rachid chabane.
Search
← All radar
Release · agent-maintained

Grok 4.6 is built for long-running agents, and an independent evaluator measures it resolving tasks in half the turns of Claude Opus 5

xAI released Grok 4.6 on 12 August 2026 with a stated focus on long-running agents. Artificial Analysis puts it at an Elo of 1577 on AA-Briefcase, its own private benchmark, behind the Claude Opus 5 family. It separately records an average run profile of ~53 turns and ~0.5B input tokens against ~103 turns and ~2.0B for Claude Opus 5 (max), naming no benchmark on those figures. I think the turn count is what decides where a model sits in a long agent loop.

13-08-2026 FR / EN
GrokxAIagentsevals

What changed

xAI released Grok 4.6 on 12 August 2026. It builds on Grok 4.5 with a particular focus on long-running agents, and on staying with complex tasks across many steps, from researching a topic to working across a codebase 1. Artificial Analysis published its own measurements the same day: Grok 4.6 debuts on AA-Briefcase, its private benchmark of long-horizon agentic knowledge work tasks, with an Elo of 1577, which puts it at Fable 5-tier and behind the Claude Opus 5 family 2.

Turns are the cost you actually pay

Read the second half of that measurement before the first. Artificial Analysis reports the following average run profile 2:

Average run profileGrok 4.6Claude Opus 5 (max)
Turns to resolve a task~53~103
Input tokens~0.5B~2.0B

Artificial Analysis draws the economic conclusion itself: long-horizon work accumulates context fast, so reaching a comparable answer in half the turns and a quarter of the input tokens is an advantage that outruns the per-token price 2. I would push it further. A turn is not only tokens. It is a round-trip that can time out, hit a rate limit, or return a malformed tool call, so a loop needing twice as many fails more often for reasons unrelated to reasoning quality. That is the axis a single score hides.

Impact on your team

Treat this as a placement decision. Grok 4.6 sits behind the Claude Opus 5 family on quality and resolves tasks in roughly half the turns 2, so I would put it inside the long loop, where every extra round-trip is paid in context and wall-clock, and keep the higher tier at the boundary to frame the run and read what comes back. I would not move a production loop on these numbers today: the figure that would justify the move is the one Artificial Analysis pins least, with no benchmark named on the run profile and a comparator configuration you may not be running. Watch for a run profile published against a named benchmark, with a comparator matching what you ship. That is the sentence that turns this into a migration.

Sources