projects. Its capacity to produce immersive, interactive environments ranging from virtual worlds to high-fidelity flight simulators holds promise for education and entertainment applications.

Strategic Shift Following Gemini 3.5 Pro Cancellation

The suspected Gemini 4 Pro test follows significant changes in Google's model rollout strategy. 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"1

1

. Google instead released four successive Flash models since May, culminating in Gemini 3.8 Flash on September 2. This strategic pivot reflects Google DeepMind's focus on delivering meaningful performance improvements rather than incremental updates. Google DeepMind announced on July 21 that it had begun "our most ambitious pre-training run yet, for Gemini 4," with Alphabet CEO Sundar Pichai repeating that statement on the company's second-quarter earnings call the following day1

1

. Community estimates point to a possible public release around October 2026, though Google has not acknowledged the Arena testing or confirmed specifications, release timing, or whether the listing was part of internal evaluation1

1

2

2

.

Recursive Self-Improvement and Innovation Trajectory

Jasjeet Sekhon, Google DeepMind's chief strategy officer, stated at a Berkeley event that recursive self-improvement had become "a key component of the rationale behind massive AI investments"1

1

. Unverified claims suggest that Gemini 4 pre-training benefited from an RSI closed loop, potentially enabling the model to autonomously refine its algorithms for smarter, more efficient outputs over time. This capability represents a forward-thinking approach to AI development, positioning Gemini 4 Pro as a model that evolves alongside user needs. However, the innovation trajectory faces scrutiny as Google DeepMind works to address persistent challenges. Critics noted that earlier Google AI models occasionally produced simplistic or blocky visuals, raising questions about whether Gemini 4 Pro has fully resolved these shortcomings. The model must demonstrate consistent performance across varying conditions to secure widespread adoption in a market where GPT-6 Astra is recognized for deep contextual understanding, Claude Fable 5.1 excels in creative outputs, and Opus 5.2 is valued for efficiency and adaptability2

2

. Watch for official announcements regarding specifications, pricing confirmation, and performance validation as the October 2026 timeline approaches.