News / Model Watch · Kimi agent platform GA
Kimi Hosted Agents Goes GA: Moonshot Bills the Agent Harness Itself
Moonshot AI (月之暗面) moved Kimi Hosted Agents (托管智能体) to general availability on Nov 5, 2026, turning the agent runtime — the harness — into a cloud product: 7×24 hosted sandbox environments, long-running sessions and multi-agent collaboration that companies rent instead of build. For API buyers the headline is the bill, not the brochure: Kimi's pricing page prices a session as sandbox runtime hours × ¥0.072 plus model inference at standard token rates plus per-call plugin fees (¥0.001-0.45 per call). That makes Moonshot the first Chinese model lab to publish a rate card that bills the agent runtime itself as a separate object.
Verified directly (Nov 7): the official pricing page states the formula total = sandbox hours × ¥0.072/h + model inference + Σ(plugin calls × plugin price), with each session in a dedicated cloud sandbox. The quickstart documents the REST surface — api.moonshot.cn/v1 agents, environments, sessions, artifacts and session filesystem, SSE event streams with Last-Event-ID resume, header kimi-api-version: 2026-09-01-beta — plus the kha CLI (runs inside Kimi Code, Claude Code, Codex) and a web console that needs no API key. An Agent = model + system prompt + tools + MCP servers + skills/plugins; third-party credentials bind via Vault; outputs land as downloadable artifacts (the official example produces a .pptx). Official counts: 140+ skills, 30+ plugins, 8 first-party finance agent templates (investment modeling, deep research, morning briefing, IPO tracking, holdings monitoring). Kimi Inference Hook — a pre-inference check on task inputs and tool returns under enterprise-defined rules — is separately gated to enterprise-certified users.
Attributed (Netease Tech reading the official pricing page and changelog): plugin price points run ¥0.001-0.45 per call — web search ¥0.03, image search ¥0.045, financial-data plugins ¥0.35/¥0.45 — with Wind, THS iFinD and S&P Global named as data sources; the changelog shows a September beta that targeted domestic enterprise-certified users and lists session history, sandbox, failure recovery, timeout retry and multi-agent orchestration. (The pricing page renders its tables client-side; our direct read captured the formula and the ¥0.072 sandbox rate but not the rendered plugin table.)
Not established: which model drives hosted sessions — inference is billed "at the chosen model's rate" and the docs name no default; no USD rate card — our Nov 7 check of the international platform (platform.kimi.ai) surfaced no hosted-agents pricing page, so the ¥-denominated card is the only published one; whether idle (non-executing) sandbox time bills — the page prices "runtime hours" without an idle policy; any independent benchmark or review — none exists yet.
The facts, graded (Nov 7)
- The launch, verified. Moonshot's official open-platform WeChat account announced GA on Nov 5, 2026 — "Harness as a Service", 7×24 hosted environments, long-horizon tasks, multi-agent collaboration. 36Kr's newsflash ran 12:28 that day; Shanghai Securities News 14:13; Beijing Business Today and People's Finance the same day. Shanghai Securities News frames it as Moonshot's B-side step "from selling model-inference API to selling agents" (attributed analysis, not an official claim).
- The rate card, verified by direct read. Official pricing page: sandbox runtime × ¥0.072/hour + model inference at standard rates + per-call plugin prices. A sandbox session running 30 days straight costs ¥51.84 in runtime fees alone (our arithmetic: 720 h × ¥0.072 — matches Netease Tech's estimate; inference and plugins extra).
- Plugin price points, attributed. ¥0.001-0.45 per call: web search ¥0.03, image search ¥0.045, financial data ¥0.35/¥0.45 (Wind, THS iFinD, S&P Global named). Source is Netease Tech's read of the official page; the tables render client-side and were not in our direct fetch.
