Grok 4.20 Beta
VS
Kimi K2.6
Grok 4.20 Beta vs Kimi K2.6: which wins at real work?
22 task areas · same graded test runs · rank comparison only, so 0–100 and Elo collections never mix raw scores.
Kimi K2.6 wins 21 of 22 task areas we tested; Grok 4.20 Beta takes 1. Kimi K2.6 costs 2.0× less per token ($4.07 vs $8 per 1M).
Kimi K2.6 costs 2.0× less per token ($4.07 vs $8 per 1M).
Task by task
| Task area | Grok 4.20 Beta | Kimi K2.6 | Winner |
|---|---|---|---|
| Summarization & Meeting Notes |
#95
/ 110
Excellent
|
#3
/ 110
Excellent
|
Kimi K2.6 |
| Legal & HR |
#106
/ 110
Strong
|
#22
/ 110
Excellent
|
Kimi K2.6 |
| Creative & Comedy | #92 / 110 | #11 / 110 | Kimi K2.6 |
| Content & Brand |
#99
/ 124
Usable
|
#22
/ 124
Strong
|
Kimi K2.6 |
| RAG, Safety & Grounding |
#93
/ 113
Strong
|
#18
/ 113
Excellent
|
Kimi K2.6 |
| Coding |
#77
/ 115
Strong
|
#5
/ 115
Excellent
|
Kimi K2.6 |
| Customer Support |
#72
/ 113
Usable
|
#4
/ 113
Strong
|
Kimi K2.6 |
| Data & Analytics |
#91
/ 110
Excellent
|
#23
/ 110
Excellent
|
Kimi K2.6 |
| Translation & Localization |
#90
/ 110
Strong
|
#22
/ 110
Excellent
|
Kimi K2.6 |
| Research & Competitive Analysis |
#86
/ 110
Needs editing
|
#19
/ 110
Excellent
|
Kimi K2.6 |
| Knowledge & Docs |
#86
/ 110
Usable
|
#22
/ 110
Strong
|
Kimi K2.6 |
| Executive Assistant |
#69
/ 112
Strong
|
#8
/ 112
Strong
|
Kimi K2.6 |
| Presentations & Decks |
#56
/ 110
Strong
|
#2
/ 110
Excellent
|
Kimi K2.6 |
| Investor & Pitch |
#43
/ 66
Usable
|
#1
/ 66
Strong
|
Kimi K2.6 |
| Landing Pages |
#60
/ 72
Usable
|
#21
/ 72
Strong
|
Kimi K2.6 |
| Chef / Home Cooking |
#38
/ 126
Usable
|
#7
/ 126
Strong
|
Kimi K2.6 |
| Sales |
#46
/ 110
Usable
|
#15
/ 110
Strong
|
Kimi K2.6 |
| Training & Education |
#76
/ 110
Strong
|
#45
/ 110
Excellent
|
Kimi K2.6 |
| AI Strategy |
#49
/ 126
Strong
|
#19
/ 126
Strong
|
Kimi K2.6 |
| Structured Output |
#43
/ 113
Excellent
|
#14
/ 113
Excellent
|
Kimi K2.6 |
| Frontend & Landing Pages |
#32
/ 109
Needs editing
|
#8
/ 109
Usable
|
Kimi K2.6 |
| Product & Project Management |
#31
/ 110
Excellent
|
#53
/ 110
Strong
|
Grok 4.20 Beta |
Rank = position among every model config we tested in that task area (lower is better). Sorted by biggest gap first.
Same task, both models — judged
Both models answered the same test case; an independent judge graded each. Quotes are the judge's actual rationale.
Legal & HR
Non-compete enforceability (Tradewinds staff) (Contract Clause Review)
“The model failed to include appropriate caveats by asserting a definitive answer and omitting the recommendation to consult a licensed attorney.”
“The model perfectly executed the prompt's instructions. It provided a highly accurate, general overview of non-compete enforceability principles without fabricating any statutes or cases. It correctly flagged the specific risks of the clause (2-year duration, nationwide scope, broad industry ban) and included excellent caveats regarding jurisdiction-dependence and the need for licensed legal counsel.”
Content & Brand
Technical blog intro with no hype (Brief Adherence Test)
“The response failed the hard constraint for word count (130-180 words).”
“The response perfectly follows all instructions and constraints, including the strict word count. The tone is expertly tailored to engineers, and the concrete failure mode is highly realistic, specific, and useful. It is difficult to improve upon this output.”
Data & Analytics
Percentage points vs percent (Northwind) (Metric Calculation)
“The model fails to provide the 20% relative increase and commits a severe analytical error by explicitly labeling the absolute change ('2 percentage points') as the 'relative change'. It hits the exact failure mode of not providing both framings and failing to distinguish between them correctly.”
“The model correctly calculates and distinguishes between the absolute change (2 percentage points) and the relative change (20% relative gain). It also provides excellent business context by noting that the overall rate is still very low (nearly 9 in 10 pickups missing the window).”
Weighted conversion rate (Cedar & Sage) (Metric Calculation)
“The model correctly pools the numerators and denominators to get 139 / 1,200 and calculates 11.5833...%, but then bizarrely rounds this to 12.3% as its final answer. Because the final answer is incorrect, correctness is scored 0.”
“The model correctly pools the numerators and denominators to calculate the overall conversion rate, arriving at the correct answer of 11.58%. It provides a clear step-by-step explanation and correctly notes why an unweighted average would be wrong, adding excellent business context.”
Frequently asked
Is Grok 4.20 Beta better than Kimi K2.6?
Across 22 task areas we benchmarked, Kimi K2.6 ranks higher in 21 and Grok 4.20 Beta in 1.
Which is cheaper, Grok 4.20 Beta or Kimi K2.6?
Kimi K2.6 costs 2.0× less per token ($4.07 vs $8 per 1M).
What is Grok 4.20 Beta better at?
Grok 4.20 Beta out-ranks Kimi K2.6 at Product & Project Management.
What is Kimi K2.6 better at?
Kimi K2.6 out-ranks Grok 4.20 Beta at Summarization & Meeting Notes, Legal & HR, Creative & Comedy.
More comparisons
This page is Spring Prompt, running in public
We just did this for every model. Do it for your prompt.
The rankings above come from running real tasks through real models and scoring every output. Spring Prompt is that same engine — pointed at your prompt, your test cases, and your definition of good.
- Generate test cases from your prompt — no eval set required to start.
- Compare models side by side with quality, cost and latency in one matrix.
- Optimise the winner until the scores say it's ready to ship.
Prompt × model results
12 test cases · 3 evals