Grok 4.7: Full Benchmark Table and Pricing at Launch
2 USD per million input tokens with a new base model. Seven benchmarks, three wins, triple advantage in legal work. Launched September 21, 2026.
👁 123 reads
Grok 4.7 launched on September 21, 2026, and the most interesting number isn’t in any benchmark. The model got a new, larger base model and longer training, yet it’s served at the same price as Grok 4.6: 2 dollars per million input tokens and 6 per million output tokens. For comparison, top-tier competitors charge 10 and 50.
Verdict in One Paragraph
This isn’t a model that wins every benchmark - it’s a model that wins on performance-to-price ratio. In coding it trails the strongest competition, but costs five times less on input and over eight times less on output. However, there are three areas where Grok 4.7 wins outright: electronics, legal work, and safety. In the legal test, the difference is threefold in its favor.
Pricing Alongside Results - Full Table
The maker did something rarely done: they put pricing in the same table as benchmarks. Below is the complete set, as of launch day.
| Grok 4.7 xHigh | Grok 4.6 High | GPT-5.6 Sol Max | Fable 5.1 Max | |
|---|---|---|---|---|
| Input (per million tokens) | 2 USD | 2 USD | 4 USD | 10 USD |
| Output (per million tokens) | 6 USD | 6 USD | 20 USD | 50 USD |
| CursorBench 4.0 - coding | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 | 71.0%* | 65.2% | 72.7% | 70.0% |
| EEBench - electronics | 64.0% | 53.0% | 39.4% | 56.4% |
| AA Briefcase - office work | 1,657 | 1,546 | 1,487 | 1,678 |
| Terminal-Bench 4.0 | 38.0% | 20.3% | 37.3% | 57.9% |
| Harvey - legal work | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% |
The asterisk by DeepSWE indicates a result in high-effort mode.
Three Things This Table Tells You
1. It’s Not the Best at Coding and Nobody’s Hiding It
On CursorBench 4.0, Grok 4.7 scores 46.3% versus a competitor’s 51.8%. On Terminal-Bench the gap is much larger: 38.0% versus 57.9%. These aren’t results you can gloss over - and the maker doesn’t gloss over them, because they published them alongside their own.
What makes a difference, though: this same Grok 4.7 jumped from Grok 4.6’s 20.3% to 38.0% on Terminal-Bench - nearly double, within a single generation.
2. Legal Work Is an Anomaly Across the Entire Table
Harvey Legal Agent Benchmark: 19.6% for Grok 4.7, 6.7% for the strongest competitor, and 2.5% for the next one. That’s a threefold and eightfold difference, not a few percentage points.
This is an opinion: when one model beats the rest by an order of magnitude on a narrow benchmark, it usually means it was deliberately trained for that type of task. This doesn’t diminish the result - but before making a purchase decision, you should check whether your work resembles what this particular test measures.
3. Electronics Without Competition
EEBench: 64.0% versus 56.4% and 39.4%. On electrical engineering tasks, the advantage is clear, and the gap versus one competitor is 25 percentage points.
What Changed Under the Hood
The maker lists four things, and all are consistent with what you see in the results:
- New, larger base model than in Grok 4.6
- Longer reinforcement training on a harder task set, weighted toward problems taking many hours
- Better self-verification and long-context handling
- Native understanding of its own agentic environment - hence the improvement in conversational and office tasks
Weighting training toward multi-hour tasks explains the jump on Terminal-Bench and in office work. These are exactly the tests where persistence matters, not brilliance in a single step.
Safety - and Here Are Concrete Numbers
This is usually the foggiest part of launch announcements, but here two numbers were given:
- 62.4% on a biological safety test conducted by an external entity - the highest score in the comparison
- 3.3% of risky dual-use queries passed on HackerBench, while rarely blocking legitimate security work
That second number is more interesting than it looks. It’s easy to build a model that refuses everything - harder to build one that refuses dangerous things but doesn’t interfere with a security professional’s normal work. Here, both are claimed at once.
What It Actually Costs
Let’s convert to something tangible. Assume a task consuming one million input tokens and 200 thousand output tokens - roughly a longer session working with large context.
| Model | Input | Output | Total |
|---|---|---|---|
| Grok 4.7 | 2.00 USD | 1.20 USD | 3.20 USD |
| GPT-5.6 Sol | 4.00 USD | 4.00 USD | 8.00 USD |
| Fable 5.1 | 10.00 USD | 10.00 USD | 20.00 USD |
The difference on one task is 16.80 USD versus the most expensive option - with a hundred such tasks per month, that becomes 1,680 USD.
A fast variant is also available: double output speed for double the price.
Where You Can Use It Starting Today
The model is available in Cursor and in the maker’s build environment, through API, in external coding tools, and through model routers and cloud platforms.
Who It’s For and Who It’s Not For
| Situation | Verdict |
|---|---|
| High query volume, budget matters | Yes. The price advantage is decisive here |
| Legal work, document analysis | Yes. Threefold advantage in the benchmark |
| Electrical engineering | Yes. Best result in the comparison |
| Long agentic tasks | Worth checking. Big jump versus predecessor |
| Most difficult programming tasks | No. Competition clearly higher, though more expensive |
| Medical applications | No. HealthBench favors competition |
Frequently Asked Questions
How Much Does Grok 4.7 Cost?
Starting at 2 USD per million input tokens and 6 USD per million output tokens - the same as Grok 4.6. The fast variant costs twice as much and has twice the output speed.
Is Grok 4.7 Better Than Grok 4.6?
In all seven published benchmarks, yes. The biggest jump is Terminal-Bench: from 20.3% to 38.0%. And the price hasn’t changed.
Is Grok 4.7 the Best Model for Coding?
No. On CursorBench 4.0 it scores 46.3% versus 51.8% for the strongest competitor. But it’s significantly cheaper, so the question is really: how much are you willing to pay extra for those five points.
Where Can I Use It?
In Cursor, in the maker’s build environment, through API, in external coding tools, and on cloud platforms and model routers.
Is There a Free Version?
The maker provides free access through their build environment - details on the project page.
Sources and Date Checked
All pricing, benchmark results, and description of changes come from the maker’s official announcement of September 21, 2026, accessed the same day: x.ai/news/grok-4-7. We did not conduct our own tests of this model - the numbers given are the maker’s declarations, including results for its competitors, which it published itself.
More on this from QUANTSLANT
ai tools·Ideogram 4.5 Image Editing: 10 Before & After Examples10 real edits in Ideogram 4.5: label changes, banner translation, menu prices, time of day, text removal. Precise mode vs regular mode tested.