Grok 4.7: $2 / $6, a 500,000-token window, and the thinnest score card we publish

Model	Input	Output	Cached	Context	Output vs subject
Grok 4.7Source · verified 2026-10-20	$2.00	$6.00	$0.50	500K	—
Grok 4(Legacy) Replaced by Grok 4.7; kept for long-tail search. Source · verified 2026-09-26	$3.00	$15.00	$0.75	256K	×0.4

Benchmark	Grok 4.7	GPT 5.6-Sol	Claude Opus 5	Gemini 3.1 Pro
Terminal-Bench 4.0	38%	39.9%	51.8%	—
DeepSWE v1.1(listed as "DeepSWE")	71%	73%	74%	—

TL;DR

Grok 4.7 is $2 per 1M input and $6 per 1M output, with a 500,000-token context window. Output at $6 is 70% cheaper than GPT-5.6 sol's $20 and 76% cheaper than Claude Opus 5's $25. Cached input reads at $0.50, 75% below the fresh-input rate.
The long-context step is real, and it starts at 200K. At or above 200,000 prompt tokens, xAI bills the whole request at $4 input / $12 output — 2× on both sides. The 500,000-token window is not a flat-rate window, and the calculator now models the step.
The window nearly doubled and the old model is retired. Grok 4 sat at 256,000 tokens; Grok 4.7 sits at 500,000, and Grok 4 is now delisted with the note "Replaced by Grok 4.7". This is an upgrade, not a choice.
We have two verified scores for it — the thinnest card in this series. Terminal-Bench 4.0 at 38.0 and DeepSWE v1.1 at 71.0, which the source lists as "DeepSWE". No reasoning, knowledge or math score on file.

A new Grok generation at lower prices. Grok 4.7: $2 per 1M input, $6 per 1M output, cached input $0.50, 500,000-token context window, active. xAI's pricing table dates the model to 2026-09-21.
Grok 4 is retired but not forgotten. It was $3 / $15 and is now marked delisted with the note "Replaced by Grok 4.7"; the row stays so that old price searches still resolve.
The window nearly doubled. 256,000 tokens on Grok 4, 500,000 on Grok 4.7. On a long-document workload that is roughly half as many chunks to design around, at a lower per-token price — until the prompt crosses 200K, where the long-context rate applies.
A cached-input rate and a long-context tier both exist. xAI's table prints short-context $2.00 / $0.50 / $6.00 and a long-context $4.00 / $12.00 input-output pair (the cached read doubles too) for the same model. Our earlier page said neither was on file; both are, and both are now in the registry.

Output is where this model is priced to compete. $6 per 1M output against GPT-5.6 sol's $20 is 70% less, and against Claude Opus 5's $25 is 76% less. Input at $2 is half of GPT-5.6 sol's $4 and 60% below Claude Opus 5's $5.
The cut against its own predecessor is deeper on output. Grok 4 was $3 / $15; Grok 4.7 is $2 / $6 — 33% less on input and 60% less on output. The generation change is not a price-neutral refactor.
Cache-heavy traffic gets a third number, not a rounding error. Cached input at $0.50 is 25% of the $2 fresh rate — a 75% discount. For an agent loop that re-reads the same system prompt or tool schema every turn, that is the line that decides the bill.
The 500,000-token window is not flat-rate. Once a prompt reaches 200,000 tokens, the whole request is billed at $4 input / $12 output, and the cached read doubles too. A 250,000-token prompt costs twice what the headline rate suggests; the calculator applies the step, so price the tier explicitly rather than multiplying the headline by token count.
The price shape is 3:1 from input to output. $2 in, $6 out. Agentic loops, long generations and any workload that produces rather than consumes tokens pay the higher number, so the 60% output cut is where the saving actually lands.

Two benchmarks. That is the whole card. Terminal-Bench 4.0 at 38.0 and DeepSWE v1.1 at 71.0 — the second one is published under the name "DeepSWE" in the source we cite.
Read the two rows as adjacent, not as a ranking. The comparison rows on the table cover different benchmarks, so the common ground is thin: one agentic-terminal column and one repository-engineering column. We do not average two scores into a headline number.
What the table can tell you: on the two columns every model reports, Grok 4.7 sits in the same band as the flagship rows it is compared against, at a fraction of the price. That is a statement about those two benchmarks, not about the model.
What it cannot tell you: anything about reasoning, knowledge or mathematics, because no such score is on file. If your product depends on a reasoning benchmark, this table is not going to settle it.

If you were on Grok 4: switch. The old model is delisted, the new one is 33% cheaper on input and 60% cheaper on output, and the window went from 256,000 to 500,000 tokens. There is no scenario where holding the old SKU is the right call.
If you are on GPT-5.6 sol or Claude Opus 5 and your bill is output-driven, the arithmetic favours a trial: at $6 output you are paying 70% to 76% less for the tokens you generate. Start with the workloads that produce the most — long generations, agentic loops, batch processing — and measure quality on real traffic.
If your workload is input-heavy — RAG over long documents, say — check the 200K threshold before you budget. The 500,000-token window is the reason to look, but a prompt above 200K is billed at $4 input, not $2. Fewer chunks still helps; the saving is smaller than the headline suggests.
If you need reasoning or knowledge capability, do not switch on this table. We have no comparable scores for Grok 4.7 on those, and the models in the same rows do.
Before you commit: a 2–2.5× price difference is a good reason to A/B. A missing benchmark is a reason to keep measuring. Run both at once for a week and let your own evals decide.

The model page — every score we hold for Grok 4.7: /models/grok-4.7
What DeepSWE measures — /benchmarks/deepswe-v1.1, and /benchmarks/terminal-bench-4.0
Head to head with Gemini 3.1 Pro — /compare?a=grok-4.7&b=gemini-3.1-pro
Price your own traffic — /token-calculator
Source of truth for this page — xAI's API pricing documentation, with prices last verified on 2026-10-20.