China's Moonshot AI formally unveils "Kimi K3.1 Pro," priced at half of Anthropic's top model, 3 days after a leak

2.8 trillion parameters, handling 1 million tokens

Chinese AI company Moonshot AI on Oct. 23 (Oct. 24 Japan time) formally announced its large model "Kimi K3.1 Pro." It made the model available the same day in its own "Kimi" chat service, the development tool "Kimi Code," the business offering "Kimi Work" and its API. It also prepared "Kimi K3.1 Pro Highspeed," which processes 2 to 3 times faster, and said the weights (trained parameters) would be released by Nov. 4.

The model is a "mixture of experts" (MoE) design that uses only some of many expert components, with 2.8 trillion total parameters and 104 billion used to process a single token. It handles a context of 1 million tokens at once and can take in text and images. It uses the same architecture as the previous "Kimi K3," with additional training.

For this release, Moonshot removed cyber-related data from the training data to curb cyber capabilities that could be used in attacks. It is the first such step by the company, which also pledged that it will not release weights that advance the frontier of cyber capabilities. How that pledge applies to the weights to be released on Nov. 4 is not clear.

API priced at $2 for input, half of Opus 5.5

API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens, down from $3 for input and $15 for output for the previous K3. That is half the price of Anthropic's "Claude Opus 5.5" ($4 input, $20 output) and one-fifth that of OpenAI's "GPT-6 Astra" ($10 input, $50 output). The Highspeed version is double, at $4 for input and $20 for output. Compared with "V4.1-Flash-1017" from China's DeepSeek, however, the per-token price is up to about 17 times higher.

Ahead on long-running development, short of Opus 5.5 on composite index

In evaluations released by Moonshot, it scored 51.7 on SWE-Marathon, which measures long-running development work, above Claude Opus 5.5's 46.3 and GPT-6 Astra's 42.5. It also scored 70.0 on HLE, which tests answering difficult questions with tools, above both models. On Terminal-Bench, which covers command-line operations, it scored 65.7, slightly short of Opus 5.5's 66.4, and on DeepSWE, which solves development tasks, it came in below DeepSeek's 77.0.

On the composite index from third-party Artificial Analysis it scored 55, up 11 points from K3 but short of Opus 5.5's 58. Most of the evaluations were measured by Moonshot itself, and other companies' scores were collected from various sources.

The Information reported on Oct. 28 (Oct. 29 Japan time) that K3.1 Pro and DeepSeek's new version were trained with heavy reliance on the output of OpenAI's GPT-6 Astra.

Name leaked through inference software, released 3 days later

The K3.1 name had leaked three days before the release. On Oct. 20 Japan time, a change proposed to the inference software "sglang" carried the name "Kimi K3.1," covering image input and levels of thinking depth. @aitrackerbot, which tracks the appearance of models, reported that the proposal was closed after seven minutes.

On Oct. 22 Japan time, "Kibitz," an anonymous model that appeared on "Arena," where users compare models, was determined to use the text segmentation specific to Kimi. On Oct. 23, Moonshot officially posted a teaser video showing "Model: Kimi K3.1" on the Kimi Code screen. Early on Oct. 24, "Pro" and "Pro Highspeed" became selectable in the Kimi app, and the formal announcement came about an hour later.