By Evan Vega
Moonshot AI released Kimi K3.1 Pro on 23 October 2026, describing it as its “most capable 3T-class open-weight
model to date.” On the independent Artificial Analysis Intelligence Index it scored 55.1, fourth of 215 models, ahead of
GPT-6 Astra and Gemini 4 Argon. But at launch it is available only through Moonshot’s own app and API: the weights are
promised “by Nov 4, 2026,” and the licence that will govern them has not been published.
Novel Cognition checked the release against Moonshot’s live API, independent evaluations and the first day of
community reaction. The full analysis, with sources and our own test conditions, is at
k31pro.novcog.us.com.
Moonshot AI’s Kimi K3.1 Pro ranks fourth of 215 models on the Artificial Analysis index and topped an independent Terminal-Bench 4.0 run, but on launch day it is API-only with weights promised by 4 No
Moonshot says it traced rising vulnerability-research capability to cyber-related content in general training data
and removed it before release. On its internal xdev-bench-hard test, the unfiltered checkpoint solved 90.5% of problems
and the shipped model 81.0% — still well above K3’s 33.3%. Artificial Analysis’s independent Cyber Index, which measures finding and fixing vulnerabilities, rates it
the highest publicly accessible model, second only to a restricted-access OpenAI system, partly because
rivals’ declined tasks score zero; meanwhile Moonshot’s public
cyber benchmarks place it behind the leading proprietary models. Moonshot’s stated principle is not to release open
weights that “extend the absolute frontier of cyber capability.” Whether the capability can be fine-tuned back once the
weights are public is not addressed. Details: the cyber cut.
Moonshot’s own table puts K3.1 Pro at 65.7% on Terminal-Bench 4.0, behind Claude Opus 5.5 (66.4%, Anthropic’s
figure) and ahead of GPT-6 Astra (58.2%). Nearly every coding score in that table was run inside Anthropic’s Claude Code
harness, which drew criticism. Artificial Analysis then ran the model in a neutral harness and measured 67.2%, the top
result on its board, so the harness did not inflate the score on that test. Other independent results are more modest:
14th on LiveBench 2.0, a preliminary 8th on Arena WebDev, and one third-party coding test that found it no better than
K3. Moonshot also stopped reporting at least nine benchmarks it had published for K3. Details:
benchmarks.
List prices are $2 per million input tokens and $10 per million output, down from K3’s $3 and $15 — half of
Opus 5.5 and a fifth of GPT-6 Astra per token. Measured per task, the cut mostly disappears: the model produced 150
million output tokens on Artificial Analysis’s suite against an 87 million median, and costs about $2 per task, roughly
what K3 cost. At its capability level it is priced like Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon. The
Highspeed variant, at twice the price, is advertised as “2-3×” faster; Artificial Analysis measured 1.5 to 1.8
times. Our own test, the day after launch, found about 1.7 times the throughput. Details:
speed and cost.
If the weights ship as promised, K3.1 Pro becomes the strongest open-weight model by roughly nine points on the
Artificial Analysis index, narrowing the gap to the best closed model from twelve points to three. Until then it is
served by Moonshot alone. Community reaction is documented at reactions
and the market picture at ecosystem; every source is listed at
sources.
This analysis was produced under the Hidden State Drift method, which
separates vendor claims from independent measurement. The ten pre-publication checks behind it are published at
hiddenstatedrift.com/verify. Novel Cognition is not affiliated with
Moonshot AI or any lab discussed.
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily Colorado News
