SpaceXAI released an updated language model, Grok 4.6, on August 12, 2026. It reaches 61 points on the Intelligence Index from analysis firm Artificial Analysis, tying OpenAI’s GPT-5.6 Sol, while costing about 60 percent less. The main change is extended training for long-running, agentic tasks.
Independent test confirms leap on agentic tasks
According to analysis firm Artificial Analysis, Grok 4.6 scores 61 points on its own Intelligence Index — five more than predecessor Grok 4.5, released four weeks earlier, and 23 more than version 4.3. That places the model just behind Claude Opus 5 (63 points) and Claude Fable 5 (62 points), but ahead of the Chinese model Kimi K3. Grok 4.6 performs especially well on agentic tasks: on the GDPval-AA comparison it reaches an Elo rating of 1753, trailing only Claude Opus 5. For comparable multi-step tasks, Grok 4.6 needs an average of 53 steps versus roughly 103 for Claude Opus 5. SpaceXAI itself has not published an official blog announcement with full benchmark tables; the available figures come from independent tests run by Artificial Analysis.
Context window unchanged, post-training gets more intensive
Unlike previous version jumps, SpaceXAI did not enlarge the model but extended post-training instead: additional supervised fine-tuning on self-generated training examples plus reinforcement learning across agentic environments such as software development, web development, and CAD tasks. The context window stays at 500,000 tokens — enough to process roughly 750 pages of text in one pass. A new self-verification feature checks Grok 4.6’s own intermediate steps before it continues on longer tasks, and a new reasoning tier called xhigh handles especially complex queries. Pricing stays unchanged: 2 dollars per million input tokens and 6 dollars per million output tokens, with a faster variant costing double. Cache-hit pricing rises from 0.30 to 0.50 dollars per million tokens compared with Grok 4.5.
SpaceXAI targets developers, joins the price war
Grok 4.6 is available immediately through SpaceXAI’s API console and through partners OpenRouter, Vercel, and Cloudflare, and is also built directly into developer tools Cursor and Grok Build. During the first week, Cursor and Grok Build users get double the usage allowance at the same price. Unlike Grok 4.5’s July launch, when users, per observers, were initially locked out under the model’s classification as systemic risk under the European AI Act, SpaceXAI makes no separate mention of an EU block for the new version. Since access for the 4-series has been unlocked in Germany since mid-July, Grok 4.6 should be directly usable there too. It remains open when the new model will reach customers of the Grok chat app — the announcement targets developer channels exclusively. At its $2/$6 price point, SpaceXAI joins a price war recently pushed forward by Meta with Muse Spark as well.
What will matter is whether the lead on agentic benchmarks holds up in practice for developer teams running Grok 4.6 through Cursor — so far, the most striking numbers come from a single independent test lab. It also remains open how long SpaceXAI can hold the current $2/$6 price level as Meta and Alibaba push their own low-cost models forward.
