Gemini 3.5 Pro Leak: Google’s Next Flagship AI Drops Today to Challenge Claude Fable 5

After weeks of vague posts, mysterious benchmark sightings, and intense developer speculation, Google DeepMind’s worst-kept secret is out: Gemini 3.5 Pro is expected to release publicly today.

The build-up to this release has been a rollercoaster. While early developer checkpoints felt lackluster, recent teasers from Logan Kilpatrick and several DeepMind engineers reignited excitement. Leaked reports—covered in detail across NokiaPowerUser’s Gemini coverage—indicate that Google spent the last few months quietly refining this release through extensive partner testing and stealth deployments.

Key Takeaways from the Gemini 3.5 Pro Leak

• Public Rollout Today: Following months of closed testing, A/B experiments, and fine-tuning, the public launch is widely expected today.

• Beating Fable 5 on Value: Benchmark leaks suggest Gemini 3.5 Pro outperforms Anthropic’s Claude Fable 5 on visual code generation, spatial reasoning, and agentic tasks—all at a significantly lower cost per token.

• Tested in Stealth Mode: Power users discovered the model running inside Google Antigravity and blind evaluation arenas under the temporary Gemini 3.1 Pro label.

• Refined Performance: Google focused heavily on post-training alignment to eliminate hallucination issues present in early test builds.

From Rough Checkpoints to Fine-Tuned Powerhouse

When early internal builds of Gemini 3.5 Pro circulated earlier this year, initial feedback was mixed. The model struggled with consistency on complex reasoning tasks and showed context degradation across large code repositories.

Rather than rushing out an incomplete model, DeepMind pulled back to overhaul its training pipelines. According to NokiaPowerUser’s deep-dive leak report, the engineering team focused on three key areas:

1. Enhanced “Deep Thinking” Reasoning: Multi-pass verification steps were added to eliminate logic errors during complex coding, math, and architectural queries.

2. 2 Million-Token Context Window: The massive context window was optimized to maintain high retrieval accuracy across deep, multi-file codebases.

3. Optimized Agent Workflows: The model’s integration with IDEs, command-line environments, and external tool APIs was heavily refined.

The Antigravity Secret: Stealth Testing in Plain Sight

One of the most interesting aspects of this launch is how Google battle-tested the architecture under real-world conditions. Rather than relying solely on synthetic benchmarks, DeepMind deployed early iterations directly into live production pipelines.

Developers on Google’s Antigravity platform noticed sudden jumps in code quality and design logic while using endpoints labeled Gemini 3.1 Pro. Telemetry later revealed that backend routing was secretly sending those inference calls to the unreleased Gemini 3.5 Pro model.

This shadow rollout provided Google with crucial real-world usage data across complex debugging sessions, UI designs, and agentic tasks before competitors realized what was happening.

Target in Sight: Beating Claude Fable 5 at a Lower Price

Anthropic’s Claude Fable 5 recently set a high benchmark for repository-scale engineering, but its high token pricing remains a major barrier for automated workflows.

As detailed in NokiaPowerUser’s benchmark comparison, Gemini 3.5 Pro directly challenges Fable 5 by matching its reasoning capabilities while offering superior performance in visual code generation and frontend design.

By leveraging Google’s custom TPU infrastructure, Gemini 3.5 Pro delivers top-tier intelligence at a price point designed for high-volume developer usage.

What to Watch for at Launch

As official announcements roll out, monitor Google AI Studio, Vertex AI, and the Gemini API for direct access. For ongoing updates, live benchmark teardowns, and official confirmation, follow the NokiaPowerUser Gemini Hub.