Moonshot cuts Kimi K3.1 Pro prices by a third, now $2 and $10 per 1M tokens
Moonshot AI dropped Kimi K3.1 Pro on October 23 and released its weights on Hugging Face in early November. The Chinese model costs $2 per million input tokens and $10 per million output tokens — half of Claude Opus 5.5 and a fifth of GPT-6 Astra. It lands in the global top four on external metrics, though the license isn't free for big commercial users.
- Original source: note.com.
- The source published it on Nov 15, 2026 at 4:00 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Moonshot AI pushed out Kimi K3.1 Pro on October 23, 2026, slotting the model into its web service, app and developer API. The Beijing company followed up on November 4 by dumping the model's weights — the trained numbers that make it work — onto Hugging Face in roughly 1.56TB of files, with some reports spotting them on November 5 Japan time. Moonshot's own test results put it level with or just behind the big US models, and external evaluators' combined metrics ranked it in the global top four on release day.
The price is the headline. Kimi K3.1 Pro runs $2 per million input tokens and $10 per million output tokens, a cut of about one-third from K3's $3 and $15 — or 13 yuan in and 68 yuan out. That puts it at half the cost of Anthropic's Claude Opus 5.5, released September 22 at $4 and $20, and one-fifth of OpenAI's GPT-6 Astra, out in September at $10 and $50. A Highspeed version that processes two to three times faster costs double, $4 and $20, landing right on Claude Opus 5.5's level. Cache fees match too: $0.20 per million tokens on both K3.1 Pro and Claude Opus 5.5.
K3.1 Pro is K3 with a heavy round of extra post-training layered on, and its release partly rolls back the price spike that arrived with K3 on July 16, when Moonshot announced a roughly 2.8-trillion-parameter model it called the largest with publicly released weights. K3's price had climbed from K2.6's $4 output to $15 — about 3.75 times higher — which German outlet the decoder read as the end of ultra-cheap Chinese AI. Moonshot's own numbers claim training gains made it about 2.5 times more efficient at turning compute into performance than K2.
The weight release comes with strings. The license is proprietary, and per reports, businesses serving the model to third parties with more than $20 million in revenue over 12 consecutive months must sign a separate commercial deal with Moonshot. The architecture itself explains the cheapness: K3 and K3.1 Pro use a mixture-of-experts setup where only about 104 billion of the roughly 2.8 trillion parameters fire on any single pass — under 4% of the total.
Moonshot has been hiking prices steadily since K2 landed in July 2025 at $0.60 input and $2.50 output, the release that helped cement the cheap-and-fast Chinese model reputation. K2.5 followed in January 2026 at $3 output and K2.6 in April at $4, both around 1 trillion parameters. The ground had been laid earlier still: DeepSeek's R1 reasoning model in January 2025, whose low development cost sent Nvidia down roughly 17% in a single US session on January 27, wiping about $589 billion in market value in what was reported as the largest one-day drop on record.
TopicsMoonshot AI · Kimi K3.1 Pro · Anthropic · Claude Opus 5.5 · OpenAI · GPT-6 Astra · DeepSeek · Nvidia
Source: note.com
Written by AI from verified reporting. Not financial advice.
More stories
- Tech & AI blockchain.news 0 viewsGrid Robotics paper pitches 'agentic robotics' powered by GPT-6 Astra
- Tech & AI x.com 0 viewsNvidia director warns board of losing "irreplaceable talent" if Huang isn't reinstated as CEO
- Tech & AI blockchain.news 0 viewsNVIDIA's Vera Rubin aces MLPerf debut with 10x agentic efficiency gain
- Tech & AI blockchain.news 0 viewsOpenAI Adds Computer-Use Agents and 20+ Plugins to ChatGPT for Healthcare

