Back to blog

Breaking: Gemini 4 Pro Leaked Onto the Arena — It Crushes Astra and Fable

A Gemini 4 Pro checkpoint disguised as gemini-3.8-flash slipped onto LMSYS Arena: DeepSWE 88.x% at No.1, the only model above 2000 Elo on GDPval-AA v2, Terminal-bench 95.3%, OSWorld-2.0 86.8% — at just $2.25/M input tokens. Plus front-end/3D/game tests and three China access routes.

Sep 18, 2026Maynor
Breaking: Gemini 4 Pro Leaked Onto the Arena — It Crushes Astra and Fable

Breaking: Gemini 4 Pro Leaked Onto the Arena — It Crushes Astra and Fable

Talk about hiding your power level.

This time, Google has finally thrown its deepest nuclear card onto the table.

Over the last couple of days, a mysterious new model quietly appeared on the global LLM arena, LMSYS Arena, wearing the alias gemini-3.8-flash.

But after only a few rounds of sparring, top AI figures and developers alike gasped:

This is no ordinary Flash build — it is very likely Google DeepMind's next-generation ultra-flagship, six months in the making: Gemini 4 Pro!

Arenas Tracker: gemini-3.8-flash variant spotted Caption: The LMSYS Arena tracker caught the mysterious gemini-3.8-flash variant — its real-world performance badly contradicts a lightweight label.

And then there's the benchmark ladder chart now circulating everywhere — a genuine dimensional strike.

Gemini 4 Pro is running rampant — on core metrics like coding agents, deep reasoning, and general tasks, it plainly crushes the rumored OpenAI GPT-6 Astra and Anthropic's Claude Fable 5.1!

Benchmark comparison: Gemini 4 Pro dominates Astra / Fable Caption: The leaked benchmark chart — Gemini 4 Pro leads across the board on agent and terminal-interaction dimensions.

Developer Vidhi says he was outright stunned after a deep session.

With nothing but the classic "generate a cat SVG" prompt, the gap is obvious: GPT-6 Astra produced a mangled, cat-adjacent blob, while Gemini 4 Pro delivered vector art that's structurally elegant and aesthetically on point — a generational lead.

Vidhi: discussion on Gemini 4 crushing GPT-6 Astra max Caption: Developer Vidhi praising Gemini 4 as comprehensively superior to GPT-6 Astra Max.

Cat SVG comparison: Gemini 4 Pro vs GPT-6 Astra vs Gemini 3.8 Flash Caption: Live-rendered cat SVG comparison — Gemini 4 Pro is markedly better on composition control and geometric fidelity.

Just as industry rumors suggested: Google got the RSI (Recursive Self-Improvement) flywheel spinning internally first, which is how Gemini 4 Pro can go head-to-head with Astra and Fable in this posture.

The long-dormant AI empire's giant has finally made a real comeback.


1. Anonymous Skin, Flagship Benchmark Slaughter

The AI world hasn't been this collectively electric in a long time.

Google's last major version bump, the Gemini 1.5/3.1 series, was months ago. The transitional version planned for summer slipped repeatedly or was scrapped entirely — the Wall Street Journal reported the core reason: the mid-cycle iteration didn't improve as much as expected.

To win this comeback, Google skipped the transitional state entirely and pushed all its compute onto Gemini 4.

Now, the Gemini 4 Pro checkpoint wearing a gemini-3.8-flash sheepskin has shown up — and its measured performance has stunned everyone.

Gemini 4 Pro brand banner Caption: The leaked Gemini 4 Pro official banner and its core positioning.

Core Benchmark Performance at a Glance

From the hard benchmark data leaking out, Gemini 4 Pro has delivered a dominant report card:

Benchmark Core scenario Gemini 4 Pro Competitive edge
DeepSWE v1.1 End-to-end agentic software engineering 88.x% (No. 1) ~2% ahead of Astra; industry-first
GDPval-AA v2 Real-world complex knowledge work 2064 Elo The only model above 2000 Elo in the field
Terminal-bench 2.1 CLI commands and low-level system ops 95.3% Extremely high-precision system-level interaction
OSWorld-2.0 Computer use 86.8% Beats Astra and Fable across the board; the ceiling for agent operation

Beyond raw performance, on the API cost developers care about most, Gemini 4 Pro also pulled out a price-performance killer:

  • Input price: just $2.25 / million tokens
  • Output price: just $11.25 / million tokens
  • Versus Astra and Fable, it smashes straight through the cost floor of ultra-flagship models!

yara David: gemini-3.8-flash on Arena with benchmarks / pelican animation collage Caption: Community developers reverse-testing the anonymous model across dimensions and confirming a next-gen architecture.

AI researcher Qwinah dug into the backend and pulled out jaw-dropping system parameters:

  • Massive context: 10 million token input limit + 256K token output limit
  • Native capabilities: permanent cross-session memory and low-level native real-time web access without an API.

Leaked DeepMind-style internal terminal: Flags and Reasoning Trace Caption: Leaked internal terminal logs exposing the long-context flag and deep reasoning trace.


2. Real-World Tests Explode: Front-End, 3D, and Interaction All Leap

Benchmarks are just numbers. Actual generation is the real test of a model.

Today's News: Gemini 4 Pro checkpoint tops Arena.ai Caption: Overseas tech-trending coverage — the Gemini 4 Pro checkpoint tops Arena.ai by a generational margin.

1. Front-end UI and interaction design: a qualitative shift in taste

Developer Bee exclaimed after trying it: "The design ability Gemini 4 Pro shows is unreal…"

In just 14 minutes, the model autonomously built a full high-end creative showcase site with a pencil-graphite texture. Even better, it natively wove in a "Scroll-to-Draw" design philosophy: as you scroll down, the strokes naturally deepen — the dynamic presentation holds up against a top-tier senior front-end designer.

