Only Anthropic rows outrank Kimi K3.1 Pro’s 55.1 at $2.11 a task

Three rows on the Intelligence Index score higher than Kimi K3.1 Pro, and all three are Anthropic’s. Moonshot AI’s model, released October 23, 2026 by Artificial Analysis’s dating and priced at $2.00 and $10.00 per million input and output tokens on both its model page and OpenRouter’s listing, scores 55.1 at $2.11 per task in the snapshot retrieved Oct 27, 2026, 2:30 PM UTC: fourth of 113 configurations with a measured cost per task, and on the chart’s cost frontier. That makes it the highest-scoring configuration from any creator other than Anthropic, and the cheapest row above it costs 1.6x as much. No row on the coding-agent chart runs Kimi K3.1 Pro yet.

One row, scored under one harness

The Artificial Analysis Intelligence Index at version 4.3.2 is one composite over 10 evaluations (AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1), weighted 30% agents, 20% coding, 20% scientific reasoning, and 30% general capability, with every model run through its API under the same harness; the cost beside each score is the weighted average API bill for one task, which the chart keeps in five parts, non-cached input, cache reads, and cache writes on the input side and reasoning and answer tokens on the output side.

The snapshot holds a single Kimi K3.1 Pro configuration with no effort label, under the settings Artificial Analysis names in the row, so every figure below is that single row.

Fourth of 113, behind three Anthropic rows

Among the 113 comparable configurations in the snapshot retrieved Oct 27, 2026, 2:30 PM UTC, Kimi K3.1 Pro ranks fourth on the index at 55.1. Three configurations score higher, and every one of them is Anthropic’s: Claude Opus 5.5 (Max, Default Fallback) at 57.6 for $5.98 per task, 2.5 points higher at 2.8x the cost, Claude Sonnet 5.5 (Max, Default Fallback) at 56.0 for $5.46 per task, 0.9 points higher at 2.6x the cost, and Claude Opus 5.5 (Xhigh, Default Fallback) at 56.0 for $3.46 per task, 0.9 points higher at 1.6x the cost. The cheapest configuration that scores higher is Claude Opus 5.5 (Xhigh, Default Fallback), at 1.6x the Kimi K3.1 Pro cost for 0.9 points more.

The row is on the chart’s cost frontier: no comparable configuration scores at least as high at the same or lower cost per task. Walking the frontier from the top, the vertex above it is Claude Opus 5.5 (Xhigh, Default Fallback) at 56.0 for $3.46 per task: the step up buys 0.9 points for 1.6x the cost. The vertex below it is Claude Opus 5.5 (High, Default Fallback) at 53.6 for $1.82 per task: the step down gives up 1.5 points and pays 0.9x the Kimi K3.1 Pro cost.

On cost per task the row sits in the upper part of the chart: 94 of the 113 cost less, and 18 cost more. Within one index point of the 55.1 sit two configurations: Claude Opus 5.5 (Xhigh, Default Fallback) at 56.0 for $3.46 per task and Claude Sonnet 5.5 (Max, Default Fallback) at 56.0 for $5.46 per task.

The highest row from outside Anthropic

Because every row above it is Anthropic’s, Kimi K3.1 Pro is the highest-scoring configuration from any creator other than Anthropic. The table takes the six other creators with the highest-scoring rows, one row each, and measures that row against Kimi K3.1 Pro. The next creator’s best row is GPT-6 Astra (Max) at 52.7 for $3.26 per task, 2.4 points lower at 1.5x the cost. Of the five rows in the table that score lower, two are cheaper: Gemini 4 Argon (High) at 0.8x the cost for 2.5 points fewer and MiMo-V2.6-Pro at 0.06x the cost for 8.8 points fewer.

Inside the $2.11 per task

Input-side cost makes up $1.56 of the $2.11 per task, 74% of the bill, and output the remaining $0.552. Ranked by size, the five parts are cache reads at $1.06, cache writes at $0.462, reasoning tokens at $0.349, answer tokens at $0.203, and non-cached input at $0.044. Cache reads alone are 50% of the total; that cache-read bill is the 13th-largest of the 113 comparable configurations, while the input side as a whole is the 18th-largest and the output side the 17th-largest. The 55,195 output tokens the model writes per task are 63% reasoning tokens.

The table splits the bill of each row ranked above Kimi K3.1 Pro into its two sides, cheapest first. Every one of them spends more than Kimi K3.1 Pro on input as well as on output.

The prices behind these figures are $2.00 per million input tokens and $10.00 per million output tokens with a 90% cache discount, as Artificial Analysis’s model page, captured October 27, 2026 UTC, records them. The same page rounds the row to 55 at $2.11 per task, lists the model as proprietary and released October 2026, and writes that it “supports text and image input, outputs text, and has a 1M tokens context window.” On volume it counts 150M output tokens to run the whole index, “very verbose in comparison to the median of 81M,” at 41.4 output tokens per second. Its verdict reads: “Kimi K3.1 Pro is amongst the leading models in intelligence and reasonably priced when comparing to other models of similar price.” It adds: “It's also notably slow and very verbose.” A footnote reads: “This page shows the reasoning version of this model. A non-reasoning variant may also exist.”

