Kimi K3.1 Pro British English Test: How It Compares to Claude Opus 5.5, GPT-6 Astra and DeepSeek
I tested Kimi K3.1 Pro across 12 UK-specific tasks via OpenRouter. Scoring 70.5/72, it matched the leaders in formal letters, UK/US localisation and NHS number code, dropping only 1.5 points on wordplay and brief code comments.

👁 112 reads
I tested Kimi K3.1 Pro in British English across 12 purpose-built local tasks on 24 October 2026, comparing it via OpenRouter against Claude Opus 5.5, GPT-6 Astra, DeepSeek V4.1 Flash 1017, Qwen3.8 Max 0902, Gemini 4 Argon and GLM-5.3. Scoring 70.5 out of 72 points, Kimi K3.1 Pro finished fourth on paper, but just one point behind the highest-scoring model in a single-run evaluation where 1-2 point gaps are effectively a tie.
- Score and rank: Kimi K3.1 Pro scored 70.5 / 72 (4th out of 7 models), trailing Claude Opus 5.5 (71.5) by 1.0 point and tied runners-up DeepSeek V4.1 Flash 1017 and GPT-6 Astra (71.0) by 0.5 points.
- Strongest UK tasks: Full marks (6.0 / 6.0) in 10 of 12 categories, including a formal council parking appeal, bidirectional UK-to-US localisation, UK VAT calculations (20/5/0%), ~30k-token context retrieval and proofreading.
- Only points dropped: Lost 1.0 point on task b (idioms and wordplay) for a forced original pun on the word “pitch”, and 0.5 points on task g (NHS number validation code) because one judge found the docstring and comments too brief.
- Speed and API cost: Kimi cost $0.151 for the 12 British English tasks at a median 84 tok/s (133 s total) on the
moonshotai/mxfp4endpoint, whereas DeepSeek V4.1 Flash 1017 delivered 71.0 points for $0.023 at 386 tok/s. - Scope warning: All results come from a single API trial per task; they do not represent the Kimi web interface and cannot be generalised to other languages or workloads.
British English Ranking, Speed and API Cost
Every model completed the same 12 British English tasks written from scratch for the UK market. Each task was worth 0 to 6 points (three criteria scored 0-2), giving a maximum of 72 points. Because I ran one trial per task without variance testing, score differences of 1 to 2 points across a language set are effectively a tie.
| Model (OpenRouter Endpoint) | Score / 72 | Cost (USD) | Median tok/s (Total s) |
|---|---|---|---|
Claude Opus 5.5 (anthropic) |
71.5 / 72 | $0.384 | 172 (99 s) |
DeepSeek V4.1 Flash 1017 (together) |
71.0 / 72 | $0.023 | 386 (38 s) |
GPT-6 Astra (openai/flex) |
71.0 / 72 | $0.252 | 195 (46 s) |
Kimi K3.1 Pro (moonshotai/mxfp4) |
70.5 / 72 | $0.151 | 84 (133 s) |
Qwen3.8 Max 0902 (alibaba) |
69.0 / 72 | $0.150 | 56 (278 s) |
Gemini 4 Argon (google-ai-studio) |
66.0 / 72 | $0.105 | 126 (67 s) |
GLM-5.3 (z-ai/fp8) |
62.5 / 72 | $0.060 | 165 (63 s) |
In this British English evaluation, Kimi K3.1 Pro achieved its narrowest gap to the leaders across my entire 10-language study. At $0.151 for the 12 tasks, its local run cost less than Claude Opus 5.5 ($0.384) and GPT-6 Astra ($0.252 on the half-price flex tier), though DeepSeek V4.1 Flash 1017 was both faster and over six times cheaper at $0.023.
Timing figures require caution. Requests ran in parallel (up to 20 at once), and GPT-6 Astra was routed through the discounted openai/flex endpoint outside the Polish set, making its latency not directly comparable to standard priority tiers. Furthermore, I only tested the moonshotai/highspeed endpoint in Polish, so I cannot infer Highspeed latency or behaviour for British English.
