Moonshot AI Kimi API Review 2026: K3.1 Pro and the Full Model Lineup
Moonshot AI is the Beijing-based lab that has shipped frontier-class models under the Kimi brand since 2024. With the late-October 2026 release of Kimi K3.1 Pro, the company now fields a 3-trillion-class open-weight flagship with a 1-million-token context window, published per-million-token pricing in both USD and CNY, and an API that speaks both the OpenAI and the Anthropic wire format. We verified every price, limit, and endpoint in this review against the Moonshot AI international documentation, the Chinese platform tables, and the live OpenRouter catalog on November 4-5, 2026.
The headline is that K3.1 Pro is a serious long-context coding model at roughly mid-tier Chinese-lab pricing, with an unusually aggressive cache discount. The fine print matters more than usual here: the official and OpenRouter price lists agree to the cent on the flagship but diverge sharply on the older K3 and K2 models, and the entire K2 generation is being retired on a published schedule. This review separates what is stable, what is changing, and where the cheapest way to call each model actually is.
What Moonshot AI actually sells
The Kimi API Platform is a pay-as-you-go inference service with no subscription plan, and Moonshot is explicit that it should not be confused with the Kimi consumer membership or the Kimi Code developer product. Billing is per million tokens consumed, with separate meters for cache writes, cached input hits, uncached input, and output. Repeated request prefixes are cached automatically; the default cache time-to-live is 5 minutes, with a paid 1-hour option for workloads that revisit a prefix less frequently.
The catalog spans two generations. The current K3 series is Kimi K3.1 Pro (the new flagship), Kimi K3.1 Pro Highspeed (the same model at roughly 2-3x the speed for double the token price), and the original Kimi K3. All three support a 1,048,576-token window. The legacy K2 series is Kimi K2.7 Code (plus its Highspeed twin) and Kimi K2.6, each capped at 262,144 tokens. On October 31, 2026, Moonshot stopped opening K2 models to newly registered users, and it plans to retire every K2 model by the end of December 2026. Existing accounts keep access during the transition, but a project started today should target K3.
Official pricing: USD and CNY
The table below is the official international (USD) price list, per 1M tokens, excluding taxes. The same page publishes a CNY table for the domestic platform; the two are not a clean currency conversion, so we report both rather than deriving one from the other.
| Model | Context | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
| Kimi K3.1 Pro | 1M | $2.00 | $0.20 | $10.00 |
| Kimi K3.1 Pro Highspeed | 1M | $4.00 | $0.20 | $20.00 |
| Kimi K3 | 1M | $3.00 | $0.30 | $15.00 |
| Kimi K2.7 Code | 256K | $0.95 | $0.19 | $4.00 |
| Kimi K2.6 | 256K | $0.95 | $0.16 | $4.00 |
On the Chinese platform the same slots are priced in renminbi: K3.1 Pro is CNY 13 input and CNY 68 output with a CNY 1.30 cache-hit rate; K3 is CNY 20 and CNY 100 with CNY 2.00 cached; K2.7 Code and K2.6 are CNY 6.50 and CNY 27. Cache writes are billed separately on the K3 series at two TTL tiers, for example CNY 13 for a 5-minute write and CNY 26 for a 1-hour write on K3.1 Pro, so a long-lived prefix is not free to populate.
Two structural points stand out. First, the cache-hit discount on K3.1 Pro is 90 percent: $0.20 per million against a $2.00 base input price. For an agent that carries a large, stable system prompt or a repeated repository prefix across many turns, effective input cost drops by an order of magnitude. Second, the Batch API on kimi-k3.1-pro and kimi-k3 bills every consumed token at 60 percent of the synchronous rate, which stacks on top of cache-hit pricing for offline evaluation and data-enrichment pipelines.
A worked cache-cost example
To see why caching is the lever rather than list price, consider a coding agent that sends a fixed 200,000-token repository prefix plus 5,000 tokens of new instructions on every one of 100 turns. Uncached, the prefix alone costs 100 x 200,000 / 1,000,000 x $2.00 = $40.00 in input tokens on K3.1 Pro, with the changing tail negligible. Because the platform caches the repeated prefix automatically, the first turn pays a one-time cache-write charge while every later turn reads that prefix at $0.20 per million: roughly 99 x 0.2 x $0.20 = about $3.96 plus the write, instead of $40.00. The saving scales directly with prefix size and turn count, which is why ordering the request correctly matters.
