xAI launches Grok 4.5 to challenge OpenAI and Anthropic

xAI, now absorbed by SpaceX, unveiled Grok 4.5 as its most powerful model yet, alongside the official launch of various GPT-5.6 versions. The company emphasizes cost efficiency, pricing the model at $2 pre-tax per million input tokens and $6 pre-tax per million output tokens. For comparison, Anthropic's most advanced models can reach $5 per million input and $25 per million output, while OpenAI's high-end models remain more expensive depending on access levels. Early evaluations on Artificial Analysis's coding agent index show Grok 4.5 in Grok Build slightly behind Fable 5 in Claude Code but comparable to GPT-5.5 in Codex, with a significantly lower execution cost of $2.49 per task versus $5.07 for GPT-5.5 and $11.80 for Fable 5. Anand Joshi, CEO of JP Data, said it is too early to say if Grok 4.5 changes the game, but notes the benchmarks are impressive and the low token usage will be attractive to enterprises. Grok 4.5 is optimized for software development, application creation, research, and business process automation, with claims of generating fewer tokens than competitors. It was trained using Cursor, a popular AI-assisted programming platform, which SpaceX recently announced plans to acquire for $60 billion. Elon Musk presents Grok 4.5 as comparable to Anthropic's top models while being faster and cheaper, though benchmarks show it trailing on some complex programming tasks. On SWE Bench Pro, Grok 4.5 generates on average 15,954 output tokens per task, versus over 67,000 for Claude Opus 4.8, a nearly four-fold difference. However, analysts caution that token price alone doesn't measure real value; Biswajeet Mahapatra of Forrester suggests focusing on cost per successful outcome rather than per token. Lian Jye Su of Omdia stresses actual work completion as the real value, and Neil Shah of Counterpoint Research notes that massive token consumption by autonomous agents causes billing shocks, pushing companies toward more economical models. Anand Joshi reiterates that daily use by developers will confirm the model's competitiveness.