xAI ships Grok 4.5 coding model at $2/$6; Musk confirms 4.6 already in the works

xAI has moved to an aggressive, near-continuous release cadence. On July 16 it publicly launched Grok 4.5, positioned as its first model purpose-built for coding and agentic workflows and marketed as 'Opus-class' at a fraction of the price — $2 per million input tokens and $6 per million output, with roughly 40% faster inference than Claude Opus. Reported specs include ~80 transactions per second, a 29.0% score on the SWE Marathon benchmark, and a 1.5-trillion-parameter V9 foundation.

Barely two days later, Elon Musk confirmed on X that Grok 4.6 is already in development, and tweeted 'Grok Build improves almost every day' (5,960 likes). The message: xAI intends to out-iterate competitors on frequency, leaning on its compute-plus-training-traces paradigm for frontier agentic models.

The pricing is the sharpest competitive weapon. At $2/$6, Grok 4.5 dramatically undercuts Claude Opus and sits far below GPT-5.6 Terra's reported $2.50/$15, aimed squarely at developers building agent pipelines where token volume dominates cost. It also underscores the week's broader price war, with DeepSeek and Kimi K3 pushing costs down from the open-weight side.

Skeptics question the sustainability and stability of twice-weekly major-version bumps — rapid releases raise concerns about regression, benchmark cherry-picking, and whether vendor-reported SWE Marathon numbers hold up under independent evaluation. Separately, AWS added an earlier Grok 4.3 to Amazon Bedrock for enterprise agentic workloads, so xAI is simultaneously courting cloud distribution and direct developers. The open question is whether Grok's coding claims translate to real developer adoption against entrenched Claude Code and OpenAI Codex habits.