Elon Musk says xAI plans to release two new frontier models in August, extending an unusually fast product cycle only weeks after the company introduced Grok 4.5. In a post on X, Musk said Grok 4.6 is expected to launch around August 7 as a 1.5 trillion-parameter model with significantly improved supervised fine-tuning and reinforcement learning. Grok 4.7, a larger 2.1 trillion-parameter system, is scheduled to follow several weeks later.
Interesting.
Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL.
Grok 4.7 will be the 2.1T model released a few weeks later. This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better… https://t.co/dJjXVnFwoA
— Elon Musk (@elonmusk) July 28, 2026
The announcement provides a release window and headline model sizes, but few other details. xAI has not published benchmark results, pricing, context length, architecture information, safety documentation, or confirmed whether both models will launch simultaneously across X, grok.com, mobile apps, and the company’s API.
The two releases would follow Grok 4.5, which xAI introduced earlier in July as a model focused on coding, agentic tasks, and professional knowledge work. The current model is priced at $2 per million input tokens and $6 per million output tokens, with xAI emphasizing speed and token efficiency rather than claiming an outright lead on every intelligence benchmark.
If Musk’s schedule holds, xAI will have moved from Grok 4.5 to Grok 4.7 in roughly six weeks. That cadence is significantly faster than the major-version release schedules traditionally used by frontier AI laboratories and suggests the company is treating model development as a continuous sequence rather than a series of isolated launches.
Grok 4.6 Will Focus on Post-Training Improvements
Musk’s description of Grok 4.6 places particular emphasis on supervised fine-tuning and reinforcement learning, the stages used to shape a base model after its initial training.
Supervised fine-tuning exposes a model to carefully selected examples of desired behavior, helping it produce more useful, accurate, and consistently formatted responses. Reinforcement learning then uses reward signals to improve how the system reasons, follows instructions, uses tools, and handles difficult tasks.
Those improvements can sometimes matter more to users than a larger parameter count. A model may gain better coding performance, stronger instruction following, or more reliable tool use without a dramatic change in its underlying size, provided the post-training process is sufficiently effective.
xAI has not explained what data, evaluators, reward models, or agent environments were used for Grok 4.6. It also has not said whether the model will introduce changes to reasoning, multimodal capabilities, coding, web research, or computer use.
The official Grok 4.5 release focused heavily on real-world coding economics. xAI said that model used fewer output tokens than several competitors and was trained with data from Cursor, giving it exposure to how developers work across real codebases.
Grok 4.6 may build on that foundation, but Musk’s brief announcement does not establish which capabilities will improve or by how much. Until xAI publishes a model card and independent evaluators test the system, the size and post-training claims should be treated as a roadmap rather than evidence of performance.
The company’s rapid release schedule follows its rebrand as SpaceXAI after the integration of Musk’s AI and space businesses. The combined structure gives model development access to the Colossus computing clusters and a wider infrastructure strategy spanning data centers, chips, networking, and potentially orbital compute.
Grok 4.7 Raises the Parameter Count to 2.1 Trillion
Grok 4.7 is expected to raise the reported parameter count to 2.1 trillion. Musk did not say whether the model will use a dense architecture, in which all parameters participate in each response, or a sparse Mixture-of-Experts design that activates only part of the network at a time.
That missing information makes direct comparisons difficult. Moonshot AI recently released Kimi K3, which contains 2.8 trillion total parameters but activates about 104 billion for each token through a sparse expert architecture. Its headline size is therefore not directly comparable with a dense model of the same scale.
Parameter count can indicate capacity, but it does not by itself determine intelligence, speed, accuracy, or cost. Training data, architecture, post-training, inference software, context management, and tool integration can all matter as much as the total number of weights.
The larger model nevertheless signals that xAI is continuing to scale aggressively. Musk has previously described plans for increasingly large training runs on the Colossus infrastructure, which the company has expanded with hundreds of thousands of advanced accelerators.
According to a report in the Economic Times, the announcement reflects xAI’s effort to maintain a rapid-fire release cycle as it competes with OpenAI, Anthropic, Google, and a growing group of Chinese model developers.
That strategy carries advantages and risks. Frequent releases can move research improvements into products quickly and provide more feedback from real users. They can also make it harder for enterprise customers to evaluate models, update prompts, complete safety reviews, and decide which version should become a stable production standard.
xAI has not said how long Grok 4.5 will remain available after the new models arrive or whether 4.6 and 4.7 will be offered at different prices. It also remains unclear whether Grok 4.7 will replace 4.6 or serve as a higher-capability option for more demanding workloads.
The distinction could matter commercially. A 1.5 trillion-parameter model optimized for speed and efficiency may be more attractive for high-volume coding or agentic tasks, while a 2.1 trillion-parameter model could target difficult reasoning jobs where customers accept higher latency and cost.
For now, Musk’s announcement is best understood as a timetable rather than a full product launch. Grok 4.6 is planned for early August, Grok 4.7 is expected several weeks later, and xAI is promising larger models and stronger post-training at a pace few competitors have attempted.
The real test will come when xAI publishes technical documentation and the models reach users. Their significance will depend not on how quickly they arrive or how many parameters they contain, but on whether they deliver measurable gains in reliability, capability, and cost over Grok 4.5.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
