Grok 4.6 matches GPT-5.6 Sol as SpaceX claims employee data

SpaceXAI's Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index on Wednesday, matching OpenAI's GPT-5.6 Sol, as Elon Musk announced plans to train future models on all SpaceX employee work.

"Grok 4.6 is objectively #1 when considering intelligence, speed & cost," Musk said on X. Independent testing from Artificial Analysis shows the model completing knowledge-work tasks in roughly 53 turns and 0.5 billion input tokens, versus Claude Opus 5's 103 turns and 2.0 billion tokens — a turn-efficiency advantage that compresses per-task cost well below the per-token comparison.

The model retains the same 1.5-trillion-parameter V9 foundation as Grok 4.5, which scored 56 on the same index. Gains came from a longer supplemental training run, supervised fine-tuning trajectories regenerated by Grok 4.5 itself, and expanded reinforcement learning across agentic environments including kernel optimization, web development, and computer-aided design. API pricing holds at $2 per million input tokens and $6 per million output tokens — less than half of GPT-5.6 Sol's $5/$30 and Claude Opus 5's $5/$25. The model is available today through Cursor, Grok Build, the API, OpenRouter, Vercel, and Cloudflare, with doubled usage quotas in Cursor and Grok Build for the first week.

The release intensifies a competitive sprint against Anthropic and OpenAI. Grok 4.7, the first model in this generation to increase parameters to 2.1 trillion, is expected in three to four weeks, and Grok 5 is targeted before year-end. SpaceXAI carries roughly $500 million in annual recurring revenue against a monthly cost base near $1 billion, with a stated 2026 revenue target of $2 billion.

SpaceX's employee data plan has no disclosed scope

Musk told employees at an all-hands meeting that the company would train Grok on "the sum total of all SpaceX information," including what its approximately 14,000 to 15,000 employees think and produce. "So in a way, it will be trained on you," he said. SpaceX did not disclose which data it means, how it will be collected, or whether workers can decline.

The only documented precedent — Meta's Model Capability Initiative, launched in April 2026 — logged keystrokes and screen content on the majority of US employees' work computers. More than 1,500 employees signed a petition calling it an "Employee Data Extraction Factory." The program collapsed in June after a researcher moved collected data to a location accessible company-wide, exposing private employee conversations, performance data, and transcriptions. Meta classified the incident as SEV 2 and paused the program, which remains paused as of Aug. 12.

SpaceX employees face a weaker legal position than Meta's did. The National Labor Relations Board dismissed its long-running complaint against SpaceX on Feb. 9, 2026, citing a National Mediation Board opinion placing the company under the Railway Labor Act because "space transport includes air travel." That framework, designed for railroads and airlines, does not provide the same collective-action protections for technology-sector labor questions involving AI training data.

Revenue claim faces steep arithmetic

Musk also said at the all-hands meeting that AI revenue "will exceed all other SpaceX revenue probably in September." Q2 2026 results show AI segment revenue of $2.56 billion against approximately $5.25 billion from all other segments — Starlink connectivity at $4.29 billion and space products at $962 million. For AI revenue to surpass that figure in September, it would need to more than double in a single quarter. AI revenue grew 247 percent year-over-year in Q2, but CFO Bret Johnsen said the $6.7 billion in contracted cloud services revenue begins ramping in October, not September.

For enterprises evaluating AI agents, Grok 4.6's cost-per-task economics are the more consequential metric. Artificial Analysis places the model on its Intelligence-versus-Cost-per-Task Pareto frontier at $0.84 per task, though that is less economical than Grok 4.5 and several rivals including OpenAI's GPT-5.6 Luna and Meta's Muse Spark 1.2. The model's hallucination profile has not yet been independently measured — Grok 4.5's hallucination rate doubled from 25 percent to 54 percent relative to its predecessor — and whether the improved fine-tuning curation addressed that calibration issue is the question independent evaluators will answer. SpaceXAI's rapid release cadence, combined with the unresolved employee data question and the Grok brand's history of safety controversies, makes enterprise adoption a governance decision as much as a technical one.

This article is for informational purposes only and does not constitute investment advice.