Latest in AI
Kimi K3.1 Pro: Pricing, API, Benchmarks and Highspeed Explained
Moonshot's new flagship costs less than K3 through the API, while its faster version makes the latency tradeoff more explicit.
Moonshot AI has launched Kimi K3.1 Pro and K3.1 Pro Highspeed across Kimi, Kimi Work, Kimi Code and its API. The new flagship keeps a million-token context window while targeting stronger coding and agent work. Standard API input and output rates are a third lower than K3's.
The official announcement describes a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token. Although Moonshot calls it open-weight, the weights are promised by November 4, 2026, not available to download at launch.
Kimi K3.1 Pro API pricing
The global API price sheet lists these US-dollar rates per million tokens, excluding applicable taxes:
| Model | Input | Output | Cached input | 5-minute cache write | 1-hour cache write |
|---|---|---|---|---|---|
| K3.1 Pro | $2.00 | $10.00 | $0.20 | $2.00 | $4.00 |
| K3.1 Pro Highspeed | $4.00 | $20.00 | $0.20 | $4.00 | $8.00 |
| K3 | $3.00 | $15.00 | $0.30 | $3.00 | $6.00 |
Highspeed doubles the standard model's input and output rates, but not its cache-hit rate. Cache writes are charged separately, so the cheap cached-input column does not describe the cost of initially loading a large project.
That matters for an agent returning to the same repository repeatedly. Reusing an eligible prefix can change the economics considerably; rebuilding that prefix or letting it expire changes them again. The platform defaults to a five-minute cache lifetime.
What changed from Kimi K3
According to Moonshot's technical report, K3.1 Pro is a post-trained K3 with substantially more post-training compute. The work focuses on reinforcement learning across many task environments, rather than introducing a larger context window.
Moonshot reports 65.7% on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. Its internal long-horizon coding benchmark rises from K3's 55.95% to 61.2%.
Those results need their evaluation conditions attached. Moonshot used the Claude Code harness for its Terminal-Bench run and drew competitors' scores from published reports or the leaderboard. The internal coding test is company-owned. Neither establishes which model will finish your repository's tasks most reliably.
The launch also includes 3D Make on Kimi's website and app, with online model editing and exports suitable for 3D printing. It gives the model's spatial and coding capabilities a more concrete interface than another chat demo.
API access and Kimi Code are different products
The API quickstart uses kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Thinking stays enabled, with low, high and max effort levels; max is the default. Vision input supports images and video, but public image URLs are not supported. The documentation also warns that its web-search integration is being updated and is not recommended for production workflows in the near term.
Kimi Code uses k3.1-pro and k3.1-pro-highspeed. Its model guide puts standard access on Plus or legacy Moderato plans and above. Pro or legacy Allegretto unlocks the full million-token context and Highspeed.
Start a fresh coding session when switching to preserve quality and cache reuse. Check the model ID carefully: a typo can silently route requests to K2.8 Preview without thinking.
When Highspeed is worth paying for
Moonshot claims two to three times faster output with unchanged capabilities, consuming about twice the Kimi Code quota. Faster output does not make every agent task two to three times faster: tests, tool calls and external services still take their own time.
For an interactive coding session, shortening repeated waits can justify the premium. For an overnight task dominated by test execution, the standard model deserves consideration first. The useful comparison is cost per accepted result and total completion time, not tokens per second alone.
K3.1 Pro makes that comparison unusually clear: a cheaper default model, a faster premium route, and self-hosting still waiting on the promised weights. Teams can measure the first two now without treating the third as something already delivered.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn launches and product updates into clear stories developers and founders want to read and share.