Google made an unusual choice: rather than ship a nearly finished AI model, the company discarded months of development work and started over from scratch. The decision to restart pre-training, the most expensive phase of building a frontier AI model, suggests Google's engineers found structural problems that fine-tuning alone couldn't fix. The rebuilt Gemini 3.5 Pro is expected to arrive around mid-July, competing directly with newly launched models from OpenAI and other labs.

What Went Wrong With the Original Version?

According to third-party reporting citing unnamed internal sources, the initial version of Gemini 3.5 Pro exhibited two critical failure modes that made it unsuitable for release. The model struggled with two specific tasks that matter enormously for real-world AI applications:

• SVG Layout Generation: The model could not maintain structural consistency when generating complex, multi-layered SVG scene layouts, a capability increasingly important for design and visualization tasks.
• Recursive Tool-Calling: The model broke down under complex, recursive tool-calling environments, where an AI agent calls one tool, uses the result to call another, and repeats this chain across dozens of sequential decisions.

These are not minor technical glitches. Recursive tool-call stability is the defining requirement for agentic coding models, which is the exact use case Google has staked the entire 3.5 generation on. If Gemini 3.5 Pro had shipped with these limitations, it would have represented a regression compared to the faster Gemini 3.5 Flash model, which already demonstrated strong performance on these benchmarks.

Why Did Google Choose to Rebuild Instead of Shipping?

The decision to scrap a near-complete model and restart pre-training reveals how seriously Google viewed the gap between what it had built and what competitors had already shipped. Pre-training is the initial training run on vast datasets that gives a model its fundamental capabilities, and restarting this phase represents an enormous investment of computing resources and time.

Google's own Gemini 3.5 Flash model already demonstrated that the company could build a model that excels at these tasks. Flash achieved 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas, outscoring the older Gemini 3.1 Pro on both benchmarks. A Pro model that performed worse on these same tasks would have been an embarrassment, not an upgrade. That competitive reality likely drove the decision to rebuild.

What Should Developers Expect From the Rebuilt Model?

Third-party sources report that the rebuilt Gemini 3.5 Pro will arrive with three capabilities designed to differentiate it from the Flash version. However, it is important to note that Google has not officially confirmed any of these specifications as of mid-July:

• Context Window: A reported 2-million-token context window, double that of Gemini 3.5 Flash's 1-million-token limit, allowing the model to process roughly 1.5 million words at once.
• Deep Think Reasoning Layer: A multi-step logical problem-solving capability that has already demonstrated strong performance on reasoning benchmarks, including an 84.6% score on ARC-AGI-2.
• Autonomous Workflow Capabilities: The ability to manage multi-file coding tasks and tool chains with minimal human direction, addressing the exact recursive tool-calling weakness that triggered the rebuild.

The 2-million-token context window figure deserves scrutiny. While widely reported, independent research published in 2025 by Chroma tested 18 frontier models and found that every single one degrades as context grows. The study found that models typically break down 30 to 40 percent before their claimed limit. Gemini 3.1 Pro already illustrates this pattern: it ships with a 1-million-token window, but independent benchmarking shows reasoning quality falls sharply past 128,000 tokens, dropping to roughly 26% at the full 1-million-token range.

How Does This Rebuild Affect the Competitive Landscape?

The timing of Gemini 3.5 Pro's arrival matters significantly. The same week, GPT-5.6 Sol launched publicly on July 9, and Grok 4.5 opened to the public the same day. DeepSeek's V4 family is separately targeting mid-July for its stable release. Three of the most-watched model events of 2026 converged on the same week, which means any further delay for Gemini 3.5 Pro would carry greater competitive consequence than the earlier June slip did.

If the July 17 date holds and the rebuilt model delivers on its reported capabilities, Gemini 3.5 Pro enters a market where Google has a narrow technical window to establish positioning before developers stabilize their stacks around the models now available. If the date slips again, the story becomes one of a lab that correctly diagnosed a problem but underestimated how long the fix would take.

How to Evaluate Vision Language Models When They Launch

When new vision language models arrive, developers and enterprises should look beyond headline specifications to assess real-world performance. Here are key evaluation steps:

• Test Long-Context Stability: Run structured retrieval benchmarks across the full claimed context window, not just at the advertised ceiling, to verify that reasoning quality holds throughout the range.
• Verify Recursive Tool-Calling Performance: Test multi-step agent chains where the model calls one tool, uses the result to call another, and repeats this pattern across dozens of decisions to ensure stability under complex workflows.
• Benchmark Against Specific Use Cases: Evaluate the model on tasks directly relevant to your application, such as SVG generation, code modification, or document analysis, rather than relying solely on general-purpose benchmarks.
• Compare Independent Evaluations: Wait for third-party benchmarking results from research organizations, not just vendor-provided numbers, to get an unbiased assessment of capabilities.

The Gemini 3.5 Pro rebuild underscores a broader reality in AI development: shipping a weaker model on schedule is often worse than delaying a stronger one. Google's decision to restart pre-training, despite the enormous cost, reflects the competitive intensity of the vision language model market and the high stakes of getting flagship models right.