New evidence from the Grok iOS app confirms that Grok 3.5 is on the verge of release. Configuration files now list “grok_3_5” alongside previous models like “grok_3_2” and the current public release, “grok_3_1.” The updated files also include a new model selector UI, letting users switch between Grok 2 and Grok 3, with space reserved for Grok 3.5.
According to TestingCatalog, the latest app strings refer to “Early access to Grok 3.5 and new features,” suggesting the model is already in staged rollout or internal testing. These strings are tied to a subscription upsell prompt, hinting that Grok 3.5 may launch first for SuperGrok users, the premium tier on X.
What Makes Grok 3.5 Different
What sets Grok 3.5 apart isn’t just scale but smarter training. It focuses on deep conceptual reasoning, allowing it to solve complex problems using fundamentals rather than surface-level patterns. Elon Musk described it as “figuring things out,” not just regurgitating internet data. This makes Grok 3.5 well-suited for fields where logic, creativity, and strategy are key.
Unlike ChatGPT or Gemini, which lean on live web data, Grok 3.5 operates with search independence. It can generate insights on its own, without copying or paraphrasing from existing content. That independence makes it useful for writing original research summaries, solving novel engineering problems, and making high-level business decisions.
Grok 3.5 also includes a major upgrade to its Think Mode. The original Big Brain Mode introduced in Grok 3 has evolved into Think Mode 2.0. This feature enables multi-step reasoning, error correction, and clear explanations of its logic. When you ask it to design a rocket engine or analyze a business plan, it doesn’t just answer. It walks you through the process, weighing options and citing relevant principles.
Launch Plans and Future Roadmap
Early benchmarks suggest Grok 3.5 outperforms GPT-4, Claude 3, and DeepSeek V3 in areas like advanced math, science, logic puzzles, and business case studies. Independent tests are still in progress, but early data looks strong.
The model runs on xAI’s Colossus supercomputer, reportedly equipped with over 200,000 NVIDIA H100 GPUs. That gives it more training power than OpenAI had for GPT-4, allowing for faster inference, better fine-tuning, and reduced latency.
For now, Grok 3.5 is in beta and accessible only to SuperGrok subscribers. Wider access is expected soon, including web and mobile apps, developer APIs, and integration with X for features like real-time chat and context-aware posts.
Elon Musk has also teased Grok 4, planned for mid to late 2025, with support for image, video, voice, and long-term memory. He claims it could reach AGI-level capabilities, though that remains to be seen.
Grok 3.5 isn’t without critics. It still hallucinates at times, its training costs raise questions about scalability, and its closed-source nature has drawn backlash from open-source advocates. Still, the model’s ambition and architecture signal a real push toward next-generation AI.