Task-by-Task Breakdown: Kimi K3.1 Pro vs Top Rivals
Structured and automated tasks proved largely saturated in British English: almost every model scored 6.0 / 6.0 on UK VAT maths (task e, checking 20%, 5% and 0% rates per gov.uk), local dates and numbers (f), logic grid puzzles (h), JSON extraction with discount and null traps (i), UK factual queries (j) and finding three needles inside roughly 30k tokens of residents’ association minutes (k). Separation came down to linguistic nuance and code documentation.
| Task | Kimi K3.1 Pro | Best Rival | Top Scoring Rivals |
|---|---|---|---|
| a - Formal letter / politeness | 6.0 | 6.0 | DeepSeek V4.1 Flash, GPT-6 Astra |
| b - Idioms and wordplay | 5.0 | 6.0 | GPT-6 Astra, Claude Opus 5.5 |
| c - Summary with local context | 6.0 | 6.0 | DeepSeek V4.1 Flash, GPT-6 Astra, Opus 5.5 |
| d - UK <-> US localisation traps | 6.0 | 6.0 | DeepSeek, Gemini, GPT-6, Opus, Qwen |
| e - Invoice: currency and UK VAT | 6.0 | 6.0 | All 6 rivals |
| f - Local date and number formats | 6.0 | 6.0 | All 6 rivals |
| g - NHS validation code & comments | 5.5 | 6.0 | DeepSeek, Gemini, Opus 5.5, Qwen |
| h - Logic puzzle | 6.0 | 6.0 | All 6 rivals |
| i - JSON extraction from text | 6.0 | 6.0 | All 6 rivals |
| j - Local UK fact | 6.0 | 6.0 | All 6 rivals |
| k - Long context (~30k tokens) | 6.0 | 6.0 | All 6 rivals |
| l - Proofreading and error correction | 6.0 | 6.0 | DeepSeek, Gemini, GPT-6, Opus, Qwen |
Notice that in task a (formal letter), the average score across rivals was only 4.6 / 6.0, yet Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017 and GPT-6 Astra achieved a perfect 6.0. Kimi also beat the rival average in task c (6.0 vs 5.6), task d (6.0 vs 5.8) and task l (6.0 vs 5.7).
Exact Excerpts from Kimi K3.1 Pro’s British English Outputs
To see how Kimi K3.1 Pro handled UK conventions and where it lost its 1.5 points, here are verbatim excerpts from my test logs alongside their exact scores.
Task a: Formal PCN Challenge to Nottingham City Council (Score: 6.0 / 6.0)
This task required challenging a Penalty Charge Notice using proper British formal letter layout, salutations, factual grounds and enclosure references within a strict word limit. Both AI judges (GPT-6.1 Sol and Gemini 4 Argon) awarded Kimi 2/2 on all three criteria:
Mr Daniel Whitfield
14 Mansfield Road
Nottingham
NG5 2BT24 October 2026
Parking Services
Nottingham City CouncilDear Sir or Madam
**Re: Penalty Charge Notice NT48201937**
I am writing to challenge the above Penalty Charge Notice, issued on 9 October 2026 in Arboretum Street for allegedly being “parked in a residents’ bay without a valid permit
Task d: UK-to-US and US-to-UK Localisation Traps (Score: 6.0 / 6.0)
For the UK test, task d replaced foreign-language translation with bidirectional localisation between British and American English, packed with false friends (floor numbering, date ordering, everyday nouns and idioms). Kimi scored 6.0 / 6.0 for converting both passages naturally:
1: Our apartment is on the second floor, but the elevator’s broken again, so I’m waiting in line at the drugstore with my cellphone in the trunk of the car. Want to get takeout tonight? It’s on me. The game was awful, though-they couldn’t hit the broad side of a barn from the edge of the field.
2: I’ll be on holiday for two weeks starting 3 December 2026. Please post the cheque to my flat on the first floor, and get some chips and biscuits for the kids. The tap’s leaking and the bin’s full-can you call the caretaker
Task b: British Idioms and Wordplay (Score: 5.0 / 6.0)
Kimi accurately defined and exemplified four British idioms and explained the double meaning of “interest” in a banking joke (scoring 2/2 on the first two criteria):
c) All fur coat and no knickers** - impressive or glamorous on the surface, but lacking real substance or quality.
*Example:* “The restaurant looks luxurious, but it’s all fur coat and no knickers-the food is distinctly average.”**d) To be a bit of a curate’s egg** - to have both good and bad parts; to be mixed in quality.
*Example:* “The hotel was a bit of a curate’s egg: the room was lovely, but the service was dreadful
However, both judges gave Kimi 1/2 on the third criterion, where models had to invent an original pun using the word “pitch”. They noted that Kimi juxtaposed a business presentation and a sports pitch separately rather than landing a true double-meaning punchline, whereas GPT-6 Astra and Claude Opus 5.5 scored 6.0 / 6.0.
