Kimi K3.1 Pro closes in on Claude Opus 5.5, but doesn’t quite catch it

Kimi K3.1 Pro is a post-trained version of Kimi K3, the 2.8-trillion-parameter open-weight model Moonshot released back in July. On Moonshot’s own benchmarks, the new version comes close to Claude Opus 5.5 without quite catching it: 65.7% on Terminal-Bench 4.0 against Opus 5.5’s 66.4%, and 92.7% on BenchCAD with tools against Opus 5.5’s 96.2%. Both scores also sit ahead of GPT-6 Astra, which Moonshot reports at 58.2% and 95.9% respectively on the same two tests. That’s a real improvement over K3, not a win over Anthropic’s top model, close parity rather than the lead some of the social media framing around this release has suggested.

The more interesting number might be hallucinations rather than raw capability. Moonshot says its claim-level hallucination rate on an internal adversarial stress test dropped from 23.4% on K3 to 8.3% on K3.1 Pro, though the company is upfront that the test is deliberately adversarial and doesn’t reflect typical day-to-day error rates. The model keeps K3’s 1-million-token context window and adds image and video input, runs in thinking-only mode (requests with thinking disabled get rejected outright), and ships alongside a faster K3.1 Pro Highspeed variant that Moonshot describes as 2 to 3 times quicker at the same capability level, at roughly double the quota cost inside Kimi Code.

Access right now runs through Kimi.ai, the mobile apps, Kimi Work, Kimi Code, and the API under the model names kimi-k3.1-pro and kimi-k3.1-pro-highspeed, gated behind Plus-tier subscriptions or higher. Full open weights are scheduled for November 4, and Moonshot says it stripped cyber-related material from the training data before that release, a version that kept that data scored 90.5% on the company’s internal xdev-bench-hard test against 81% for the model it’s actually shipping.