Maximum benchmark performance.
Claude Fable 5 High
AnthropicLeads the supplied Arena document comparison. Prioritize it when document benchmark performance matters most.
Choose your workload
Models for extracting, summarizing and reasoning over document content.
Maximum benchmark performance.
Leads the supplied Arena document comparison. Prioritize it when document benchmark performance matters most.
A smarter performance-to-cost balance.
Kimi is the current supplied value recommendation for documents, with lower input and output token reference prices.
Arena · supplied snapshot·Supplied Arena snapshot · date pending·How we compare
Input and output savings use the separate supplied token rates. Combined cost uses 1M combined tokens: 75% input / 25% output; your workload mix changes the combined savings. Arena scores and prices are from the supplied snapshot. The leaderboard source and snapshot date still need confirmation.
The same workload, with the details that help you choose.
| Comparison | Claude Fable 5 HighBest Performance | Kimi K2.6Best Value |
|---|---|---|
| Arena score | 1496 Supplied Arena snapshot · date pending | 1451 Supplied Arena snapshot · date pending |
| Relative score | Performance reference | 3.0% lower Arena score |
| Input / 1M tokens | $10.00 | $0.95 |
| Output / 1M tokens | $50.00 | $4.00 |
| Arena difference | Performance reference | -45 / -3.0% |
| Cost difference | Price reference | 90.5% lower input cost · 92% lower output cost |
| Supported modalities | text, image, document → text | text, document → text |
| Main reason to choose | Choose this when maximum benchmark performance matters more than inference cost. | Consider this for production workloads where the reference price matters. Evaluate the benchmark gap against your own workload. |
| Best fit | Maximum benchmark performance for reasoning, coding, visual understanding and documents | Document extraction, summaries and document reasoning at volume |
| Billing details | Supplied reference prices per 1M input and output tokens. No additional percentage discount is applied. | Supplied reference prices per 1M input and output tokens. No additional percentage discount is applied. |
Input and output savings use the separate supplied token rates. Combined cost uses 1M combined tokens: 75% input / 25% output; your workload mix changes the combined savings.
Implementation
Copy a model test, replace the API key with your own, and run.
Copy the Python or cURL test for Kimi K2.6. Replace the API-key placeholder with your own key and run.
Document text test · The sample document is included in the request as text.
import json
from urllib.error import HTTPError
from urllib.request import Request, urlopen
API_KEY = "YOUR_ROTEN_API_KEY"
def send(url, payload=None):
data = json.dumps(payload).encode("utf-8") if payload is not None else None
request = Request(url, data=data, headers={"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"})
try:
with urlopen(request, timeout=180) as response:
return json.load(response)
except HTTPError as error:
raise SystemExit(f"HTTP {error.code}: {error.read().decode('utf-8')}")
payload = json.loads(r'''{
"messages": [
{
"role": "user",
"content": "Summarize this sample project memo in two bullets.\n\nThe product team is evaluating an AI API for customer-support replies. The trial will compare response usefulness and request cost. Engineering will record the results before selecting a model."
}
],
"max_tokens": 256,
"stream": false,
"model": "kimi-k2.6"
}''')
result = send("https://api.roten.ai/v1/chat/completions", payload)
print(json.dumps(result, indent=2, ensure_ascii=False))API_KEY='YOUR_ROTEN_API_KEY'
curl --fail-with-body --silent --show-error 'https://api.roten.ai/v1/chat/completions' \
-H "Authorization: Bearer $API_KEY" \
-H 'Content-Type: application/json' \
--data '{
"messages": [
{
"role": "user",
"content": "Summarize this sample project memo in two bullets.\n\nThe product team is evaluating an AI API for customer-support replies. The trial will compare response usefulness and request cost. Engineering will record the results before selecting a model."
}
],
"max_tokens": 256,
"stream": false,
"model": "kimi-k2.6"
}'Python tests use the standard library. Video cURL tests also use Python 3 to read task responses.
A clear reference point
Recommendations combine the supplied Arena scores with the supplied model prices. A benchmark score is one signal for choosing a model for your workload.
Arena · supplied snapshot·Supplied Arena snapshot · date pending·How we compare
Built for production
Talk to us about volume pricing, model selection and migration.