Kimi K3.1 Pro Lands: 3T-Class Open Weights That Beat GPT-6 Astra on Terminal-Bench

Moonshot AI released Kimi K3.1 Pro on October 23, and the number that will get quoted is Terminal-Bench 4.0: 65.7%, ahead of GPT-6 Astra’s 58.2% and behind only Claude Opus 5.5’s 66.4%. K3.1 Pro is a post-trained evolution of Kimi K3 with substantially larger post-training compute and joint reinforcement learning across hundreds of domains, described as Moonshot’s most capable 3T-class open-weight model to date, with a 1M-token context window and a Highspeed variant delivering 2 to 3 times the throughput. It is live now on Kimi’s apps, API (with zero data retention), and desktop tools; full weights are promised by November 4.

What the benchmark card claims, and what it concedes

The card is unusually specific about both wins and losses. OfficeQA Pro v2 hits 64.4, claimed as state of the art, and Kimi’s own hallucination benchmark shows claim-level hallucination falling from 23.4% in K3 to 8.3%, which matters more for agentic use than any single accuracy point. BenchCAD with tools lands at 92.7% against Opus 5.5’s 96.2% and Astra’s 95.9%, an honest gap. Internal finance evaluations show +7.2 points on financial modeling and roughly 4x K3 on business operations. Kimi Code Bench 3.0 long-horizon rose from 55.95% to 61.2%. Against this month’s open-weights field, the trajectory is the story: DeepSeek’s 1017 refresh moved agent benchmarks 10-35% in eight days, and K3.1 Pro now holds the open-weights lead on the benchmark every agent team is watching.

The cybersecurity deletion is the most interesting line

One section of the card deserves more attention than any benchmark. Moonshot removed cyber-related data from mid-training and RL environments before the open-weight release, and quantified the cost: xdev-bench-hard drops from 90.5% for a checkpoint without the cleaning to 81.0% for the shipping model. That is a deliberate, measured capability sacrifice, justified by stated principles: avoid extending the offensive cyber frontier, and progressively raise open-weight capability over time. After a month in which a single operator ran an agent stack that breached 27+ companies for $25 each, a major open-weights lab publishing the price tag of its own safety tax is a precedent worth more than the 9.5 points it cost. Whether other labs follow, and whether users can literally measure the gap, is now an accountability question.

What to watch

Three things decide whether K3.1 Pro consolidates Moonshot’s position. First, the weights on November 4: the license terms will determine whether this competes for self-hosting or stays an API story, and given that the 3T-class size puts it beyond most inference budgets, expect quantized community builds to be the real distribution channel. Second, independent verification of the Terminal-Bench number, since Moonshot’s own harness results have historically been directionally right but optimistic at the margins. Third, whether the Highspeed variant’s 2-3x throughput claim holds at production latencies, because the month’s pricing war has moved competition from capability to cost-per-task, and Moonshot has not yet published API pricing for either variant. The Chinese open-weights labs are now trading the benchmark lead weekly; K3.1 Pro holds it as of today.