Task g: NHS Number Validator in Python (Score: 5.5 / 6.0)
Models had to write a Python function validating a 10-digit NHS number (including the 3-3-4 spaced format and modulus 11 check digit) with clear British English documentation. In automated execution (`python -I`), Kimi passed 5/5 valid and rejected 6/6 invalid hidden test cases:
import re
_NHS_NUMBER_PATTERN = re.compile(
r"(?:[0-9]{10}|[0-9]{3} [0-9]{3} [0-9]{4})"
)
def is_valid_nhs_number(value: str) -> bool:
"""Return True if *value* is a valid NHS number, otherwise False.
The number may contain exactly ten digits or use the usual 3-3-4
grouping with single spaces. Non-string and otherwise invalid input
returns False rather than raising an exception
While GPT-6.1 Sol awarded 2/2 for the English comments, Gemini 4 Argon gave 1/2 because the docstring omitted the modulus 11 algorithm explanation and reduced inline comments to two brief remarks, resulting in a averaged 5.5 / 6.0 score.
Methodology, Judging Rules and Limitations
To keep the comparison transparent, several methodological constraints from my 24 October 2026 run must be kept in mind:
- API-only evaluation: All models were called via OpenRouter with
provider.onlyenforced (no fallbacks),reasoning_effort: "medium"andmax_tokens: 16000. I did not test Kimi through its consumer web UI, and API endpoint results cannot prove identical behaviour in the web application. - Cross-family blind judging: Subjective criteria (tasks a, b, c, d, plus comment quality in g and style in l) were graded blindly by GPT-6.1 Sol and Gemini 4 Argon. To prevent self-preference, GPT-6 Astra answers were scored only by Gemini 4 Argon, Gemini 4 Argon answers were scored only by GPT-6.1 Sol, and all other models (including Kimi K3.1 Pro and Claude Opus 5.5) received the average of both judges. Across 975 criterion pairs in the full study, the two judges gave the exact same score 89% of the time.
- Benchmark design disclosure: The task prompts and grading rubrics were generated with the assistance of a Claude Opus 5.5 agent. Although most points are automated and subjective criteria were scored by non-Anthropic models, this is not an independent human benchmark and carries a potential stylistic bias in favour of Opus 5.5.
- Endpoint stability across the wider study: In British English, Kimi K3.1 Pro answered every prompt properly. However, across the full 120-task, 10-language test on
moonshotai/mxfp4, Kimi produced 8 completely off-topic responses (in Italian, German, Turkish and Japanese) that counted as 0 in the official rankings. Diagnostic reruns outside the ranking suggested an endpoint serving issue rather than a core capability gap, though two trials cannot prove this conclusively. No method corrections were required for the British English task set (unlike minor puzzle or regex adjustments made in German, Italian, Japanese, Dutch and French).
For the full multi-language breakdown and cross-market scores, see my Kimi K3.1 Pro benchmark hub. If you want to try the platform directly, you can access it via this Kimi referral link (affiliate link - I may earn a commission at no extra cost to you).
Frequently Asked Questions
Did Kimi K3.1 Pro beat GPT-6 Astra and Claude Opus 5.5 in British English?
On raw points, Kimi K3.1 Pro scored 70.5 / 72, finishing 0.5 points behind GPT-6 Astra and DeepSeek V4.1 Flash 1017 (both 71.0) and 1.0 point behind Claude Opus 5.5 (71.5). Because this was a single-trial test per task, a 0.5 to 1.0 point difference is effectively a tie in practical quality.
Where did Kimi K3.1 Pro lose points in the UK test?
It lost points in only two places: 1.0 point in task b (idioms and wordplay) because both judges found its original pun on the word “pitch” slightly forced, and 0.5 points in task g (NHS number validation code) because one of the two judges considered its Python docstring too brief regarding the modulus 11 formula.
Did Kimi K3.1 Pro suffer any off-topic API failures in British English?
No. While the moonshotai/mxfp4 endpoint returned 8 off-topic responses across Italian, German, Turkish and Japanese in the wider 120-task test, all 12 British English responses were on-topic and scored between 5.0 and 6.0 out of 6.0.
Which model was the most cost-effective for British English API tasks?
In my local 12-task run, DeepSeek V4.1 Flash 1017 (via Together) scored 71.0 / 72 at a total cost of $0.023 and a median speed of 386 tok/s. Kimi K3.1 Pro scored 70.5 / 72 for $0.151 at 84 tok/s.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
All benchmark scores, latencies, costs and text excerpts come from my own OpenRouter API test executed on 24 October 2026. For general background on the model and vendor documentation, see:
More on this from QUANTSLANT
ai tools·Convert PDF to Excel and check the table before using the numbersImport a PDF table through Power Query, spot repeated headers and broken rows, and validate totals before relying on the Excel output.

