Elon Musk says SpaceXAI plans to release its next flagship AI model, Grok 4.6, around 7 August, with a more powerful Grok 4.7 scheduled to arrive only weeks later as competition in the frontier model race intensifies.
In a post on X on 28 July, Musk said Grok 4.6 is being built as a 1.5 trillion‐parameter model with upgraded supervised fine‐tuning and reinforcement learning aimed at improving response quality and instruction-following.
He added that Grok 4.7 is expected to launch “a few weeks later” as a 2.1 trillion‐parameter system, though he did not give a firm date or reveal how either new model will be priced or made available to users.
Describing the step up from 4.6, Musk wrote: “This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better token efficiency.”
Both timeframes are targets rather than fixed appointments, with Musk acknowledging that schedules for training, testing and deployment can still shift before the models are ready for general use.
Push to build on Grok 4.5 momentum
The announcements follow the release earlier in July of Grok 4.5, which SpaceXAI positioned as a model geared towards coding, agent-style workflows and knowledge-intensive tasks.
SpaceXAI currently charges $2 per million input tokens and $6 per million output tokens for Grok 4.5, a price point that has drawn attention in recent independent benchmarking.
Musk outlined his roadmap for 4.6 and 4.7 after Vercel chief executive Guillermo Rauch publicly assessed Grok 4.5 as the best value cybersecurity model in Vercel’s latest tests when performance was weighed against cost.
According to Rauch’s analysis, Grok 4.5 came in at 10 times cheaper than OpenAI’s GPT‐5.6 Sol, 5.7 times cheaper than Anthropic’s Opus 5 and 2.2 times cheaper than Moonshot’s Kimi K3, while delivering results close to Kimi’s performance.
Even so, Rauch still ranked Sol as the strongest overall frontier model, placing it ahead of Opus 5. His conclusions are based on Vercel’s own methodology and do not amount to a definitive league table across all types of AI tasks.
SpaceXAI has not yet indicated whether Grok 4.6 and Grok 4.7 will keep Grok 4.5’s pricing structure. Any shift in service latency, token efficiency and API costs will determine whether the newer systems can sustain a similar price‐performance advantage.
Chinese competition and open‐weight pressure
The Grok upgrade plan is emerging as Chinese developers increase pressure on US firms at the top end of the AI market.
On 27 July, Moonshot AI released the full weights for its Kimi K3 model, allowing third parties to download, adapt and run the 2.8 trillion‐parameter system on their own infrastructure. Kimi K3 includes native vision capabilities and a one‐million‐token context window, and its open‐weight approach contrasts with the largely closed, API-based distribution used for many US frontier models.
SpaceXAI, OpenAI and Anthropic are already competing directly in areas such as coding support, cybersecurity, complex reasoning and the cost of delivering each task. Grok 4.7’s larger architecture may enhance its capabilities in some of these domains, though Musk has acknowledged that it will respond more slowly than Grok 4.6.
Parameter counts alone are not seen as a decisive measure of quality. Overall performance also depends on training data, model architecture, post‐training techniques, inference infrastructure and how efficiently a system uses tokens.
Federal scrutiny and safety framework
Alongside the technical race, leading US developers are working under growing federal scrutiny.
OpenAI and Anthropic have already agreed to provide US officials with early access to unreleased frontier models for safety evaluations. Those deals helped form the basis of a wider voluntary federal review framework, which now also covers other major players including SpaceXAI, Google DeepMind and Microsoft.
President Donald Trump formalised that framework via an executive order in June. Under the arrangement, participating companies can give the federal government access to covered models for up to 30 days before releasing them to trusted partners. The order specifies that the scheme does not amount to a compulsory licensing or pre‐approval regime.
However, there is no consensus in the industry over how Washington should approach open‐weight systems. In an open letter dated 24 July, Nvidia, Microsoft, Meta, OpenAI, Palantir and other organisations urged US lawmakers not to impose sweeping restrictions on open models, arguing they promote competition, lower deployment costs and enable developers to operate and adapt AI on their own hardware.
Moonshot allegations fuel policy dispute
That argument has taken on new urgency since Moonshot AI made Kimi K3’s 2.8 trillion‐parameter weights public.
Trump administration officials have accused Moonshot of training Kimi K3 using outputs from Anthropic’s Fable model via large‐scale distillation, characterising the alleged practice as an attempt to acquire proprietary US technology. White House technology adviser Michael Kratsios has raised those concerns, while Treasury Secretary Scott Bessent has said the administration is weighing potential sanctions and the possibility of placing Moonshot on a trade blacklist.
Moonshot has not publicly accepted the allegations. Nonetheless, the row has placed Kimi K3 at the centre of a broader policy battle over whether the US should focus on targeted measures addressing specific security and intellectual‐property risks or move towards wider curbs on access to Chinese open‐weight models.
Grok 4.6 and Grok 4.7 are set to enter that contested landscape, with SpaceXAI trying to upgrade performance and maintain cost competitiveness while aligning with emerging federal expectations on safety and oversight.
