THE SHORT ANSWER
● Higher score — Claude Opus 5.5 at 58 on the Artificial Analysis Intelligence Index, against K3.1 Pro at 55
● Lower cost — Kimi K3.1 Pro at $2 input and $10 output per million, with $0.20 cache hits
● Open weights — Kimi, promised November. Opus 5.5 stays closed
● Science reasoning — Opus 5.5 by a wide margin - 63.3 against 38.1 on Terminal-Bench Science 0.1
Moonshot released K3.1 Pro on 23 October and it entered the Artificial Analysis Intelligence Index at number 4 with a score of 55. Claude Opus 5.5 holds 58. That is the whole headline gap, and it is worth sitting with before looking at anything else: a Chinese open-weight-bound model is now three points off the frontier on an independent index.
The numbers side by side
| Measure | Kimi K3.1 Pro | Claude Opus 5.5 |
|---|
| AA Intelligence Index | 55 (4th) | 58 |
| GDPval-AA Elo | 1,846 | 1,866 |
| Terminal-Bench 4.0 (AA measured) | 67% | not stated in the same run |
| Terminal-Bench 4.0 (Moonshot self-reported) | 65.7% | 66.4% |
| SWE-Marathon (Moonshot sheet) | 51.7 | 46.3 |
| Terminal-Bench Science 0.1 | 38.1 | 63.3 |
| BenchCAD with tools | 92.7 | 96.2 |
| Input per 1M tokens | $2.00 | higher |
| Output per 1M tokens | $10.00 | higher |
| Context window | 1M tokens | large, not matched on this figure |
| Weights | November | Closed |
The two Terminal-Bench numbers are both real
This trips up every roundup covering the launch. Artificial Analysis independently measured K3.1 Pro at 67% on Terminal-Bench 4.0, ahead of Claude Sonnet 5.5 at 64%. Moonshot separately self-reported 65.7%, comparing against Opus 5.5 at 66.4%. Those are different runs with different comparison sets, not a contradiction. A third figure doing the rounds - 65.7% against GPT-6 Astra at 58.2% - matches neither pairing.
The practical reading: on terminal and agentic coding work the two models are within a point or two of each other, and the choice does not turn on that benchmark.
Where Opus 5.5 is genuinely ahead
Science. Terminal-Bench Science 0.1 puts K3.1 Pro at 38.1 against Opus 5.5 at 63.3, and GPT-6 Astra at 68.1. That is not a three-point gap, it is a 25-point gap, and it is the clearest capability difference in the sheet. If your workload is scientific reasoning rather than code or documents, the index score is misleading you.
Opus 5.5 also holds BenchCAD with tools at 96.2 against 92.7, and GDPval-AA at 1,866 against 1,846.
Where Kimi wins outright
Price, and it is not close. $2 input and $10 output with $0.20 cache hits, each a third below Kimi K3 before it. A 1 million token context window. And weights in November, at which point the self-hosting option opens and Artificial Analysis expects this to be the highest-scoring open-weights model available.
Moonshot also reports claim-level hallucinations dropping from 23.4% on K3 to 8.3% on its internal hard hallucination benchmark - a vendor-measured figure, but a large enough move to matter if K3 burned you on that.
The caveat nobody is pricing in
For the open-weight release Moonshot deliberately stripped cyber-related data from mid-training and reinforcement learning. On its own xdev-bench-hard the shipped model solves 81.0% at pass@3, down from 90.5% for an unfiltered checkpoint. That is a vendor publishing a weaker model on purpose, which is new, and it means the weights you download in November are not the model that scored 55 on every axis. Nobody has yet measured what else moved.