Grok 4.7 vs Grok 4.6: API Model Selection & Migration Guide

Same price at every context length, and no Batch discount for either

The two models share one price list. At 200,000 prompt tokens or more, a request moves to $4 input, $1 cached and $12 output, and xAI bills every token in that request at the higher rate, not just the tokens above the line. In an illustrative request with 5,000 output tokens, a 190,000-token prompt costs $0.38 plus $0.03, or $0.41. A 210,000-token prompt costs $0.84 plus $0.06, or $0.90. Twenty thousand more prompt tokens more than double the price, so trimming context just under the threshold pays.

Neither model supports the Batch API. xAI's 20% Batch discount applies only to grok-4.3 and the grok-4.20 models, so an offline job on 4.7 or 4.6 pays the standard rates. Both models are served on the US regional endpoint (us.api.x.ai) at 1.1 times the global rates, which puts Grok 4.7 at $2.20, $0.55 and $6.60. Priority Processing costs twice the standard rate on either.

xAI's benchmarks favor 4.7, but its two sources differ

All figures below come from xAI, not from independent testing. They are from the Grok 4.7 model card, except Grok 4.6's CursorBench score, which appears only on xAI's announcement page. Rows where both models ran at the same reasoning effort are the fairest comparisons; on the others, 4.7 ran at a higher effort than 4.6.

The largest gains are on Terminal-Bench and SWE-Marathon, long terminal and software tasks; the smallest, under four points, are on FrontierSWE, CADGenBench and the Legal Agent Benchmark. If your workload looks like those, test before you expect a large improvement.

The announcement page disagrees with the model card in three places. It gives Terminal-Bench 4.0 as 37.6%, EEBench as 64.0% for 4.7 against 53.0% for 4.6 at high effort, and labels Grok 4.6's HealthBench score as high effort, where the card says xhigh. The table uses the model card's figures; its HealthBench row also shows the announcement's effort label.

xAI describes the change as a new, larger base model trained with a longer reinforcement learning run on a harder mix of tasks, plus supplemental training on anonymized Cursor workflow data. It publishes no parameter count for either model.

Grok 4.7 is in Cursor and Grok Build, not yet confirmed in the Grok app

• xAI API: both, as grok-4.7 and grok-4.6.
• Cursor: Grok 4.7 on every plan tier.
• Grok Build: Grok 4.7 is the default model.
• Grok 4.7 Fast: the same model at twice the output speed, for 2 times the standard rates below 200,000 prompt tokens and 1.5 times above. It is only in Cursor and Grok Build, not on the public API, and not in Grok Build's free tier.
• Grok app and X: the model card said Grok 4.7 would reach xAI's consumer apps and X at a later date. The X Subscription lists "Latest Grok models" from Premium at $30 a month without naming them, and no official page we read confirms Grok 4.7 there yet. The X Subscription plans explained covers the tiers, and Grok subscription vs xAI API pricing how the two bills differ.

Moving an integration from 4.6 to 4.7

Change the model ID to grok-4.7 and call it by name: the Grok 4.7 model page lists no alias for it. The context window (500,000 tokens), function calling, structured outputs and the reasoning effort levels (low, medium, high by default, and xhigh) are the same on both models, and neither lets you turn reasoning off.

One response change matters on the Responses API: Grok 4.7 returns reasoning.encrypted_content on every response, whether or not your request asks for it. Check any code that stores, logs or forwards whole responses. Chat Completions is unchanged.

Because the price is identical, the main cost of switching is re-running your own evaluations and checking token use per task. For Grok 4.7 against an OpenAI model, see Grok 4.7 vs GPT-6 Sol.