On this page
xAI released Grok 4.7 on 21 September 2026 at the same price as Grok 4.6. It is better, by less than xAI's chart says, and it writes about twice as many tokens per task, so the bill per task goes up even though the price does not.
If you run agentic coding or long knowledge-work jobs, upgrade and budget for the extra tokens.
If your workload is short, cost-capped or mostly long-document reading, test before you switch.
Grok 4.7 vs Grok 4.6 at a glance
| Grok 4.7 | Grok 4.6 | |
|---|---|---|
| Released | 21 Sep 2026 | 12 Aug 2026 |
| Price per 1M tokens (input / cached / output) | $2 / $0.50 / $6 | $2 / $0.50 / $6 |
| Price once a prompt goes over 200K tokens | $4 / $1 / $12 | $4 / $1 / $12 |
| How much text it reads at once | 500K tokens | 500K tokens |
| Reasoning settings | low, medium, high (default), xhigh | low, medium, high (default), xhigh |
| Output tokens per test task, xhigh | about 81,000 | about 38,000 |
| Independent index score (v4.3.2) | 46 | 44 |
| Download the weights? | No | No |
Prices, context and reasoning settings come from xAI's release notes for 21 September. Token counts and scores come from Artificial Analysis, an independent testing firm, which published its results the same day.
What changed in Grok 4.7?
Grok 4.7 is a new, larger base model with a longer training run, and the API around it did not change. xAI's announcement says training focused on hard problems that take many hours. It also says the model checks its own work better and handles long context better.
Your code barely changes. The model name goes from grok-4.6 to grok-4.7. Both models take text and images and return text. Both default to the "high" reasoning setting. The new model's knowledge runs to May 2026, per its model page.
Two details are new. Grok 4.7 Fast is the same model at twice the speed and twice the price, and it is sold only inside Cursor and Grok Build. On the Responses API, the new model always returns its encrypted reasoning, even when you do not ask for it.
Is Grok 4.7 cheaper in practice?
No. The price per token is identical, but the new model writes far more tokens per task, so a finished job costs more.
On Artificial Analysis's test tasks at the xhigh setting, Grok 4.7 wrote about 81,000 output tokens per task. Grok 4.6 wrote about 38,000. That is a little over twice the output.
At $6 per million output tokens, that moves output cost from about 23 cents a task to about 49 cents. Input and cache add more on top. VentureBeat reported the testers' full cost per task as about $3.74 on xhigh and $2.73 on high.
The pricing cliff is unchanged. Once a prompt goes over 200K tokens, xAI's release notes list double rates for input, cached input and output alike, so the whole request costs more, not just the tokens past the line. A model built for long jobs will cross that line more often. Our post on cost per solved task shows how to price this: divide the cost of an attempt by how often it succeeds.
What do the independent tests say vs xAI's chart?
Independent tests show a real gain, but a much smaller one than xAI's chart. xAI ran Grok 4.7 on xhigh and Grok 4.6 on high, so its chart pits a model allowed to think longer against one that was not. These are xAI's own numbers:
| Task (xAI's own claim) | Grok 4.7, xhigh | Grok 4.6, high |
|---|---|---|
| Finishing multi-step jobs in a command-line terminal | 37.6% | 20.3% |
| Long, multi-file software engineering tasks | 71.0% (on high) | 65.2% |
| Everyday coding help inside an editor | 46.3% | 40.4% |
| Professional health questions | 56.7% | 48.5% |
| Legal agent tasks | 19.6% | 15.8% |
| Real professional work, rated against experts (Elo) | 1,695 | 1,605 |
The terminal test shows the gap best. On the same test version, Artificial Analysis measured a gain of 4.5 points, with the new model at about 26%, per VentureBeat. That is a gain of under five points, where xAI's chart shows seventeen. xAI does not say how it set up its test, so we cannot say what explains the rest.
The testers' index tells the same story. The new model gained two points over the old one, to 46. On launch day, the leading closed models from Anthropic and OpenAI scored 53 on that index, per The Decoder. A day later, Claude Opus 5.5 scored 58.
The gains sit in agentic work. Per Artificial Analysis, the coding-agent score rose from 47 to 56. Wrong answers on its made-up-facts test fell from 34% to 29%. The new model slipped slightly on reading long documents and on automating business workflows.
Read the version number on every score. Version 4.0 of the terminal test is a different test from version 3.0. Our earlier Grok 4.6 comparison quoted a version 3.0 score, so it cannot sit beside any number here. Our post on benchmark version drift explains why.
Should you upgrade from Grok 4.6 to Grok 4.7?
Yes, for agentic coding and long analyst-style work, because that is where independent tests found the gain. The switch is a one-word change at the same price per token. The bill is what moves.
