Kimi K3.1 Pro: Artificial Analysis independent intelligence index, cost, and speed snapshot

One-sentence takeaway

In this page snapshot, Kimi K3.1 Pro has an Artificial Analysis Intelligence Index of 55 (#4/215), placing its quality signal near the top; however, 80.7 token/s, 150M index output tokens, and approximately $2.11 per task indicate that waiting time and output volume are the main costs of long-chain reasoning.

Use cases

• Tasks these results can help assess: Comparing overall intelligence, the 10 evaluation sub-scores, output speed, evaluation time, output volume, and task cost within the same Artificial Analysis snapshot.

• Tasks these results should not be extrapolated to: Production billing, success rates for all Chinese-language or multimodal tasks, a fixed SLA, or absolute rankings across different evaluation harnesses.

• Applicable model version: The Kimi K3.1 Pro reasoning version; the page notes that a non-reasoning variant may exist.

• Test environment or client: Artificial Analysis Intelligence Index; the speed and latency data are labeled as Kimi first-party provider data. The page charts also show No publicly available API, while the FAQ says there are 2 API providers; both are simply page states at the time of collection.

• Reasoning setting and parameters: The page specifications mark Reasoning: Yes; Prompt Options provide Single/Multiple (Multiple only supports 1k tokens) and input lengths of 1k, 10k, and 100k tokens. The options were not switched for this collection, and no rerun was performed.

Evaluation method

The Artificial Analysis page defines Intelligence Index v4.3.2 as the combined result of 10 evaluations:

1. AA-Briefcase v1.1

2. GDPval-AA v2.1

3. AutomationBench-AA

4. Terminal-Bench 4.0

5. SciCode

6. Humanity’s Last Exam

7. GDP.pdf

8. CritPt

9. AA-Omniscience

10. AA-LCR v1.1

The chart descriptions say that the evaluation results were independently measured by Artificial Analysis and run on dedicated hardware. The Intelligence Index chart on this page shows 27/709 models, while the AA-Omniscience chart shows 27/582 models. These denominators belong to the model sets for each chart on that day and should not be treated as fixed leaderboard sizes.

The page makes the evaluation names, chart definitions, and some model-level data public, but does not disclose Kimi’s complete inputs, prompts, execution logs, failure handling, random seeds, or a complete reproduction package. This article records only the page snapshot and does not claim to have rerun or reproduced Artificial Analysis’s scores.

Key results

Raw data

The page dates the model's release to October 2026, with October 23, 2026 in the FAQ. This describes the model's release date, not the publication date of the evaluation page.

Page data for the 10 evaluations

The table below shows the model-level results visible in the page’s structured data. The scores retain the approximate precision of the page fields and cannot be used to derive the weights of the composite index.

The page also shows capability indices: Finance & Accounting 56.5, Strategy & Ops 59.4, Legal 55.8, Engineering 55.9, and Economics 60.0. These are snapshots from separate capability charts, not substitute scores for the 10-item index.

Speed and latency breakdown

The long-prompt profile in the page’s structured data is: TTFT 2.97 s, reasoning 24.78 s, answer output 6.19 s, for a total of 33.94 s; the first answer token arrives at 27.74 s. The end-to-end time for the 100k-input profile is 42.11 s, showing that prompt length changes waiting time.

Boundary between page labels and official status

• Artificial Analysis labels the model Proprietary and shows No publicly available API in several charts.

• These are classification and availability snapshots from the Artificial Analysis test page. They are not statements by the official provider about its plans for releasing weights or its API status.

• This article does not use these labels to rewrite the weight-release date in Kimi’s official plans, and does not treat the page label as a conclusion that the current API is unavailable.

Conclusions and limitations

1. In the page snapshot from 2026-10-24, Kimi K3.1 Pro has a composite score of 55 and ranks #4/215, placing it near the top of the model set currently shown on that page.

2. Cost cannot be assessed from the $2/$10 list prices alone: the weighted task cost for running the index is $2.11, and the run produced approximately 148M output tokens in total. Reasoning output volume amplifies both waiting time and cost.

3. The speed metrics come from Kimi first-party API data. The 80.7 token/s and 2.97 s TTFT figures represent only the page’s testing conditions; different providers, regions, concurrency levels, and prompt lengths can change the results.

4. Proprietary and No publicly available API can only be recorded as page collection labels; they cannot replace the official weight-release plan or API documentation.

5. Dynamic leaderboards, model sets, prices, providers, and evaluation versions may change. When citing these results again, record the date, version, denominator, and Prompt Options again.

Reproduction notes

1. Fix Artificial Analysis Intelligence Index v4.3.2, the Kimi K3.1 Pro reasoning version, the collection date, and the provider.

2. Record the 10 evaluation names, the denominator of the chart’s model set, the Prompt Options (Single/Multiple, 1k/10k/100k), and the price snapshot given on the page.

3. If using your own Kimi API for regression testing, compare only task success rate, input/output/reasoning tokens, TTFT, end-to-end time, tool calls, and the actual bill. Do not claim to have reproduced the page’s score of 55.

4. Record public list prices, cache discount, blended price, per-task index cost, and full index cost separately. Their formulas differ and they cannot be substituted for one another.

5. If the page still shows No publicly available API, retain the test-page snapshot and separately check the official API status. Do not infer the official current status from the page label.