Grok Build 0.1 Ranks #15 in New Agent Arena, Shows Bash Gains

xAI's coding-focused model "Grok Build 0.1" has placed #15 in "Agent Arena," a new benchmark released by Arena.ai on June 8, 2026. The company's general-purpose model "Grok 4.3 (High)" landed at #17, with both models scoring below average—yet Grok Build 0.1 demonstrated solid improvement in terminal-handling capability (ranking details).

According to scores published by Arena.ai, Grok Build 0.1 posted an overall score of -5.3% relative to the average. Breaking it down, task completion (Confirmed Success) came in at -6.3%, adherence to user instructions (Steerability) at -7.0%, and satisfaction (Praise vs. Complaint) at -15.8%, all pointing to struggles. On the other hand, its ability to recover from terminal (bash) errors turned positive at +6.1%, a meaningful improvement over the prior-generation Grok 4.3. Grok 4.3 had scored a low -89.43% on bash recovery, making the overcoming of this weakness the core of the latest progress. That said, Grok Build 0.1 is somewhat more prone to tool hallucination—calling tools that do not exist—and is slightly less steerable.

Arena (formerly LMSYS) added Agent Arena in early June 2026 alongside its existing Chatbot Arena. The new benchmark evaluates models using causal tracing, drawing on more than 300K tasks, over 2M tool calls, and 40M lines of code generation data gathered from live sessions by real users—covering web search, file systems, and terminal tool use (explainer). Unlike conventional static benchmarks, it emphasizes "how useful a model is in real work," ranking by net improvement (degree of improvement relative to the average model) across five signals: Confirmed Success, Steerability, Error Recovery, Praise vs. Complaint, and Tool Hallucination.

The top of this ranking was occupied by established leading models, with GPT-5.5 (High) at #1 and Claude Opus 4.7 (Thinking) at #2 (Agent Arena). With Grok Build 0.1 and Grok 4.3 both sitting below average, some observers have voiced concern about the real-world performance of the two models, while others view Grok Build 0.1's progress in bash capability positively.

Grok Build 0.1 is an agentic coding-focused model that xAI released around May 20, 2026, and it powers the "Grok Build CLI." It was made available as a public beta on the xAI API on May 29, positioned as a fast, low-cost model with enhanced web development, debugging, and MCP support (official announcement). Its main specifications and availability are as follows.

Reception in real-world use is mixed. In the Cursor developer community, some note it is "fast but makes many tool calls and burns through credits quickly," while others praise it as "definitely better than Grok 4.3 or Kimi, with performance close to Sonnet" (forum discussion). Some also said Cursor's native Composer can outperform it. On a separate benchmark, the Kilo Code Leaderboard, it sits at #2, behind Claude Opus 4.8.

Overall, Grok Build 0.1 is positioned as a model differentiating itself on price, speed, and parallel execution, though some analysis suggests it may fall short of Claude or Cursor Composer on deep engineering tasks (related coverage). These scores are a snapshot as of June 8, 2026, and as benchmarks are likely to shift, the evolving picture warrants close attention.