Kimi K3.1 Pro: agentic benchmarks and the routing call

Published on 27 October 2026.

A model whose weights have not been released entered fourth place among 221 models, with 55 points, and finished ahead of Claude Sonnet 5.5 on the evaluation that measures terminal work. Kimi K3.1 Pro is not news for the placing.

Whoever signs the inference invoice recognizes the question that remains: where does the closed model still pay for itself?

What did Moonshot release on 23 October?

Kimi K3.1 Pro is the model Moonshot AI released on 23 October 2026 and that Artificial Analysis placed 4th of 221 on the Intelligence Index v4.3.2, with 55 points. On agentic tasks: 67% on Terminal-Bench 4.0, 73% on AutomationBench-AA, 1,846 Elo on GDPval-AA and US$ 2.11 per task on the model page, read on 27 October 2026.

Image: Moonshot AI

Image: Moonshot AI, reproduced from the Kimi K3.1 Pro launch announcement (23 October 2026).

The Artificial Analysis article of 23 October 2026, read on 27 October 2026, is the primary source for this reading. The index is the easy part. Ahead of it sit three models: Claude Opus 5.5, with 58 points, Claude Sonnet 5.5 and Muse Wonder Full, with 56.

The piece reads the three instruments separately and derives the routing frontier by task family.

What does the spec sheet explain and what does it not explain?

The spec sheet repeats the K3 family: 2.8 trillion total parameters, 104 billion active per request, a 1 million token window, text and image input. None of that is new. K3.1 Pro is a post-trained successor, and the measured gain comes from there, not from an architecture the source does not describe.

The piece on the K3 open weights already described the KDA attention and the licence of the previous generation. Here the architecture enters as inherited: the jump from 13% to 67% on Terminal-Bench 4.0 is not explained by parameter, but by the tuning after pre-training.

The model page, read on 27 October 2026, records availability through a first-party API. The list price is US$ 2 per million input, US$ 10 output and US$ 0.20 cache read, 33% below Kimi K3 on all three rates.

What is missing is the weight: the source classifies K3.1 Pro as proprietary on the date of reading, and no licence is published.

What does Terminal-Bench 4.0 measure that the others do not?

It measures terminal work with a long horizon: a sequence of commands, state that persists between steps and error correction along the way. Kimi K3.1 Pro scored 67% and the predecessor, 13%, in the article of 23 October 2026. This is where the model with unpublished weights already contests the route.

Thirteen per cent is the score of a model that describes the right command. Sixty-seven per cent is the score of a model that gets it wrong, reads the error and tries again until the environment is correct.

The test demands continuity. The task chains commands inside the same environment: the state of one step is the input of the next, and an early error contaminates what comes after.

On the same source and the same date, Claude Sonnet 5.5 scored 64%. The article calls the 67% the highest score of any model on that evaluation. The article of 23 October does not carry the Opus 5.5 score on that test.

In code and terminal with a long horizon, the closed model stopped being the automatic choice: Kimi K3.1 Pro, with weights still unpublished, is ahead of Sonnet 5.5 by 67% to 64% and charges US$ 2.11 per index task against US$ 5.46. Both cost numbers come from the pages read on 27 October 2026.

What does 73% on AutomationBench-AA say about SaaS workflow?

It says the model chains actions inside business software: fills a field, respects a permission, handles a form exception. Kimi K3.1 Pro rose from 58% on K3 to 73%, according to the article of 23 October 2026. Running a command is one thing. Operating a system that was not designed for an agent is another.

An error in a terminal leaves a red test. An error in a SaaS flow writes the wrong record, sends the email to the customer or moves inventory, and the data stays.

Seventy-three per cent is not 100%, and the distance between the two numbers defines whether the flow can run without human review.

Here the route is still mixed: the article of 23 October does not carry the score of a closed model on this evaluation. Kimi K3.1 Pro enters with human acceptance and a record of each action; a closed model stays in the flow where the exception costs customer data.

Why is GDPval-AA read in Elo and not in percentage?

