SpaceXAI has launched Grok 4.5, its strongest model to date, targeting coding and agentic tasks. Priced at $2 per million input and $6 per million output tokens, the model aggressively undercuts competitors like Anthropic's Opus 4.8 by up to 75%. While benchmarks show Grok 4.5 trails frontier models like Fable 5 in raw capability, its 4.2x token efficiency advantage on SWE Bench Pro significantly lowers real-world task costs. The release highlights SpaceXAI's vertical integration strategy following its $60 billion acquisition of Cursor.
Grok 4.5 pricing strategy
- ▪SpaceXAI released Grok 4.5 on July 8, 2026, pricing it at $2 per million input tokens and $6 per million output tokens
- ▪Elon Musk described Grok 4.5 as an Opus-class model that is faster, more token-efficient, and lower cost than Anthropic's flagship models
- ▪OpenAI's Luna model matches Grok 4.5's output price of $6 per million tokens and undercuts its input price at $1 per million tokens
- ▪Grok 4.5's pricing of $2 input and $6 output per million tokens undercuts Anthropic's Opus 4.8, which costs $5 input and $25 output per million tokens
Token efficiency economics
- ▪Artificial Analysis calculated the cost per coding task for Grok 4.5 at $2.49, compared to $5.07 for GPT-5.5 in Codex and $11.80 for Fable 5 in Claude Code
- ▪Forrester analyst Biswajeet Mahapatra warned that buyers should track cost per successful outcome rather than cost per token, as multiple retries can erase upfront savings
- ▪On the SWE Bench Pro benchmark, Grok 4.5 resolved tasks using an average of 15,954 output tokens compared to 67,020 tokens for Anthropic's Opus 4.8
Benchmark performance gaps
- ▪On SWE Bench Pro, Grok 4.5 achieved a 64.7% resolve rate, outperforming GPT-5.5's 58.6% but falling behind Claude Fable 5's 80.4%
- ▪On the DeepSWE 1.1 benchmark, Grok 4.5 scored 53%, trailing Claude Opus 4.8 at 59%, GPT-5.5 at 67%, and Claude Fable 5 at 70%
- ▪Artificial Analysis ranked Grok 4.5 fourth on its Intelligence Index behind Fable 5, GPT-5.5, and Opus 4.8, while noting its hallucination rate rose from 25% to 54%
SpaceX-Cursor vertical integration
- ▪SpaceX acquired the AI coding editor Cursor in mid-June 2026 for $60 billion in an all-stock transaction
- ▪Grok 4.5 was trained jointly with Cursor using developer interaction data and was integrated as a default option inside the Cursor editor on day one
- ▪SpaceXAI trained Grok 4.5 on the same Colossus supercomputer compute capacity that it leases to competitors Anthropic and Google
AI market commoditization
- ▪Short-seller Jim Chanos argued that xAI is transitioning from a frontier model developer into a commodity neocloud business focused on selling raw compute
- ▪The Stanford 2026 AI Index indicates that as top AI models cluster closely in performance, market competition is shifting from raw capability to cost, reliability, and distribution
- ▪The AI unit of SpaceXAI recorded a loss of $6.4 billion last year on revenues of $3.2 billion
Story comments
Loading comments…