The same page gives the row four ranks: #4 / 221 on intelligence, #187 / 192 on speed, #92 / 121 on cost, and #98 / 122 on verbosity, each within the page’s own comparison group. This chart ranks only the configurations with a measured cost per task, which puts the same row fourth of 113. The two ranks agree on the position and differ on the denominator because the cohorts differ, so a reader quoting either should name its cohort.

Closed on the chart, with weights promised

Moonshot AI’s release post calls the model “the first model in our new K3.1 series and our most capable 3T-class open-weight model to date, with a 1-million-token context window,” and states that “Full model weights will be released by Nov 4, 2026.” OpenRouter’s listing, captured October 27, 2026 UTC, describes it the same way: “Kimi K3.1 Pro is Moonshot AI's flagship open-weight multimodal reasoning model and a post-trained successor to Kimi K3, with a 1M-token context window.” The listing dates the release Oct 23, 2026, gives a 1,048,576 token context window, and prices Moonshot AI’s endpoint at $2.00 per million input tokens, $10.00 per million output tokens, and $0.20 per million cached input tokens. The snapshot retrieved Oct 27, 2026, 2:30 PM UTC flags the Kimi K3.1 Pro row as closed: Artificial Analysis lists it as proprietary, and the chart carries that flag until Artificial Analysis changes it; the release post’s promise does not move it. Of the 113 comparable configurations, 49 are flagged open, 64 closed. The highest-scoring open-weights row is MiMo-V2.6-Pro at 46.3 for $0.133 per task, 8.8 points below Kimi K3.1 Pro at 0.06x the cost.

The snapshot lists Moonshot AI’s rows under the creator name Kimi and stores one configuration there besides it: Kimi K2.7 Code at 25.8 for $0.541 per task, 29.3 points lower at 0.3x the cost, flagged open, released June 12, 2026. The release post says that “Kimi K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute,” and the Intelligence Index snapshot stores no Kimi K3 row with a measured cost per task to set beside it, nor a Kimi K3.1 Pro Highspeed row: the snapshot has one Kimi K3.1 Pro configuration, and every figure in this note belongs to it.

Moonshot AI’s own benchmarks and the coding-agent chart

The release post reports that on Terminal-Bench 4.0 “K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%),” with a footnote that “Kimi K3.1 Pro is evaluated with the Claude-Code harness” and that, unless a result says otherwise, its figures use the reasoning effort max, temperature 1.0, and top-p 0.95. Those are figures Moonshot AI reports, and by the same footnote each competitor’s score is the higher of the figure in its own post and the official Terminal-Bench 4.0 leaderboard; nothing on this site confirms or contradicts them. The post also announces a second model, kimi-k3.1-pro-highspeed, which it says “delivers 2-3× the speed of K3.1 Pro for more responsive interactions,” and the OpenRouter listing prices it at $4.00 and $20.00 per million tokens, twice the kimi-k3.1-pro prices. Kimi’s API guide sets reasoning effort through a reasoning_effort request field with three values, low, high, and max, and a default of max. The Intelligence Index row carries no effort label, so the chart does not say which of the three it ran.

The coding-agent chart snapshots the public Artificial Analysis coding-agents comparison each day, and its unit is different: a model inside a named agent harness at one effort setting, scored on DeepSWE v1.1, Terminal-Bench 4, and SWE-Atlas-QnA and costed as the mean API bill for one task in that harness. In the coding-agent snapshot retrieved Oct 27, 2026, 7:14 AM UTC, no row runs Kimi K3.1 Pro in any harness, so this note has no coding-agent score for the model and does not estimate one. The Moonshot AI row it does store is Kimi Code CLI · Kimi K3 (default) at 41.4 on AA Index for $5.17 per task, a different model in a different harness from the Terminal-Bench run the release post reports, so the two do not compare.

Limits

• Artificial Analysis measured the 55.1, the $2.11, and the token counts through its own harness on the retrieval date under version 4.3.2 of the index; a different harness, prompt set, or effort setting would give different figures.
• aicharts computed the rank, the three configurations above it, the frontier steps, the per-creator table, the side split, the component ranks, and the open and closed counts from that snapshot. The next snapshot can move any of them, and the headline rests on three Anthropic rows that a rescoring or a new release can change.
• The weights flag is Artificial Analysis’s classification on the retrieval date. The release post’s promise of weights by November 4, 2026 is Moonshot AI’s, and nothing here checks whether it is kept.
• No per-evaluation score is stored, so nothing here says which of the 10 evaluations lifts or drags the composite.
• Kimi K3.1 Pro inside Kimi Code CLI, Claude Code, or any other harness is a configuration the coding-agent snapshot does not store, so this note says nothing about it.
• The Terminal-Bench 4.0 figures, the Highspeed speed claim, the context window, and every sentence about the weights release are Moonshot AI’s or OpenRouter’s. aicharts did not run Kimi K3.1 Pro.