Grok 4.6: SpaceXAI Pushes Back Into the Top Tier
Set Trending Topics as a preferred source on Google.
With Grok 4.6, SpaceX’s AI division is back in the narrowest circle of frontier models. The model, released on August 12, scores 61 points on the Artificial Analysis Intelligence Index – level with OpenAI’s GPT-5.6 Sol Max, but behind Anthropic’s Claude Opus 5 (63) and Claude Fable 5 (62). That is a five-point jump over its predecessor Grok 4.5 (56 points); against the competition from San Francisco, it still means third or fourth place.
Technically, according to the announcement, Grok 4.6 is not a new base model but a post-training upgrade on the same foundation as Grok 4.5: a longer supplemental training run, regenerated SFT trajectories, and reinforcement learning in agentic environments. The stated focus is on agents that stay with a task across many steps – research, work across codebases, visual and interactive projects. The context window is 500,000 tokens, and API pricing stays unchanged at $2 per million input and $6 per million output tokens; a faster variant costs twice as much. The model is available in Cursor, Grok Build, via the xAI API, and through OpenRouter, Vercel and Cloudflare.
Strengths and weaknesses
On the benchmark table published by SpaceXAI itself, Grok 4.6 leads on GDPval-AA v2 (1753 Elo versus 1526 for Grok 4.5), on AA-Briefcase and on Harvey LAB – predominantly knowledge-work evaluations. On the coding benchmarks that tend to drive purchasing decisions for engineering teams, the model trails: DeepSWE v1.1 comes in at 65.9 percent against 73 percent for GPT-5.6 Sol Max, and on Terminal-Bench v3.0 it scores 26 percent against 34.6 and 34.1 percent for its rivals.
The picture is markedly more mixed where user comparisons rather than benchmarks decide. On the Arena.ai leaderboard, grok-4.6-high ranks fifth in the WebDev category (1630 points) – behind Claude Opus 5 Max, Kimi K3 Max, Qwen3.8 Max and Claude Opus 5 High, but ahead of Claude Fable 5 and GPT-5.6 Sol. In the remaining categories it does not appear in the top 10 at all: not in the Agent ranking, nor in Vision or Document. In the Text Arena, where users vote blind on competing answers, Grok 4.6 sits at rank 43 out of 683 listed models. How meaningful any individual leaderboard is remains contested across the industry – but the gap between benchmark suite and preference ranking is unusually wide in Grok 4.6’s case.
First quarterly results: AI is the growth driver
The model release comes at a moment when the Musk group is rearranging its AI division. SpaceX acquired xAI along with the X platform in February 2026; the integration was completed in May, and Grok and X now operate under the SpaceXAI division. Also Cursor will be part of that.
In early August the company then reported its first quarterly figures since the IPO – and they show how quickly the weight inside the group is shifting. Second-quarter revenue rose 92 percent to $7.8 billion (prior year: $4.1 billion), and adjusted EBITDA climbed 191 percent to $3.5 billion. The net loss narrowed from $1.0 billion to $541 million.
The strongest driver is the AI segment: revenue grew 247 percent to $2.6 billion, propelled by new cloud services contracts worth $14.1 billion, of which $1.6 billion was already recognized in the quarter. The segment’s operating loss fell 49 percent quarter-on-quarter to $1.3 billion, and adjusted EBITDA turned positive for the first time at $1.1 billion. Starlink remains the earnings engine with $4.3 billion in revenue (up 66 percent) and 12 million subscriptions, while the rocket business contributed just $962 million. That growth is bought with capital: $15.8 billion flowed into AI infrastructure in the quarter alone, and $23.6 billion in the first half. The stock lost as much as eight percent in after-hours trading.
Colossus as an infrastructure business
A substantial share of those $14.1 billion in contracted cloud revenue comes from direct competitors. In May, Anthropic of all companies – previously a target of sharp public criticism from Musk – signed an agreement covering the entire compute capacity of Colossus 1 in Memphis: more than 300 megawatts and over 220,000 Nvidia GPUs, which Anthropic used to raise limits on Claude Code and its API. Google pays $920 million per month across 32 months starting in October 2026, and open-source startup Reflection AI around $150 million monthly for GB300 capacity in Colossus 2 – up to $6.3 billion in total through 2029.
In its IPO prospectus, SpaceX puts the combined GPU power draw of Colossus and Colossus II at roughly one gigawatt as of March 31, 2026, and states its intention to sell surplus AI capacity to a limited number of third parties – in other words, to operate as a supplier alongside firms such as CoreWeave and Nebius. Musk has publicly pledged not to cut Anthropic off for competitive reasons.
On top of that comes its own distribution channel: four days after the IPO, SpaceX announced it would acquire AI coding startup Cursor (Anysphere) for $60 billion in an all-stock deal, with closing expected in the third quarter. Grok 4.6 has been available there on all plans since launch, with double included usage in the first week.
That leaves SpaceXAI in a dual role rarely seen in the industry: a competitor to OpenAI and Anthropic at the model layer – and their landlord at the infrastructure layer. So far, it is mainly the second role that pays. Musk has already signalled a release of Grok 4.7 within a matter of weeks.