Bee: Gemini 4 Pro builds a polished feature site in 14 minutes Caption: A high-aesthetic, interactive front-end project produced from a single prompt.

Faced with cyberpunk and retro-futurist site requirements, the model directly integrated 3D spatial grids, live data-visualization dashboards, and interactive WebGL models — completely shedding the "template feel" of earlier model-generated front ends.

2. 3D and complex dynamic vector (SVG): detail maxed out

On the notorious "pelican riding a bicycle" stress test, Gemini 4 Pro's output amazed the room.

Inverse kinematics (IK) solving, cadence dynamics animation, day/night lighting systems, bicycle mechanism annotations… the logic is extremely rigorous.

Pelican riding a bicycle SVG IK animation (Le Pélican Cycliste) Animated demo: the pelican-cycling animation with complex IK dynamics generated by Gemini 4 Pro.

GPT-6 Astra vs Gemini 4 Pro pelican cycling comparison (watermark blurred) Animated comparison: GPT-6 Astra vs Gemini 4 Pro on the same prompt — dynamics and physical plausibility.

Whether it's a meticulously crafted pixel-art 3D pagoda or a high-precision Airbus H145 helicopter 3D renderer assembled accurately in just 10 minutes, 4 Pro handles both effortlessly.

Pixel-art 3D pagoda, Gojū-no-tō Garden animation Animated demo: pixel-art 3D garden interactive showcase — excellent lighting and depth layering.

Airbus H145 helicopter 3D model configuration UI animation Animated demo: high-fidelity Airbus helicopter 3D rendering with a configurable web interface.

In a 3D flight-sim test, Gemini 4 Pro's physics-engine scheduling and frame rate left the older Flash model far behind.

Sh AI / Tim Jayas: discussing Gemini 4 Pro vs 3.8 Flash visual differences Caption: The community analyzing the gap between the two generations in rendering pipeline and spatial understanding.

Gemini 3.5 Flash vs Gemini 4 Pro 3D flight sim comparison (watermark blurred) Animated comparison: real-time 3D flight sim rendering — Gemini 4 Pro's smoothness and lighting are a generational improvement.

3. End-to-end game building: one prompt to a playable prototype

On single-prompt game generation — from recreating Minecraft's block-building logic to writing a 3D go-kart racing game with modern drift physics — Gemini 4 Pro shows terrifying code integration and a runtime error self-correction loop.


3. Google's Real Strategic Card: RSI (Recursive Self-Improvement)

Behind this qualitative change, it isn't just tens of thousands more TPUs. Google's ultimate bet is RSI — Recursive Self-Improvement.

As Google DeepMind's Chief Strategy Officer Jasjeet Sekhon put it: the next super-curve of the AI industry depends on whether models can recursively improve themselves.

Traditional pretraining is linear feeding. Once AI can close the loop across synthetic data construction, code reflection, and RL verification, the self-iteration speed jumps beyond exponential.

lyra: congratulations to GoogleDeepMind Caption: Industry insiders congratulating DeepMind on closing the RSI flywheel.

Combined with Google's newly published Dream-RSI architecture — letting agents continuously optimize search strategies through autonomous exploration — the industry widely speculates: Gemini 4 finished ahead of schedule because its core engine was "AI-directed AI self-iteration"!


4. The Top Table Is Being Reshaped: The "Big Three" Shift

Mehdi: summary of the Gemini 4 Pro leak aimed at Astra Caption: An analyst summarizing the core traits of the leaked Gemini 4 Pro, aimed squarely at OpenAI's Astra.

The war among top models has long since left the beginner tier of "daily Q&A."

  • OpenAI holds Astra, pushing super-agents and system integration;
  • Anthropic leans on Claude Fable 5.1, going deep on code and long scientific workflows;
  • Google, with Gemini 4 Pro, is launching a counteroffensive of full modality, ultra-long memory, and the RSI closed loop!

5. Three Ways to Use Gemini Efficiently in China (No VPN Needed)

Network conditions make reaching Gemini's official site a hassle from China. Here are three stable direct-access routes — pick what fits:

Channel Comparison

Option Platform Positioning & best for Core advantage
Option 1: API relay https://tryallapi.com/ Developers, enterprise integration, batch calls Direct from China, WeChat / Alipay supported, pay-per-token across the full model lineup
Option 2: Mirror chat site https://trygpt.asia/list/#/home Everyday personal writing, ready-to-use, deep Q&A No VPN, UI matches the official one, works out of the box
Option 3: Multi-model hub http://trymodel.chat/ Head-to-head across Gemini, ChatGPT, Claude All three majors in one place, no switching accounts

Detailed Access Guide

  • Route 1: API relay service (for development / system integration)

    • Link: https://tryallapi.com/
    • Positioning: A professional domestic API aggregation relay (MomoAPI) supporting pay-per-use calls to the latest models including Gemini.
    • Key features: In-country servers for direct connection, very low latency, domestic payment methods, standardized interfaces for individual and enterprise workloads.
  • Route 2: High-fidelity web mirror (for everyday chat and office work)

    • Link: https://trygpt.asia/list/#/home
    • Positioning: A web mirror reachable directly from China.
    • Key features: Usable with no VPN, extremely high UI fidelity, well suited to daily code debugging, paper reading, and multilingual translation.
  • Route 3: Multi-model hub (for one-stop comparison)

    • Link: http://trymodel.chat/
    • Positioning: A one-stop AI model workbench.
    • Key features: Integrates Gemini, ChatGPT, Claude, Grok and other top models — run the same prompt across models on one screen, no tedious account switching.

One-line recommendation: for production code and business workloads, pick Option 1; for light office chat, pick Option 2; for multi-model shootouts, pick Option 3.