Moonshot matches on exact prefixes, so the cache only helps if stable content is placed first and volatile content last. Keep the system prompt, tool definitions, reference material, and codebase at the front; put the user's latest question, tool results, and any timestamps or random identifiers at the end. Once any token in the prefix changes, everything after that position loses its hit. The default 5-minute TTL is renewed on every hit, so a back-and-forth agent stays warm for free, whereas a scheduled job that revisits a prefix hourly should pay for the 1-hour write tier. Caches are isolated by organization and cannot be cleared by hand; an unused entry simply expires.
Official list versus OpenRouter: read the fine print
Because Kimi is also available through OpenRouter, we compared the relay catalog against the official list. The result is a useful warning about assuming a single price. For the flagship, the two sources agree exactly: OpenRouter lists moonshotai/kimi-k3.1-pro at $2.00 input, $0.20 cached, and $10.00 output, matching the documentation to the cent. For the older models the lists diverge materially: OpenRouter shows Kimi K3 at roughly $0.19 input and $16.00 output against an official $3.00 and $15.00, K2.7 Code at about $0.60 and $3.50 against $0.95 and $4.00, and K2.6 at about $0.65 and $3.41 against $0.95 and $4.00.
There is no contradiction to resolve by picking one number. The official platform is the authoritative billing surface for a direct contract, domestic CNY pricing, the rate-limit ladder, and the Batch API; the OpenRouter listing is a live third-party market rate that can run cheaper on input or slightly dearer on output depending on the model. Budget production against the official table if you integrate directly, and treat any relay quote as a separate line item you re-check at purchase time rather than a permanent discount.
Rate limits and the top-up ladder
K3 series models require a successful recharge before they can be called, and the cumulative lifetime top-up amount sets the account tier. A small new-user credit exists for evaluation, but it cannot be spent on K3, so plan to fund the account before any production workload. The published ladder runs from Tier0 to Tier5:
| Tier | Cumulative top-up | Concurrency | RPM | TPM | TPD |
|---|---|---|---|---|---|
| Tier0 | $1 | 1 | 3 | 500,000 | 1,500,000 |
| Tier1 | $10 | 15 | 100 | 2,000,000 | Unlimited |
| Tier2 | $20 | 40 | 100 | 3,000,000 | Unlimited |
| Tier3 | $100 | 50 | 200 | 3,000,000 | Unlimited |
| Tier4 | $1,000 | 60 | 200 | 4,000,000 | Unlimited |
| Tier5 | $3,000 | 100 | 300 | 5,000,000 | Unlimited |
The biggest jump is the first ten dollars, which takes an account from 3 RPM and a 1.5M daily token cap to 100 RPM and unlimited daily tokens. Web search is metered separately by queries-per-second and does not consume these limits. Vouchers and promotional credit explicitly do not count toward the cumulative top-up total, and Moonshot reserves the right to apply risk-control throttling to abnormal traffic.
Model details and request parameters
Moonshot describes K3.1 Pro as its most capable 3-trillion-class open-weight model, post-trained from Kimi K3 with substantially larger post-training compute, with particular gains in coding, 3D creation, and financial research and modeling. Thinking is always on. You set inference depth with the top-level reasoning_effort field, which accepts low, high, and max and defaults to max; there is no way to disable reasoning on the K3 series. The model supports tool calls through tool_choice (including required), dynamic tool loading where complete tool definitions are placed in a system message, JSON Schema structured output, and the Responses, Chat Completions, and Messages APIs.
Several sampling parameters are fixed and must be omitted: temperature 1.0, top_p 0.95, n 1, presence_penalty 0, and frequency_penalty 0. Passing any other value returns an error. max_completion_tokens defaults to 131,072 and can be raised as far as 1,048,576. Vision accepts an array of content parts rather than a serialized string, and images or videos can be supplied as inline base64 data, as public URLs, or through the platform file API using the ms:// reference protocol; SVG is rejected and should instead be pasted as XML text.
The Highspeed variant is the same model with identical parameter constraints, differing only in output speed. It is worth its 2x token premium for interactive coding assistants and latency-sensitive chat, and not worth it for batch document processing, where standard K3.1 Pro plus the Batch discount is the better value.
