Summary: Grok 4.7 in 30 Seconds
- Grok 4.7 landed on September 21, 2026, ten days later than Musk's own estimate.
- Price is flat again: $2 per 1M input tokens, $6 per 1M output. Third release in a row at the same tag.
- 500K context window, four reasoning tiers (low, medium, high, xhigh), May 2026 knowledge cutoff.
- It does not win at coding, but it takes electrical engineering and legal work by a wide margin.
Short version: it shipped, it is cheap, and it is very good at things that are not coding. Musk said "ten days" in early September and the date slipped twice. Now there is an actual price list and an actual benchmark table to work from.
Grok 4.7 announcement artwork. Source: SpaceXAI
What Is Grok 4.7?
SpaceXAI's new flagship for coding, agentic tasks and knowledge work. The spec sheet:
| Property | Value |
|---|---|
| Model ID | grok-4.7 |
| Context window | 500,000 tokens (about 375,000 words) |
| Input / output | Text + image / text only |
| Knowledge cutoff | May 2026 |
| Reasoning tiers | low, medium, high (default), xhigh |
| Batch API | Not supported |
| Rate limits | 150 requests/sec, 50M tokens/min |
No parameter count was published. The only architectural statement is that the base model is larger than the one behind Grok 4.6.
Four things changed in this release: a larger base model, a longer reinforcement learning run weighted toward tasks that take hours, better self-verification between steps, and native understanding of the company's own agent framework (Grok Bot).
How Does It Compare? 📊
The table SpaceXAI published. Grok 4.7 scores are at the xhigh tier, Grok 4.6 at high:
| Benchmark | Grok 4.7 xHigh | Grok 4.6 High | GPT-5.6 Sol Max | Fable 5.1 Max |
|---|---|---|---|---|
| Input / output price (1M) | $2 / $6 | $2 / $6 | $4 / $20 | $10 / $50 |
| CursorBench 4.0 | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 | 71.0%* | 65.2% | 72.7% | 70.0% |
| EEBench (electrical eng.) | 64.0% | 53.0% | 39.4% | 56.4% |
| Terminal-Bench 4.0 | 38.0% | 20.3% | 37.3% | 57.9% |
| Harvey Legal Agent | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% |
| GDPval (Elo) | 1695 | 1605 | 1542** | 1735 |
* Scored at the high effort tier. ** The GDPval row's score is GPT-6 Astra (max).
The published scores, charted. Data source: SpaceXAI
Grok 4.7 beats its own predecessor in every row and splits the decision against rivals. It wins on domain expertise: EEBench by 7.6 points over the runner-up, and the highest Harvey legal score in the table. It loses on software engineering, where Fable 5.1 is 20 points ahead on Terminal-Bench 4.0.
Reading these numbers
Every score above is vendor-published, including the competitor numbers. No independent evaluation of Grok 4.7 exists yet, and gaps of a few points usually sit inside published confidence intervals.
Pricing and the 200K Trap 💸
Per 1M tokens:
| Item | Request under 200K | Request over 200K |
|---|---|---|
| Input | $2.00 | $4.00 |
| Cached input | $0.50 | $1.00 |
| Output | $6.00 | $12.00 |
Identical to Grok 4.6. The same announcement lists $4/$20 for GPT-5.6 Sol and $10/$50 for Fable 5.1, so Grok 4.7's output tokens cost one eighth of Fable 5.1's.
The trap: once a request crosses 200K tokens the tariff doubles, and the higher rate applies to the entire request. A 199K prompt bills at $2; a 201K prompt bills all of it at $4. The US regional endpoint also bills at 1.1x, and there is no Batch discount.
Monthly bill for an agent pipeline burning 5M input and 1M output tokens a day:
| Model | Per month (30 days) |
|---|---|
| Grok 4.7 | $480 |
| GPT-5.6 Sol | $1,200 |
| Fable 5.1 | $3,000 |
Cheap tokens are not the same as cheap work: a model that needs more steps can erase its own advantage. Put your own volumes into our LLM cost calculator to compare.
Who Should Actually Switch?
- Hardware, circuits and engineering calculations: the EEBench gap is wide (64.0% against 39.4% for GPT-5.6 Sol), the strongest case for Grok 4.7.
- Contracts, compliance and legal drafting: 19.6% on Harvey is the highest score in the table.
-
Cost-sensitive, high-volume work: classification and summarisation on the
lowtier cut the bill hard. - Terminal and repo-scale coding agents: the table still points at Claude Fable 5.1.
On safety, the company says this release ships a rebuilt safeguard stack: top of LatchBio's biosafety benchmark at 62.4%, and only 3.3% of risky dual-use prompts getting through on HackerBench v0.3.
What Happened to the Pre-Launch Claims?
For weeks, "2.1 trillion parameters" and "trained on SpaceX rocket data" were everywhere. Neither appears in the official announcement:
| Claim | What the announcement says |
|---|---|
| 2.1T parameters | No number, only "larger base model" |
| SpaceX rocket and satellite data | Not mentioned |
| Will beat every model | First in three of seven benchmarks |
| Shipping September 12 | Shipped September 21 |
If the source is not x.ai/news or docs.x.ai, do not quote the number. We tracked the wait itself in a separate post.
Where Can You Use It?
In Cursor's model picker, in Grok Build (free to try), through the SpaceXAI API as grok-4.7, and via third-party harnesses and model routers. There is a fast variant at twice the speed and twice the price. No open weights.
The API is OpenAI-compatible:
from openai import OpenAI
client = OpenAI(api_key="YOUR_KEY", base_url="https://api.x.ai/v1")
response = client.chat.completions.create(
model="grok-4.7",
reasoning_effort="high", # low | medium | high | xhigh
messages=[{"role": "user", "content": "Describe Grok 4.7 in one sentence."}],
)
print(response.choices[0].message.content)
Dropping simple calls to low is the easiest saving available, since thinking tokens bill as output. On the Responses API, reasoning.encrypted_content always comes back even when include does not ask for it.
Frequently Asked Questions
Q: When was Grok 4.7 released?
A: September 21, 2026, shipped the same day through the API, Cursor and Grok Build.
Q: How much does Grok 4.7 cost?
A: $2 per 1M input tokens, $0.50 cached input and $6 per 1M output. Requests above 200K tokens double the whole tariff.
Q: How many parameters does Grok 4.7 have?
A: Undisclosed. The "2.1 trillion" figure circulating before launch has no official source.
Q: Is Grok 4.7 better than Claude and GPT?
A: It leads on electrical engineering and legal work; Claude Fable 5.1 is clearly ahead on coding benchmarks. On price-performance Grok wins outright.
Q: Is Grok 4.7 free?
A: API use is paid. You can try it free in Grok Build, with limited free use on grok.com and in the Grok app.
Take care... 🙂
AI-Generated Content Notice
This blog was entirely generated by artificial intelligence. While AI can help create content, it may still contain errors or biases. Please verify critical details before relying on them.
Originally published on Proje Defteri, where this post is kept up to date.
More: English posts and free browser tools.
Your support means a lot! ✨ Comment 💬, like 👍, and follow 🚀 for future posts!
Top comments (0)