Because GDPval-AA compares deliverables between models, and Elo is a relative scale: it is not a test hit rate and does not convert into a percentage. Kimi K3.1 Pro scored 1,846 Elo and finished 3rd, behind Claude Opus 5.5, with 1,866, and Muse Wonder Full, with 1,857, in the article of 23 October 2026.

The gain over the predecessor: 312 Elo. What it measures is the distance between deliverables judged in pairs, not the share of tasks solved.

Twenty Elo points separate K3.1 Pro from Opus 5.5, and that is little on the scale. Kimi K3.1 Pro ties in the middle of office work, and whoever delivers a report, a proposal or a legal document pays the closed model for the guarantee.

What does the 55 index of Kimi K3.1 Pro show and what does it hide?

It shows the weighted average of ten evaluations on the Intelligence Index v4.3.2 and the 4th position among 221 models. It hides the distribution. The gain over the 44 of K3 in the max variant comes concentrated in three agentic evaluations, and the others rise little.

The others move in a short step, according to the article of 23 October 2026: CritPt from 23% to 25%, SciCode from 59% to 61%. On AA-Omniscience, the index goes from 20 to 30, hallucination falls from 53% to 44% and accuracy rises from 48% to 51%.

The version of the index that produces that 55 has Terminal-Bench 4.0 inside and AutomationBench-AA with its own weight, as the piece on index v4.3 describes.

Artificial Analysis projects that, once the weights are published, K3.1 Pro will be the highest scoring open model on the index, with the distance between the open leader and the closed leader falling from 12 to 3 points. It is a projection, not a measurement.

Why do two models with the same price cost three times more per task?

Because the list price is the same and consumption is not. Kimi K3.1 Pro and Claude Sonnet 5.5 charge US$ 2 per million input and US$ 10 output, and the index cost per task is US$ 2.11 against US$ 5.46 on the pages read on 27 October 2026. What separates the two is the spend on output tokens.

Source: Artificial Analysis, article of 23 October 2026 and model pages read on 27 October 2026.

The 150 million output tokens in the last column are for the full index. Per task, they are 55 thousand. Mixing the two invents a difference of three orders of magnitude.

The pattern is verbosity, and it does not become a formula. Four hundred and twenty million against 150 million is 2.8 times, and the cost per task is 2.59 times higher. Opus 5.5 charges double on the table, consumes 260 million and lands at US$ 5.98, not at the US$ 7.31 that the rule of three from the US$ 2.11 of K3.1 Pro would suggest.

The launch article published US$ 1.98 per task; the model page, read on 27 October 2026, carries US$ 2.11 as a weighted average, and the table uses the page number because the three models were read on the same date.

The piece of 26 October already dealt with token price against cost on the same model, and that account is not repeated here.

What do 41.4 tokens per second change in a long-running agent?

They change the queue. Kimi K3.1 Pro delivers 41.4 output tokens per second on the model page read on 27 October 2026, against 98.3 for Claude Opus 5.5 and 141.2 for Claude Sonnet 5.5, on the same source and the same date. On an isolated question that is comfort.

On an agent of dozens of steps, it is accumulated wait: K3.1 Pro appears 188th of 192 models in output speed on the same page.

Each step waits for the previous one. An agent that chains thirty calls pays the output rate thirty times, and the distance between 41.4 and 141.2 becomes wall time.

The evaluator's speed is measured in its own environment, not the rate each provider advertises.

At high volume with a short response, it enters the criterion next to price.

What does the weight calendar change in hosting?

It changes who can run the model, where the data sits and how cache and fallback are designed. The Kimi K3.1 Pro weights have not been published, and Artificial Analysis projects their release shortly after launch, with no date. Until then, the agentic gain is a first-party API number.

The projection is the evaluator's and carries no date: the weights come out, the model becomes the highest scoring open one and the distance to the closed leader falls from 12 to 3 points.

Today hosting is a choice between providers, and the cache belongs to the provider.

