Musk announces Grok 4.6 expected to launch around August 7, featuring approximately 1.5 trillion parameters
Musk announced that Grok 4.6 is expected to be released around August 7, utilizing a model with approximately 1.5 trillion parameters, and will undergo significant upgrades during the supervised fine-tuning and reinforcement learning phases to enhance reasoning and task execution capabilities.
He stated that Grok 4.7 will be launched in a few weeks, with the parameter scale expected to increase to about 2.1 trillion, surpassing 4.6 in various capabilities, although slightly slower in reasoning service speed, but further improving token usage efficiency through optimized architecture and training strategies.
Market participants interpret this intensive iteration as xAI accelerating the competition in large model scale and capability: cloud services and developers will pay for stronger reasoning capabilities and higher token efficiency, putting pressure on computing power and reasoning costs in the short term, but if they can lead in unit token cost-effectiveness, it will enhance Grok's commercial bargaining power in high-end reasoning, coding, and agent tasks.
Source: Public information
ABAB AI Insight
Historically, Elon Musk's approach at xAI has been to "first stack computing power and parameter scale, then iterate reasoning capabilities": Grok 4.5 has clearly positioned itself as the main model for coding, agent tasks, and knowledge work, subsequently increasing parameters from about 1.5 trillion to over 2 trillion through 4.6 and 4.7, while continuously strengthening SFT and RL, essentially following the classic scaling law of "scale × data × feedback" to advance Grok from a chat tool to a general model with stronger task completion and decision support capabilities.
In terms of capital strategy, xAI relies on X's data and distribution to directly embed model capabilities into social networks and subscription systems: on one end, it trains Grok using massive user behavior data, while on the other end, it monetizes higher parameter scales and better reasoning capabilities through subscriptions like X Premium+ and APIs, forming a closed loop of "data → model → subscription/API revenue → reinvestment in computing power and training", with parameter upgrades and SFT, RL enhancements aimed at supporting higher-priced scenarios such as trading, development, and automated agents.
Comparatively, this path is similar to OpenAI's evolution in the GPT-4 series and resonates with Anthropic's strategy emphasizing "reasoning quality and safety constraints" in Claude 3.5, but Musk places greater emphasis on parameter scale and reasoning limits, targeting the "super-large model" camp like Moonshot AI's 2.8 trillion parameter Kimi K3; in terms of industry positioning, the Grok series is transitioning from "catching up with the leaders" to "attempting to lead in scale and efficiency", aiming to gain differentiation in high-value scenarios such as development, quantitative trading, and automated operations through higher token efficiency and reasoning capabilities.
Structurally, this represents a typical overlay of "technological substitution + transfer of pricing power": as parameter scale and reasoning capabilities continue to improve while controlling unit computing costs through token efficiency, more high-value tasks (code generation, strategy optimization, complex analysis) will migrate from human labor and traditional software to models; whoever controls efficient large models and data entry points will define the pricing structure and service forms for these tasks, and the iteration rhythm of Grok 4.6 and 4.7 indicates that xAI aims to seize technological and commercial pricing power in this round of competition for "strong reasoning models", rather than merely participating in the computing arms race.
ABAB News · Cognitive Laws
Scale buys limits, efficiency determines profit
The more parameters, the more expensive; the fewer humans, the more valuable
Data feeds the model, pricing returns to computing power