Batch mechanics and limits
The Batch API is worth treating as its own product rather than a price toggle. A batch job is submitted as a file of requests and processed asynchronously and concurrently, so it is not bound by the real-time concurrency limits that cap interactive traffic; that makes it the practical route for bulk evaluation runs, offline labeling, and backfills that would otherwise queue for hours. Every token consumed inside a completed job, including output, is billed at 60 percent of the standard rate. The main model-pricing page flags kimi-k3.1-pro and kimi-k3 as the batch-supported models; the dedicated batch page also enumerates the K2.7 Code and K2.6 models, which are in any case being retired, so new batch pipelines should target K3.
Three operational caveats apply. First, batch work on K3.1 Pro and K3 still incurs cache-write charges, whereas the K2 models do not support cache writes; and because jobs are processed concurrently, cache hits inside a batch are best-effort rather than guaranteed, so cost the discount at list price when you forecast. Second, every job carries a completion window, and work that has not finished when the window expires is lost rather than retried automatically. Third, the 60 percent rate does not change the 1M token per-request context ceiling, so unusually large requests still need to be split ahead of submission. The pattern that works best is to develop and validate a prompt synchronously on a few examples, then fan the full set out through Batch and retrieve results when the job completes.
How Kimi K3.1 Pro compares
On the vendor-published benchmark that matters most for this audience, Terminal-Bench 4.0, Moonshot reports K3.1 Pro at 65.7 percent, ahead of OpenAI GPT-6 Astra at 58.2 percent and narrowly behind Anthropic Claude Opus 5.5 at 66.4 percent. On its internal Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scores 61.2 percent against 55.95 percent for the original K3. These are vendor-reported figures and should be re-checked as the leaderboards move, but they position K3.1 Pro inside the frontier coding-agent cluster rather than on its budget edge.
On price the positioning is clearer. GPT-6 Astra is $10 input and $50 output per million on OpenRouter, so K3.1 Pro at $2 and $10 is about one-fifth the output cost while scoring higher on the benchmark Moonshot cites. Claude Opus 5.5 is $4 and $20, roughly double K3.1 Pro on both meters, with a 0.7-point Terminal-Bench lead. Against the cheapest open-API coding models the trade-off reverses: DeepSeek V4.1 Flash also offers a 1M window but at about $0.045 input and $0.50 output per million, a small fraction of K3.1 Pro's price, while DeepSeek's older V3.2 sits at 128K context. The honest read is that DeepSeek remains the raw-cost champion, GPT-6 Astra and Opus 5.5 set the capability ceiling, and K3.1 Pro occupies the capability-per-dollar middle for long-context agentic work, especially once the 90 percent cache discount applies.
Getting started
Fund the account, create an API key in the console, and point any OpenAI-compatible SDK at the international base URL. The domestic platform uses a separate host, so use the endpoint that matches where your account and billing live.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_KIMI_API_KEY",
base_url="https://api.moonshot.ai/v1"
)
completion = client.chat.completions.create(
model="kimi-k3.1-pro",
messages=[
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write a haiku about distributed systems."}
],
reasoning_effort="high"
)
print(completion.choices[0].message.content)When you switch to Highspeed, only the model id changes; the parameter constraints are identical. Teams migrating from K2 should remove the old K2-specific thinking configuration and pass the complete assistant message, including reasoning_content, back into messages on multi-turn calls, because Preserved Thinking is mandatory on K3.
Who it is for
Kimi K3.1 Pro is a strong fit for teams building long-context coding agents, repository-scale review tools, and financial-research pipelines that benefit from a 1M-token window at a mid-tier price. The combination of a 90 percent cache discount, a 40 percent Batch discount, and published rate-limit economics makes sustained agentic workloads with a stable prefix unusually cheap to operate, and the dual OpenAI/Anthropic compatibility keeps migration friction low.
It is a weaker fit when you need a no-thinking fast response, want a usable free tier without a top-up, or are optimizing purely for token cost on tasks where a budget open-API model is already good enough. Mainland-China teams should use the CNY-denominated domestic platform rather than the international USD endpoint. For everyone else, K3.1 Pro is one of the most operationally transparent long-context models available: the prices, tiers, cache terms, and even the retirement date of the old generation are all published, which is more than most frontier labs provide.
Sources: Moonshot AI international documentation (model inference pricing, recharge and rate limits, Kimi K3.1 Pro quickstart, model parameter reference, vision input guide), the Kimi API open platform CNY pricing tables, and the live OpenRouter model catalog. All prices and limits verified November 4-5, 2026; benchmark scores are vendor-reported.