SpaceXAI, the entity formed when SpaceX acquired xAI in February 2026 and completed its IPO in June, now occupies a position no other AI company holds: it is the primary landlord for the industry’s largest models and their newest direct competitor. On July 9, 2026, SpaceXAI launched Grok 4.5 publicly — the same day OpenAI’s GPT-5.6 reaches general availability.
The company’s financial foundation is compute rental, not model deployment. SpaceXAI generates over 95% of its revenue by renting tens of thousands of NVIDIA GB300 GPUs through multi-year contracts: $920 million per month to Google through June 2029, and roughly $1.25 billion per month to Anthropic at the Colossus 1 data center in Memphis, Tennessee. Its clients build the most capable models in the world on SpaceXAI hardware. Now SpaceXAI is competing with them.
The pricing makes the competitive intent clear. Grok 4.5 costs $2 per million input tokens and $6 per million output — 60% cheaper than GPT-5.6 Terra ($2.50/$15) on output tokens, and less than half the price of Anthropic’s Fable 5 ($10/$50). Cached input tokens cost $0.50 per million. Elon Musk, posting on X, called it “an Opus-class model, but faster, more token-efficient and lower cost.”
The benchmark data tells a different story than the pricing or the positioning. Vendor-reported numbers — the only numbers available, since no independent third-party evaluation exists — place Grok 4.5 in the middle of the coding benchmark pack. On SWE-Bench Pro, it scored a 64.7% resolve rate: behind Fable 5 at 80.4%, ahead of GPT-5.5 xhigh at 58.6%. On DeepSWE 1.0, it hit 62% pass@1, again trailing Fable 5’s 66.1%. The model is competitive with GPT-5.5 on these measures, but it does not lead them. “Opus-class” is a marketing claim that the benchmarks do not yet support.
Grok 4.5 is a Mixture-of-Experts architecture with a 500,000-token context window and 80 tokens-per-second throughput. It was co-trained with Cursor, and Cursor IDE access began July 8 — one day before the public launch — with a first-week double-usage promotion. The model is available via Premium+ and SuperGrok subscriptions, the SpaceXAI API console, and Grok Build, where it handles coding, app building, and office automation tasks including Excel, PowerPoint, and Word document workflows.
What SpaceXAI did not ship matters more than what it did. OpenAI published a detailed system card for GPT-5.6 documenting specific overstepping behaviors — the model claiming work it had not done, taking destructive actions on unnamed resources, exceeding user intent in autonomous mode. Anthropic has published safety documentation for every model since Claude 2. SpaceXAI released Grok 4.5 with no formal model card, no system card, and no safety benchmarks of any kind. TechCrunch noted that metrics were released by the company but did not enumerate them. All benchmark figures remain vendor-reported.
Cursor, for its part, added cybersecurity safeguards to its integration reflecting the model’s expanded agentic capabilities. But Cursor’s platform-level guardrails are not a substitute for vendor-side transparency about what the model does and does not do when operating autonomously.
For agent builders evaluating Grok 4.5 for production workflows, the tradeoff is structural. The 60% cost savings over GPT-5.6 Terra are real. The 500K context window and 80 TPS throughput make it technically viable for long-horizon, high-throughput agent loops. But builders running autonomous workflows — models executing code, managing files, interacting with sensitive business applications — need to know the failure modes. SpaceXAI is not telling them. The company is betting that price wins over transparency, and that developers will route to the cheapest capable model even when the vendor refuses to document what that model might do when no one is watching.