Kimi K3.1 Pro climbs to 4th place on the AA Index with 55 points and a leap in agentic tasks

Moonshot has released Kimi K3.1 Pro, and according to Artificial Analysis the new reasoning model scored 55 points on the Intelligence Index. That puts it in fourth place, 11 points above Kimi K3, which scored 44. Only Claude Opus 5.5 (58 points in max with fallback mode), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) rank higher. The K3.1 Pro weights are not yet published, but Moonshot is expected to release them soon after launch. At that point the model will become the strongest open-weights model, and the gap between the best open model and the best closed one will shrink from 12 points to 3.

Agentic leap: Terminal-Bench from 13% to 67%

The most notable changes concern agentic work. On GDPval-AA, where models solve tasks close to real office and expert work, K3.1 Pro scores 1 846 Elo. That is 312 points more than Kimi K3 and third place behind Claude Opus 5.5 (1 866) and Muse Wonder Full (1 857). The gap to the leader is just 20 Elo.

On Terminal-Bench 4.0 the model scores 67% and beats every other model, including Claude Sonnet 5.5 at 64%. By comparison, Kimi K3 solved only 13% here. A jump like that in a single update points to targeted work on how the model behaves in the command line: planning steps, fixing errors and seeing tasks through to completion.

AutomationBench-AA, which tests agentic scenarios in SaaS products, rose from 58% to 73%.

Hallucinations and science: more modest progress

Gains on the other tests are moderate. Humanity's Last Exam rose from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. The agentic leap looks like the main bet of this release, not an even increase in intelligence across the board.

Answer reliability changed more noticeably. The AA-Omniscience Index climbed from 20 to 30: the hallucination rate fell from 53% to 44%, and accuracy rose from 48% to 51%. The model more often admits it does not know the answer instead of making one up. For agents that call tools on their own and write to production systems, this is an important improvement.

Cost per task and the weights question

K3.1 Pro spends 55K output tokens per Intelligence Index task. Kimi K3 used 48K, which is 14% fewer. Moonshot, however, cut prices by 33% across all three tiers: $2 per million input tokens, $10 per million output tokens and $0.20 per million cache hits. As a result, the cost of running a single task stayed almost the same, at $1.98. The model got markedly stronger, and it costs about the same to use.

Key specs:

• Intelligence Index: 55 points, 4th place (Kimi K3: 44)
• GDPval-AA: 1 846 Elo, 3rd place, +312 over K3
• Terminal-Bench 4.0: 67% (K3: 13%), best result among all models
• AutomationBench-AA: 73% (up from 58%)
• AA-Omniscience Index: 30 (up from 20), hallucinations 44%
• Size: 2.8T parameters total, 104B active
• Context: 1M tokens, input: text and images
• Price: $2 / $10 / $0.20 per 1M tokens (input / output / cache)
• Task cost: $1.98

What to check before switching to K3.1 Pro

Teams building coding and automation agents get a cheap alternative to closed models. On Terminal-Bench 4.0, K3.1 Pro beats Claude Sonnet 5.5, and on GDPval-AA it trails the leader by 20 Elo. Tokens also cost considerably less. But the model spends 14% more output tokens than its predecessor, so you should calculate the cost per task, not just the price per million tokens.

The second question concerns the weights themselves. Until they are out, the model is available only through the Moonshot API. Self-hosting a 2.8T model with 104B active parameters will require a serious cluster, so cloud providers and large companies with their own infrastructure will benefit most from open weights. For everyone else, something else matters more: pressure on closed labs' prices is growing, because the quality gap has narrowed to 3 points.

Before migrating, run the model on your own agentic workloads. Benchmarks show a model's class, but they do not replace testing on real tasks. Pay particular attention to the hallucination rate: 44% is still high.