Four checks before you switch a production workload:
- Run it on high first. High is the default, and xAI's headline numbers come from xhigh.
- Do not assume xhigh is worth it. The two top settings scored about the same on the independent index, per The Decoder.
- Measure tokens per finished job, not per request. Run the same 20 to 50 real tasks on both models and compare cost per success.
- Keep Grok 4.6 for long-document reading and workflow automation until your own tests say otherwise. Those are the two areas where the new model scored slightly lower.
We have not run either Grok model on our own decision-task benchmark, so this post quotes no internal numbers for them.
Can you run Grok 4.7 in Canada?
No. Grok 4.7 is a closed model, so xAI's own servers process your prompts, and xAI offers no Canadian endpoint. xAI's regions page says its global endpoint may route requests between regions, so "the processing location is not guaranteed". Its only regional option is in the US, at 10% more per token.
For inference in Canada at this level, the route is an open-weight model on dedicated GPUs in Canada. That means a Canadian region of your own cloud account, or hardware you own.
The largest open-weight option today is Kimi K3, from Moonshot AI, with weights on Hugging Face since July. We have not seen an independent test that puts it beside Grok 4.7 on the same setup, so run your own jobs on both before you choose. Our Kimi K3 vs Claude Opus 4.8 comparison covers what it does well.
Kimi K3 is large. The model card lists 2.8 trillion parameters, 104 billion active per token, and a 1M-token context. The download comes to about 1.6 terabytes. The card names serving software such as vLLM and SGLang but no GPU count, so your workload sets the hardware size. It ships under Moonshot's own Kimi K3 License, so read the terms before you build a product on it.
The trade-offs from our Canadian residency guide apply. GPU capacity is scarcer in Canadian regions, you pay for idle GPUs, and you own the serving.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account or on-premises. Prompts and outputs stay in the country, and we build the workflow on top. Your team runs the work and the model takes part. See how we deploy.
If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.
For the wider open-weight field, see the best open-source coding models of 2026.
FAQ
Is Grok 4.7 better than Grok 4.6 for coding?
Yes, modestly. On finishing multi-step terminal jobs, Artificial Analysis measured a gain of under five points. Its coding-agent score rose nine points. xAI's own chart shows a much larger jump, but it compares Grok 4.7 on xhigh with Grok 4.6 on high. Expect a real but smaller improvement, and test on your own repository before you rely on either number.
Is Grok 4.7 more expensive than Grok 4.6?
Per token, no. Both models charge the same rates, and both double them once a prompt goes over 200K tokens. Per task, the new model usually costs more. In Artificial Analysis's tests it wrote a little over twice as many output tokens for the same task, so the same job produces a bigger bill even though nothing on the price list moved.
What is the context window of Grok 4.7?
500,000 tokens, the same as Grok 4.6. The catch is the price step at 200,000. Once a prompt goes over that size, xAI charges double for input and output alike, so the whole request costs more, not only the tokens past the line. Plan long-context jobs around that threshold rather than the 500,000 ceiling.
Can I download Grok 4.7 and run it myself?
No. xAI has not released weights for Grok 4.7 or Grok 4.6. You reach them through xAI's API, Cursor, Grok Build or model routers. If you need to run a model on your own hardware or in your own cloud account, you need an open-weight model such as Kimi K3.
Is Grok 5 out?
No. Grok 4.7, added on 21 September 2026, is the newest model in xAI's API release notes. xAI has published no release date for Grok 5, so plan around Grok 4.7 for now.
Sources
- xAI, Grok 4.7 announcement, 21 Sep 2026
- xAI, API release notes, entries for Grok 4.6 (12 Aug 2026) and Grok 4.7 (21 Sep 2026)
- xAI, Grok 4.7 model page, Sep 2026
- xAI, Regional endpoints, Sep 2026
- Artificial Analysis, Benchmarking Grok 4.7, 21 Sep 2026 (Intelligence Index, Coding Agent Index, AA-Omniscience, AA-LCR, AutomationBench-AA)
- Artificial Analysis, Claude Opus 5.5 takes the top spot, 22 Sep 2026
- The Decoder, xAI launches Grok 4.7 at bargain prices, 21 Sep 2026 (Intelligence Index v4.3.2)
- VentureBeat, Grok 4.7 pairs coding gains with the same pricing, 21 Sep 2026
- Moonshot AI, Kimi K3 model card
- Benchmarks named in plain words above: finishing multi-step jobs in a terminal is Terminal-Bench 4.0; long multi-file engineering is DeepSWE v1.1; everyday coding help in an editor is CursorBench 4.0; professional health questions is HealthBench Professional; legal agent tasks is the Harvey Legal Agent Benchmark; real professional work is GDPval
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project