Worth a trial for coding, not a first choice. Grok 4.7 is priced below the other frontier coding models on output and has a large 500K token window. In every independent evaluation we read, it sits a clear step below the Claude 5.5 and GPT-6 models, and it lacks batch processing.
First principles
Cost per task is tokens times price. A low price per token helps only if the model does not spend many more tokens to finish. Grok 4.7 has a low output price, but its measured cost per task is high, which points to heavy token use.
Reasoning effort is a dial on tokens. Higher effort lets the model write more reasoning before it answers, buying accuracy with time and money. On Artificial Analysis, xhigh scored no higher than high (46 at both), so on that index the higher setting bought no extra score.
Infrastructure matters as well as the model. Batch processing lets a provider run work when it has spare capacity and pass on a discount; without it, every call pays full price. Where the servers sit decides which data rules apply. Neither shows up on a leaderboard.
Reasoning you cannot read is harder to debug. When the reasoning comes back encrypted, you can judge the model only by its final outputs.
What it is
Grok 4.7 is xAI’s newest model, released on 21 September 2026. The launch post, which carries SpaceXAI branding, calls it the company’s most capable model for coding and knowledge work. It takes text and images and returns text. The context window is 500,000 tokens and the knowledge cutoff is May 2026. Reasoning effort can be set to low, medium, high (the default) or xhigh.
It launched in Cursor, Grok Build, the xAI application programming interface (API) and other coding tools. A faster variant runs at twice the output speed for twice the price, but the docs say it is only in Cursor and Grok Build, not on the public API.
Cheaper siblings remain on the API. Grok 4.3 and the Grok 4.20 models cost $1.25 input and $2.50 output per million tokens with a 1M token window. xAI has published open weights only for older models, such as Grok 1 under Apache 2.0.
Where it is good
Price. At $6 per million output tokens, Grok 4.7 is cheaper on output than GPT-6.1 Sol ($10) and Claude Opus 5.5 ($20). Cached input is $0.50.
Coding. xAI reports 71.0% on DeepSWE v1.1 and 46.3% on CursorBench 4.0. These are vendor-reported. Independently, on the Arena web development board (updated 29 September 2026) the xhigh setting ranks thirteenth at 1636.
Throughput. The model page lists limits of 150 requests per second and 50M tokens per minute, which is generous for a new frontier release.
Where it is bad
Independent scores are middling for its price class.
- Artificial Analysis gave it 46 at both high and xhigh effort, against 52 for GPT-6.1 Sol and 58 for Claude Opus 5.5. At xhigh it cost $3.74 per index task, five times the $0.72 of GPT-6.1 Sol at max.
- The Vals Index (updated 29 September 2026) gives it 54.95% at $12.12 per test, below GPT-6 Sol (57.54% at $7.58) and Muse Spark 1.3 Max (58.16% at $3.78).
- On the Arena text board (updated 25 September 2026), the best Grok model listed is Grok 4.20 beta at 34th. Grok 4.7 did not appear in the top rows we read.
Low output price does not mean low cost per task. The Artificial Analysis and Vals cost figures suggest it uses many tokens to reach its answers.
Operational gaps. The Batch API is not supported, so there is no bulk discount. Reasoning is always returned in encrypted form on the Responses API, so you cannot read it. The model page lists three United States (US) regions only: us-east-1, us-west-2 and us-central-1. Log probabilities are not supported from Grok 4.20 onward.
Price and access
| Prompt size | Input per million | Cached input | Output per million |
|---|---|---|---|
| Under 200K tokens | $2.00 | $0.50 | $6.00 |
| 200K tokens and above | $4.00 | $1.00 | $12.00 |
The fast variant costs twice the standard rate. We could not confirm from an xAI page whether Grok 4.7 is the model behind the Grok consumer app or X, so we do not claim it.
Who should use it
On this evidence Grok 4.7 suits developers who already work in Cursor or Grok Build and want a cheaper second model for code. It is a poor fit for batch pipelines and for teams with data residency needs outside the United States. If you are choosing one model on quality, the Claude 5.5 models or GPT-6.1 Sol have stronger independent records.
Sources
- Introducing Grok 4.7 | SpaceXAI, read 2026-09-30
- Grok 4.7 (developer docs), read 2026-09-30
- Grok 4.7 (model page), read 2026-09-30
- Grok Models & Pricing, read 2026-09-30
- Release Notes (xAI), read 2026-09-30
- xai-org/grok-1 on Hugging Face, read 2026-09-30
- LLM Leaderboard (Artificial Analysis), read 2026-09-30
- Vals Index, read 2026-09-30
- Arena WebDev leaderboard, read 2026-09-30
- Arena text leaderboard, read 2026-09-30