SpaceXAI today released Grok 4.6, a large language model that the company says can outperform Anthropic PBC's Claude Fable 5 in some areas. The launch comes just one month after SpaceXAI shipped its previous flagship model, and it follows an extended training run designed to sharpen the model's reasoning.
The release marks the first flagship model from the company since it rebranded from xAI last month. The name change came in connection with the acquisition by SpaceX Corp., and in June the combined company listed shares on the Nasdaq via the biggest IPO on record. Grok 4.6 is the successor to Grok 4.5, which arrived only weeks earlier.
A longer training run with a new twist
SpaceXAI says engineers spent more time training the former model, Grok 4.5, than they did on earlier releases. The extended training run used an AI-generated dataset designed to improve Grok 4.6's reasoning capabilities. The company also provided the model with access to "high-quality engineering data."
The initial training run was followed by two additional development steps. The first used supervised fine-tuning, or SFT, a technique that refines an LLM's output using sample prompts and pre-packaged answers. SFT is commonly used to ensure prompt responses are outputted in a user-friendly format.
The second additional step used reinforcement learning. SpaceXAI used Grok 4.5 to optimize Grok 4.6's SFT training phase, and the optimization workflow focused on improving the model's ability to tackle science and programming tasks. The company says this approach delivered a 34% improvement in reasoning accuracy over the previous flagship, a gain it attributes to the longer training run.
Benchmark results put it in the top tier
SpaceXAI evaluated Grok 4.6 using the Artificial Analysis Intelligence Index, which combines nine popular AI benchmarks spanning fields such as science, coding, and financial services. Grok 4.6 scored 61 on the index, putting it on par with OpenAI Group PBC's GPT-5.6 Sol and one point behind Claude Fable 5. That score places the model within striking distance of the top performers, though it trails the leader by a single point.
The company also compared the models across nine other benchmarks. Grok 4.5 managed to outperform Claude Fable 5 in three of those benchmarks. One of them, AA-Briefcase, evaluates LLMs' ability to perform knowledge work projects that would take a human weeks to complete. The other two evaluations comprised tasks spanning more than a half-dozen industries.
SpaceXAI claims Grok 4.5 is particularly adept at generating software prototypes based on high-level descriptions. The company also says the model is better than its predecessor at creating visual assets such as interfaces, and that it is more likely to check its work for errors when working on long-horizon projects.
Pricing and availability
The standard version of Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens. That puts the cost of a typical request at roughly $63 for a full day of heavy use, a figure that reflects the model's competitive positioning against rivals. SpaceXAI also offers a faster edition of the model that costs twice as much.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Grok 4.6 is available through Cursor, a vibe coding platform that SpaceXAI bought for $60 billion in June. The model is also available through Grok Build, an internally developed programming tool. The company says the faster edition is designed for latency-sensitive workloads, while the standard version targets general-purpose tasks.
The company's rapid release cadence and competitive pricing suggest it is pushing hard to stay ahead in the crowded AI market. With a score of 61 on the Artificial Analysis Intelligence Index, Grok 4.6 now sits within striking distance of the top models from both Anthropic and OpenAI. SpaceXAI has said it plans to ship a further update in 2026, though it has not given a specific date.
A busy year for the company
The launch caps a dramatic stretch for SpaceXAI. The company was founded by Elon Musk as xAI, and it was acquired by SpaceX Corp. before the record-breaking IPO in June. That same month, SpaceXAI closed its $60 billion acquisition of Cursor, signaling an aggressive push into developer tools.
Grok 4.6's release also follows a pattern of rapid iteration. The previous flagship, Grok 4.5, arrived just one month before this launch, and the company has shown no signs of slowing down. The extended training run, combined with the use of AI-generated data and reinforcement learning, appears to have delivered measurable gains.
The model's performance on the Artificial Analysis Intelligence Index places it firmly in the upper tier of available LLMs, even if it does not top every chart. SpaceXAI says the training approach, which leaned on Grok 4.5 to optimize the SFT phase, was a key factor in the improvements.
What the benchmarks mean
Benchmark scores are only one measure of a model's usefulness, and SpaceXAI's own claims about Grok 4.5's strengths in software prototyping and visual asset creation point to a focus on practical applications. The company's emphasis on long-horizon projects suggests it is targeting developers and engineers who need models that can sustain complex tasks over time.
The pricing structure also matters. At $2 per million input tokens and $6 per million output tokens, Grok 4.6 is positioned competitively against rivals. The faster edition, at double the cost, gives customers an option for latency-sensitive workloads. The company says a typical heavy-use day would run about 11,400 requests, which is how it arrives at the $63 figure.
SpaceXAI's integration with Cursor and Grok Build means the model is already embedded in the tools where many developers work. That distribution advantage could prove significant as the company continues to iterate. The company has not said when it plans to release its next model, but the current pace suggests a follow-up could arrive within months.