When the weight comes out, the model can run in the buyer's environment, the cache becomes yours, and fallback stops depending on a second provider. It is not a price decision. It is a decision about where the data sits.

There is no date fixed by the source and no published licence for the K3.1 Pro weights, so no sentence about commercial use holds. The dispute between the open and closed leader is the axis of the text on the end of the duopoly.

How to route an agent with this set of numbers?

Route by the evaluation that most resembles the task, not by the position on the index. Code and terminal with a long horizon go to Kimi K3.1 Pro, with a closed model on the second attempt. High-risk office deliverables and tasks that require the top of the index stay on the closed model.

The yardstick comes before the supplier. The piece on the three agentic evaluations already showed that they measure different things, and here the difference appears inside the same model, which rises from 13% to 67% on one evaluation and from 58% to 73% on the other.

1. Choose the yardstick by the task, before looking at any ranking.
2. Measure the cost per task on your own load, with model, tokens and result in the same record.
3. Define fallback by family and test it before the first incident.

The rule in one sentence: route by the evaluation that resembles your task, never by the position on the index.

Where does Nexforce Router come into this decision?

It comes in as the place where the rule by task family is applied and audited. Nexforce Router is a gateway with one API for the main providers, routing by cost, performance, latency and context, automatic failover with configurable fallback, budget per key, agent or project and a record of each call.

A routing table is only worth it if there is a single point that applies the rule and shows what passed through it. With the Router, changing the model of a family becomes configuration, not code reintegration.

The limit is clear. The Router executes and records the defined route, and does not decide what counts as a completed task.

Frequently asked questions

What is Kimi K3.1 Pro?

It is the model Moonshot AI released on 23 October 2026, a post-trained successor to Kimi K3, with 2.8 trillion total parameters and 104 billion active per request. Artificial Analysis placed it 4th of 221 models on the Intelligence Index v4.3.2, with 55 points. The gain over the predecessor is on Terminal-Bench 4.0: 67% against 13% for Kimi K3, according to the article of 23 October 2026.

Is it already an open weights model?

No. The K3.1 Pro weights were not published on the date of the reporting, and the model page classifies it as proprietary in the reading of 27 October 2026. Artificial Analysis projects that, once the weights are published, it will be the highest scoring open model on the index. In the same projection, the distance between the open leader and the closed leader falls from 12 to 3 points. The source fixes no date.

Why does it cost less per task with the same token price?

Because it consumes less output per index task. Kimi K3.1 Pro and Claude Sonnet 5.5 charge US$ 2 per million input and US$ 10 output, and the cost per task is US$ 2.11 against US$ 5.46, on the pages read on 27 October 2026. K3.1 Pro spends 150 million output tokens on the full index, against 420 million for Sonnet 5.5.

Which benchmark should decide an agent's route?

The one that measures the same family of work as your task. Terminal-Bench 4.0 serves code and terminal with a long horizon, AutomationBench-AA SaaS software flow and GDPval-AA office deliverables. In high-volume support with a short response, the yardstick is list price and output speed: K3.1 Pro delivers 41.4 tokens per second on the model page read on 27 October 2026. None of the three replaces the cost per task measured on your own load.

References and Further Reading

• Artificial Analysis, "Kimi K3.1 Pro launches at 4th on the Intelligence Index and will lead open weights models", 23 October 2026: https://artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks (accessed on 27/10/2026).
• Artificial Analysis, model pages for Kimi K3.1 Pro, Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full: https://artificialanalysis.ai/models/kimi-k3-1-pro, https://artificialanalysis.ai/models/claude-opus-5-5, https://artificialanalysis.ai/models/claude-sonnet-5-5, https://artificialanalysis.ai/models/muse-wonder-full (accessed on 27/10/2026).

What is left for the closed model

The list of what requires a closed model got shorter. What remains is the top of the index and the deliverable where an error costs a contract, two cases in which no published benchmark answers for the risk.

When the route changes, Nexforce Router applies and records the choice, with cost and fallback in the same record.