Kimi K3.1 Pro

Pro and Highspeed serve the same model job

Moonshot calls K3.1 Pro a 3T-class model post-trained from K3; that wording is not a newly verified exact parameter count. Highspeed is described as the same model quality at roughly 2–3× speed. Treat speed and benchmark claims as Moonshot-reported results under its configurations, not independent guarantees.

API IDs, thinking and context

The paid API uses kimi-k3.1-pro and kimi-k3.1-pro-highspeed with a 1,048,576-token context. The guide documents max_completion_tokens defaulting to 131,072 and allowing up to 1,048,576 within the shared context budget. Thinking is always on, with low, high and max; the API defaults to max. Disabling thinking is rejected. Product and Code defaults can differ.

• Standard Pro API record

  Read standard ID, costs and request fields.

• Highspeed API record

  Read the separate priced variant.

Price and plan boundaries

The API platform lists Pro at $2 input and $10 output per million tokens, with cache writes of $2 for 5-minute retention or $4 for 1-hour retention, and $0.20 cache hits. Highspeed is $4 input and $20 output, cache writes of $4 for 5 minutes or $8 for 1 hour, and $0.20 cache hits. API prepaid balance and rate tiers are separate from Kimi Code subscriptions. Code uses k3.1-pro and k3.1-pro-highspeed: Pro starts at Plus/legacy Moderato, while 1M context and Highspeed require Pro/legacy Allegretto or higher. Code defaults to high thinking and says Highspeed uses about twice the quota.

A recorded 3D workflow, not a native 3D API

Moonshot’s modular-lamp recording illustrates a model working through a tool-mediated 3D task. It does not prove that the text-output model exposes a native 3D-generation API, that every asset was produced without tools, or that the task succeeds in one request. The cover is an original GetLLMs illustration; the recording remains on Moonshot’s own CDN.

What remains to verify before switching

Kimi Code quota, API balance, context entitlement and artifact access are different checks. Community discussion highlights quota exhaustion and total reasoning-token cost; these reports are useful trial questions, not representative failure rates. Run one repeated task with a budget and a checkable result before switching. For self-hosting, wait for a verifiable official checkpoint and license rather than treating a scheduled weight release as files already downloaded.

Useful first jobs

• Supervised coding and long-running tool workflows.
• Visual analysis and tool-assisted 3D work.
• Source-checked document and financial research.