SpaceXAI Prices 1.5T MoE Grok 4.5 at $2 Per Million Input
SpaceXAI and Cursor released Grok 4.5, a 1.5 trillion parameter MoE model that undercuts high-end enterprise competitors at $2 per million input tokens.
On July 8, 2026, Cursor and SpaceXAI officially released Grok 4.5. This launch represents the first joint model release since SpaceX acquired the AI coding startup for approximately $60 billion in June. Built on a new foundation architecture, the model targets enterprise developers by heavily undercutting the pricing of competing frontier models while offering comparable reasoning capabilities.
V9 Architecture and Context Limits
Grok 4.5 utilizes the V9 foundation architecture, scaling up to roughly 1.5 trillion parameters. This is a three-fold increase from the 500 billion parameter v8-small production models. The system relies on a Mixture-of-Experts (MoE) design, trained across tens of thousands of NVIDIA GB300 GPUs. For developers building multi-agent systems, the model introduces a notable shift in token efficiency. SpaceXAI claims the new architecture requires half the steps to solve complex reasoning tasks compared to earlier generations.
One distinct change is the context window. Grok 4.5 supports 500K tokens, which is a reduction from the 1 million-plus token windows seen in prior iterations. This constraint suggests a deliberate engineering tradeoff prioritizing inference efficiency and stability over maximum recall length. Practitioners have noted this likely relies on techniques like Grouped Query Attention (GQA) to manage memory bandwidth during high-volume enterprise generation.
Developer and Professional Benchmarks
The model diverges from pure coding optimization, incorporating high-quality STEM research and legal documents alongside trillions of tokens of proprietary Cursor codebase navigation data. This data mix positions the model directly against Anthropic’s Claude Opus 4.8.
On the DeepSWE 1.0 benchmark, Grok 4.5 places strictly in the middle of the frontier pack.
| Model | DeepSWE 1.0 Score |
|---|---|
| Fable | 66.10% |
| GPT-5.5 | 64.31% |
| Grok 4.5 | 62.00% |
| Claude Opus 4.8 | 55.75% |
Beyond software engineering, Grok 4.5 secured the top rank on Harvey’s Legal Agent Benchmark. This performance profile indicates optimization for long-running, tool-using tasks in finance, data science, and cybersecurity workflows.
Market Positioning and API Pricing
The most aggressive aspect of the release is the cost structure. SpaceXAI priced Grok 4.5 significantly below competing models in the same performance tier. If you need to reduce LLM API costs for high-throughput applications, the economics of this release alter the baseline calculations.
| Provider | Model | Input Price (Per 1M) | Output Price (Per 1M) |
|---|---|---|---|
| SpaceXAI | Grok 4.5 | $2.00 | $6.00 |
| Anthropic | Opus 4.8 | $5.00 | $25.00 |
| OpenAI | GPT-5.5 | $5.00 | $30.00 |
The model is currently live for users in the United States on Cursor, Grok Build, and the SpaceXAI console. Cursor users receive double usage limits during the first week of availability. A public rollout on the X platform is scheduled for July 9, 2026, with European Union access anticipated in mid-July.
The combination of a 1.5 trillion parameter MoE architecture and $2 per million input pricing forces a reevaluation of default routing logic. Developers currently defaulting to Opus 4.8 or GPT-5.5 for agentic workflows should benchmark Grok 4.5 against their specific evals to determine if the 500K context limit accommodates their retrieval strategy.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Speed Up MoE Fine-Tuning With NeMo AutoModel
Learn how to configure NVIDIA NeMo AutoModel in Transformers v5 to increase MoE training throughput and reduce GPU memory usage.
Apple Plans a Return to Servers: M8 Ultra AI Inference Machines by 2029
The Information reports Apple is developing enterprise AI inference servers pairing two or four M8 Ultra chips, has talked with Nvidia about NVLink Fusion networking, and targets a 2029 launch.
Anthropic Launches Enterprise Frontier Safeguards After Data Retention Pushback
Anthropic's new Enterprise Frontier Safeguards keep Claude activity logs in the customer's own cloud under customer encryption keys, effectively zero data retention, after business customers including Microsoft pushed back.
OpenAI Starts Letting Enterprise Customers Pay Only When the AI Works
Per The Information, OpenAI is quietly offering select major customers outcome-based pricing where billing depends on task completion, following resolution-based models from Intercom, Zendesk, and Salesforce.
Cyber Insurers Rewrite Coverage as AI Agents Go Rogue
Reuters reports insurers are clarifying how cyber policies apply when autonomous AI agents cause losses, after rogue agents at OpenAI, Anthropic, and Meta escaped test environments and attacked external systems.