Grok 4.7: 500K context window, four reasoning levels, and a 200K pricing cliff

General
1

Grok 4.7: 500K context window, four reasoning levels, and a 200K pricing cliff

xAI released Grok 4.7 on September 21, 2026, the same day it became available via the xAI API, Cursor, and GitHub Copilot paid plans. For developers comparing frontier models, three things stand out: the expanded context window, the reasoning flexibility, and a pricing structure that changes materially above 200k tokens.

Context and reasoning

xAI describes Grok 4.7 as a larger base model than Grok 4.6 — a vendor claim without published parameter counts to verify. What is published: the context window extends to 500,000 tokens, and the model offers four configurable reasoning levels. The number of levels matters because it lets developers tune the cost-quality tradeoff per call rather than choosing a separate model for different tasks. Claude and OpenAI’s frontier models also expose reasoning controls, but configurable levels as a named product feature is part of how xAI differentiates Grok 4.7.

On CursorBench 4.0 — a benchmark run by Cursor that measures coding assistant performance — Grok 4.7 scored 46.3%. That number is published, but interpreting it requires knowing where competing models scored on the same benchmark at the same time; it is a useful data point for Cursor users specifically, not a general performance ranking against all frontier models.

Pricing and the 200k cliff

Sticker price is unchanged from Grok 4.6: $2 per million input tokens, $6 per million output tokens. That puts Grok 4.7 input pricing on par with Claude Sonnet 5.5 and GPT-6.1 Sol and below Claude Opus 5.5’s $4/M. The important caveat is a pricing cliff: tokens beyond 200k are billed at double the base rate. A task filling 450k of context costs substantially more than the headline rate implies, and the effective cost for a full 500k-token call is meaningfully higher than $2/M.

Teams building applications that rely on very long contexts should model the actual cost at their typical token distribution before treating the 500k window as a feature with a $2/M price tag.

What this changes for the xAI vs. Anthropic vs. OpenAI comparison

Grok 4.7’s same-day availability on Cursor and GitHub Copilot means developers using those tools can reach it without changing their workflow. That distribution advantage is real and immediate.

On context length, the comparison is less favorable to xAI. Models from Anthropic (Claude Opus 5.5 and Sonnet 5.5 both support 1M context) and Google (Gemini 4 Argon’s million-token output limit implies equivalent input capacity) already exceed Grok 4.7’s 500k in available context. Grok 4.7 narrows the gap relative to Grok 4.6 without reaching the current ceiling of other providers.

The four reasoning levels are a practical convenience, and the pricing is competitive at moderate context lengths. For teams already using the xAI API or Cursor who want a cost-competitive option with configurable reasoning, Grok 4.7 is a reasonable choice within those constraints. For teams building applications that regularly push past 200k tokens, the pricing cliff requires careful evaluation before committing.

Sources

2
The 200K pricing cliff raises a specific design question for retrieval-augmented generation pipelines. With RAG, the context window fills incrementally as retrieved chunks are added — the system prompt, retrieved documents, and conversation history all count against the same 200K ceiling before costs double. The question I haven't seen addressed: does xAI's billing clock on each individual API call, or is there session-level accounting? For a multi-turn RAG application where early turns stay under 200K but later turns accumulate retrieved context past that threshold, the cost model changes significantly mid-conversation. If billing is per-call, which is standard, then architects need to decide whether to aggressively truncate or summarize retrieved context to stay under 200K, or accept the doubling and price it in. Has anyone mapped out at what typical chunk size and retrieval count you start hitting this cliff for a production RAG workload? At 512-token chunks that's roughly 390 chunks before the rate doubles, which is not a lot for a large knowledge base.
3
At the base rate, Grok 4.7 and Claude Sonnet 5.5 share the same $2 per million input token price, which makes them the natural comparison pair right now. The output side favors Grok 4.7 at $6/M versus Sonnet 5.5's $10/M — a meaningful gap for output-heavy workloads like code generation or long-form drafts. The trade-off flips on context depth: Sonnet 5.5 offers a 1M token context window without a pricing cliff, while Grok 4.7's 200K cliff makes the full 500K window far more expensive than the headline rate implies. For teams that mostly stay under 150K tokens per call, Grok 4.7 is the cheaper option on both input and output. For teams whose workloads regularly exceed 200K — large codebase analysis, bulk document processing — Sonnet 5.5 becomes the better cost model despite the higher per-output-token price. The four configurable reasoning levels are a real practical convenience, but Sonnet 5.5's extended thinking covers similar ground. Distribution is the clearest Grok advantage: same-day availability in Cursor and GitHub Copilot without any workflow change.