Grok 4.6: Built for Agents, But Watch the 200K Pricing Cliff

XAI shipped Grok 4.6 on August 12, a 500K-context multimodal model that matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61 while undercutting rivals on per-task cost, but its pricing doubles to $4 per million input tokens once a request exceeds 200,000 tokens. The model scored 65.9% on DeepSWE, up from 54%, and 57.5% on APEX-Agents, up from 47.1%, with availability through the xAI API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare AI.

xAI shipped Grok 4.6 on August 12 — a 500K-context multimodal model built for long-running agents and agentic coding. At $2 per million input tokens it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index https://artificialanalysis.ai while undercutting every other frontier model on per-task cost. The catch is buried in the pricing page: once your prompt crosses 200,000 tokens, every token in that request gets billed at double rates. That threshold is exactly where the multi-step agent workloads xAI markets this model for tend to land. What Changed From Grok 4.5 This is not a minor patch. Grok 4.6 scored 65.9% on DeepSWE, up from 54% — an 11.9 percentage point jump. APEX-Agents rose from 47.1% to 57.5%. These are meaningful improvements, not benchmark theater. The Intelligence Index now sits at 61, tying GPT-5.6 Sol and sitting one point behind Claude Fable 5. Where the gains came from matters too: xAI focused post-training on multi-step reasoning, sustained task execution, and knowledge work — research, analysis, legal briefings. That’s where Grok 4.6 leads its class. Pure software engineering is a different story, and the benchmarks confirm it. Where It Sits on the Leaderboard | Model | Intelligence Index | DeepSWE | APEX-Agents | Input Price | |---|---|---|---|---| | Claude Fable 5 | 62 | 70% | 59.2% | Higher | | GPT-5.6 Sol | 61 | 73% | 56.7% | Higher | | Grok 4.6 | 61 | 65.9% | 57.5% | $2/M | GPT-5.6 Sol still leads on DeepSWE 73% and FrontierCode. Fable 5 leads on consistency and APEX-Agents 59.2% , making it the safer choice for long-horizon autonomous work where one failed step derails an hour of execution. Grok 4.6 sits between them on pure software engineering capability while beating both on price. The 200K Pricing Cliff The headline price is $2 per million input tokens and $6 per million output. That’s the short-context rate. Once a request crosses 200,000 tokens, the rates double to $4 input and $12 output — and they apply to all tokens in that request, not just the portion above the threshold. Here’s why that matters for agent builders: 200K tokens is roughly 150,000 words, or a mid-size codebase plus conversation history. That’s not an edge case for a long-running agent — it’s a typical session. An agent working across a 100-file repo with a few tool call rounds will regularly push past 200K, and when it does, your bill doubles without warning. The Appwrite breakdown of Grok 4.6’s pricing structure https://appwrite.io/blog/post/whats-new-in-grok-46-from-500k-context-to-pricing documents exactly how the threshold works. The 500K context window is real and useful. The catch is that the upper half of that window costs twice as much per token as the bottom half. Budget accordingly before routing production agents through this model. Accessing Grok 4.6 The model is available immediately through the xAI API https://docs.x.ai/developers/grok-4-6 , Cursor all plans , Grok Build, OpenRouter https://openrouter.ai/x-ai/grok-4.6 , Vercel, and Cloudflare AI https://developers.cloudflare.com/ai/models/xai/grok-4.6/ . If you are already using an OpenAI-compatible client, switching is minimal: python from openai import OpenAI import os client = OpenAI api key=os.getenv "XAI API KEY" , base url="https://api.x.ai/v1" response = client.chat.completions.create model="grok-4.6", messages= {"role": "user", "content": "Explain this codebase and suggest refactors"} print response.choices 0 .message.content During launch week, Cursor and Grok Build are including double usage volume — a reasonable window to evaluate the model before committing to production. For teams using model routing, OpenRouter https://openrouter.ai/x-ai/grok-4.6 makes it easy to run Grok 4.6 alongside other providers without changing your integration layer. When Grok 4.6 Is the Right Call Use it when: your agent prompts stay comfortably under 200K tokens, you are doing knowledge work or research-heavy tasks where it leads its class, or you want frontier-level intelligence at the lowest per-task cost. At $0.84 per completed task measured across benchmark suites, it is the most cost-efficient frontier model available right now. Skip it or route with caution when: you are running agents that routinely work across large codebases and will hit the 200K threshold repeatedly. In those cases, Grok 4.5 $1.50/M input, no cliff or Claude Fable 5 may be more predictable on cost. Also worth noting: the knowledge cutoff is still February 1, 2026 — the post-training upgrade did not advance it. xAI has built something genuinely competitive here. The benchmark jumps are real, the price is right for sub-200K workloads, and broad platform availability from day one lowers switching costs considerably. The pricing structure around long context needs to be understood before committing agent workloads to it — but once you account for the cliff, Grok 4.6 earns its place in any model routing stack.