Kimi K3.1 Pro API: Moonshot's $2/$10 Flagship With 1M Context — and the K2 Retirement Clock

The verified rate card (Nov 4, 2026, per MTok)

K3.1 Pro undercuts its own predecessor by a third on both directions. The table below is read straight from platform.kimi.ai's pricing page on Nov 4; cache billing is automatic with a 5-minute default TTL, and cache-hit input never incurs an extra write charge.

Context window 1,048,576 tokens on all K3-series models; K3.1 Pro billing is flat pay-as-you-go with no context-length tiering. Prices exclude applicable taxes. Source: platform.kimi.ai pricing page and quickstart, read directly on 2026-11-04.

What shipped, graded (Nov 4)

• The launch, verified across official docs. platform.kimi.ai's model list describes kimi-k3.1-pro as the first model of the new K3.1 series and Kimi's most capable flagship to date: a 3T-class model with native visual understanding, a 1M-token context window, and stated improvements in coding, 3D creation, and financial research and modeling. Chinese press dated the launch Oct 24, 2026, with the model live the same day on Kimi.com, Kimi Work, Kimi Code and the Kimi API. Note the careful wording: K3 (2.8T parameters) remains "the world's first open-source 3T-class model" — K3.1 Pro is "3T-class", and its weights are not yet public.
• The behavior contract changed with the generation. Thinking mode is always on; a top-level reasoning_effort field accepts low, high and max (default max). Sampling is fixed: temperature 1.0, top_p 0.95, n=1, no frequency/presence penalties — scripts that tune these silently do nothing. Vision input takes base64 or ms:// file IDs, not public URLs. A cumulative top-up of at least $1 unlocks the flagship tier, with rate limits scaled to account tier.
• The price cut, with math. Versus K3: input $3.00→$2.00 and output $15.00→$10.00, both −33%; cache-hit $0.30→$0.20. The Highspeed variant doubles the base price ($4/$20) for roughly 2-3× the speed. Batch inference on kimi-k3.-pro runs at 60% of list ($1.20 in / $6.00 out) — though batch jobs are also charged cache-write fees, with cache hits on a best-effort basis only.
• The K2 retirement clock, verified from official notes. kimi-k2 retired May 25, 2026. kimi-k2.5 and the whole moonshot-v1 series (including -vision-preview and -auto) were discontinued Aug 31, 2026 and now return 404 "model not found". From Oct 31, 2026, K2-series models are no longer available to newly registered users; existing users keep access until Moonshot retires the entire K2 series by the end of December 2026. The official migration target is kimi-k3.1-pro. If you were routing cheap agent loops to K2.5 ($0.45/$2.25, verified 2026-08-23), the same volume on K3.1 Pro costs about 4.4× more per direction — re-budget before the December cutoff.
• Open weights: promised, not delivered (timestamped). Launch coverage said K3.1 Pro weights would be released by Nov 4, 2026. Our probe on Nov 4 (UTC+8): no K3.1 Pro repository under moonshotai on Hugging Face, and ModelScope returns "record not found" — while the moonshotai/Kimi-K3 repo still resolves normally as a control. Treat self-hosting as unavailable until a repo actually lands; per this site's standing rule we do not write "never", only "not as of the timestamp".
• What we did not verify. Relay availability: ChinaModelAPI has no live relay key on this machine, so we cannot confirm a kimi-k3.1-pro alias on any independent relay — confirm in your provider's dashboard. Benchmark claims (coding, 3D creation, financial modeling) are vendor-stated and carry no independent numbers in the checked docs, so they stay unlabeled until third-party evals exist.

Verification on Nov 4, 2026: platform.kimi.ai pricing page (rate card parsed from page data), model list and K3.1 Pro quickstart read directly; platform changelog read for the Batch API and K2 sunset notes; Hugging Face API and ModelScope probed for weight repos. Chinese press (Oct 24) used only for the launch date and the weights promise, attributed.

Quickstart: call kimi-k3.1-pro (official endpoint)

The OpenAI-compatible base URL is https://api.moonshot.ai/v1, and reasoning_effort is a top-level request field. The official quickstart ships the same chat-completions shape you already run; thinking output arrives as separate reasoning_content deltas before the final content when streaming. Keep complete assistant messages (not just content) across multi-turn calls and tool loops.

curl https://api.moonshot.ai/v1/chat/completions \
  --header "Authorization: Bearer $MOONSHOT_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "kimi-k3.1-pro",
    "reasoning_effort": "max",
    "messages": [{"role": "user", "content": "Introduce Kimi K3.1 Pro in one sentence."}]
  }'

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)
completion = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="max",
    messages=[{"role": "user", "content": "Prove that the square root of 2 is irrational."}],
)
print(completion.choices[0].message.content)

Both snippets mirror the official quickstart (Nov 4, 2026). The flagship tier unlocks after a top-up of at least $1; concurrency/RPM/TPM limits scale with your cumulative top-up tier.

Routing cheat sheet after the K2 sunset

Positioning against the other Chinese frontiers, with sources dated. Our GLM guide lists GLM-5.3 at $1.40/$4.40 per MTok (verified Aug 19, 2026) and our cheapest-API table lists DeepSeek V4.1-Flash off-peak at $0.22 per 1M tokens (verified Sep 2026) — both cheaper than K3.1 Pro's $2/$10. K3.1 Pro's counter is the 1,048,576-token window at a flat rate plus 3T-class scale; whether that wins per dollar depends on your harness, not on anyone's leaderboard. Budget note for K2.5 refugees: the value tier you were using is gone, and the nearest cheap Kimi-style option is batch-mode K3.1 Pro, not a K2 successor.

Overseas coverage, briefly

English-language coverage trailed the Chinese launch by days — third-party guides (e.g. FelloAI's K3.1 overview) appeared within the first week, while the verified rate card only became quotable from the official pricing page this week. Chinese financial press (Eastmoney, Sina) covered the Oct 24 launch and the weights promise; both are secondary sources used here only for dates. As of this check no Western lab-style benchmark of K3.1 Pro exists in the sources we could reach.

Access from outside China — and the usual disclaimer

This site tracks Chinese model APIs for international builders. The Kimi side of the ledger lives in the Kimi API guide; the cost arithmetic belongs in the cheapest-API guide and the comparison table. For Kimi self-hosting, the previous-generation K3 open-weight stack remains the only published option until K3.1 Pro weights land.

Independence disclaimer: ChinaModelAPI is an independent information site, not affiliated with Moonshot AI or any vendor. Rate-card figures above are from a direct read of the official pricing page on Nov 4, 2026; nothing here is a routing, availability or pricing promise, and relay aliases must be confirmed in each provider's dashboard.

Primary sources

• platform.kimi.ai — Model Inference Pricing (K3.1 Pro / Highspeed / K3 rate card, read directly Nov 4, 2026)
• platform.kimi.ai — Kimi K3.1 Pro Quickstart (model IDs, reasoning_effort, behavior contract; read directly)
• platform.kimi.ai — Model List (positioning: first K3.1-series model, 3T-class, native vision, 1M context)
• platform.kimi.ai — Platform Changelog (Batch API for kimi-k3.1-pro at 60%; K2 sunset notes; Aug 31 K2.5/moonshot-v1 404)
• Hugging Face — moonshotai organization (probed Nov 4, 2026: no K3.1 Pro repository as of the timestamp)
• Eastmoney (Oct 24, 2026) — launch report: Oct 24 release date and the by-Nov-4 weights promise (secondary, Chinese)
• 站内对照：Kimi API 指南 · K3 + GLM-5.2 开源栈拆解 · Harvey Tenet 迁移 K3 基座

FAQ (2026)