Gemini 4 Pro Leaks as 'gemini-3.8-flash' on Arena, Beats Astra and Fable Across Benchmarks

Over the past few days, a new model named gemini-3.8-flash quietly appeared on the LMSYS Chatbot Arena. Developers and AI researchers who tested it quickly concluded this is almost certainly Google DeepMind's next-generation flagship Gemini 4 Pro , disguised under a Flash-series name. The previous Gemini 3.8 Flash launched in early September, so a second model bearing the exact same name is highly unusual.

Benchmark Results Show Broad Superiority

A widely circulated benchmark screenshot indicates Gemini 4 Pro achieves comprehensive leads over GPT-6 Astra and Claude Fable 5.1:

DeepSWE v1.1 (AI agent coding tasks): 88% — nearly 2 points above Astra.

GDPval-AA v2 (real-world knowledge work): 2064 Elo — the only model above 2000.

Terminal-bench 2.1 (terminal coding): 95.3% — highest score.

OSWorld-2.0 (computer-use capability): 86.8% — beats both Astra and Fable.

If these numbers hold, Gemini 4 Pro becomes the current “strongest model on earth” across agentic coding, reasoning, and computer-control tasks.

Pricing and Technical Specifications

Pricing is reported at $2.25 per million input tokens and $11.25 per million output tokens , making it the most cost-effective option among the “big three” frontier models. Researcher Qwinah discovered the backend exposes a 10 million token input limit , 256k token output limit , permanent cross-session memory , and built-in web access without requiring an API .

Hands-On Demos Reveal Step-Change in Creative & 3D Capabilities

Developers shared extensive real-world tests:

UI/UX Design : In 14 minutes, Gemini 4 Pro built a “creative showcase” page with a “scroll-as-brush-stroke” interaction where lines deepen smoothly on scroll.

Cyberpunk/Retro-Future Worksite : Combined 3D grid, data dashboards, and an interactive 3D model on the hero screen.

Classic “Pelican on a Bicycle” SVG/3D Test : Output includes color theming, day/night toggle, headlights, anatomical annotations, and cadence control — completion quality matches top-tier models.

Pixel-Art 3D Pagoda : Immersive, explorable environment.

Airbus H145 Helicopter : Full 3D model generated in ~10 minutes.

3D Flight Simulator : Visual fidelity far exceeds the public Gemini 3.8 Flash, confirming the Arena model is a different, more capable system.

Game Generation : Produces playable games with complete interaction logic and polished visuals — e.g., a Minecraft-style voxel builder and a modern 3D kart racer — in minutes.

Developer Pankaj Kumar summarized: “Speed is very fast; SVG and 3D generation noticeably improved; complex requirements understood in a single prompt.”

Recursive Self-Improvement (RSI) as the Strategic Driver

The article connects the leap to Recursive Self-Improvement (RSI) . Google DeepMind Chief Strategist Jasjeet Sekhon recently stated at a Berkeley event that RSI has become a key pillar of AI investment logic. If the RSI loop sustains, model improvement could accelerate beyond exponential curves. Rumors suggest Gemini 4 finished pre-training early precisely because DeepMind closed the RSI loop during training. On 14 October, Google publicly released the Dream-RSI research, which explores how agents can continuously refine their own search strategies from experience — confirming DeepMind’s public commitment to RSI.

Competitive Context: The “Big Three” Race

OpenAI fields Astra ; Anthropic fields Fable 5.1 . Both have pushed competition beyond chat into long-horizon tasks, agents, code, reasoning, and autonomous research. Google’s Gemini 3.5 Pro was cancelled (WSJ reported it “couldn’t even beat Flash”), so Gemini 4 Pro must re-prove Google’s frontier leadership to defend its $4.23 T market cap. The anonymous Arena debut appears to be a live checkpoint test before a formal launch.

Reference

https://x.com/HarshithLucky3/status/2100524699780600136