- The API surface, verified by direct read. Create an environment (
{"type": "cloud"}), create a session bound to agent + environment, subscribe the SSE event stream before sending the task, collect artifacts, browse the session filesystem. Events include session.status, agent.thinking, agent.tool_use, agent.message; streams resume via cursor or Last-Event-ID. The platform is stateful — history, sandbox state and files persist server-side; sessions archive before deletion. - Beta timeline, corroborated. Netease Tech (citing the open-platform changelog) reports a September beta for domestic enterprise-certified users. The API's own version header —
kimi-api-version: 2026-09-01-beta— independently matches that window. - Skills, plugins, templates, official. 140+ skills and 30+ plugins spanning web retrieval, professional data, office documents and content generation; enterprises can attach private tools and data sources. The first batch ships 8 finance agent templates: investment modeling, deep research, morning-brief generation, IPO tracking, holdings monitoring, regulatory-case dailies and fund reconciliation among them.
- Security posture, official with one gate. Dedicated sandboxes, network access controls, operator-defined boundaries, mid-session abort, audit export and per-project billing reports. Kimi Inference Hook checks task inputs and tool returns before every model inference under enterprise-defined rules — enterprise-certified users only (contact-sales).
- Platform context from the same docs. The chat pricing page re-read today confirms: K3-series context is 1,048,576 tokens; batch inference runs at 60% of standard for kimi-k3.1-pro and kimi-k3; K2 stopped new users on Oct 31 and retires fully by end of December 2026. The hosted platform lands as Kimi's own API lineup consolidates on K3 (K3.1 Pro launch brief, Oct 24).
Verification on Nov 7, 2026: Kimi open-platform WeChat announcement, hosted-agents pricing page and quickstart read directly; chat pricing page re-read; plugin price points, September-beta scope and Microsoft/Salesforce context per Netease Tech (via Sina), 36Kr and Shanghai Securities News; no independent benchmark exists for the hosted service.
Context: the agent runtime becomes a line item
Kimi is not inventing the category — it is the first Chinese lab to put a published rate card on it. Salesforce commercialized Agentforce in October 2024 starting at $2 per conversation. Microsoft opened Copilot harness in Foundry Agent Service as a public preview on Oct 28, 2026, with usage-based billing — sandbox compute and inference priced separately — from Dec 1. DeepSeek took the research route instead: its Oct 31 Super-agent writeup and open dsh-superagent prototype document the architecture without selling a product (our coverage: DeepSeek superagent prototype, DeepSeek Harness v0.1).
The economics change shape. Token billing makes an agent's idle time free; hour billing makes a session's existence cost money before any token moves — ¥1.728 per sandbox-day at list (24 × ¥0.072), before inference and plugin fees. For always-on agents (monitoring, daily briefings, holdings watches — exactly the templates Kimi ships), the floor cost is now the sandbox, and the plugin mix is the variable. For episodic agents, tokens still dominate. Cost models written before Nov 5 don't include this line.
Timing matters too: the platform lands two weeks after K3.1 Pro (Oct 24) and its finance-template suite rides the model's finance push — the same week Tsinghua/Zhipu's HarnessPro showed harnesses themselves can be auto-evolved (Nov 2). Agent infrastructure is where Chinese labs are now competing below the model layer.
What this means for builders
Model the floor first. If your agent holds a session open, multiply hours by ¥0.072 and add the plugin mix at per-call rates (web search at ¥0.03 adds up fast at high frequency); only then stack token costs at the standard chat rates (kimi-k3.1-pro lists at $2/$10 per MTok on the international pricing page, verified Nov 4). Run one real session end-to-end and read the billing report before committing a workload.
The integration protocol, from the official quickstart: create an environment with {"type": "cloud"}, create a session binding agent_id + environment_id, open the SSE stream before sending the user message, pull results via artifacts (GET /v1/artifacts?session_id=..., then the content endpoint), and keep secrets in Vault rather than prompts. The kha CLI wraps the same flow and runs inside coding agents.
Watch five signals: a USD/international rate card; an idle-time billing clarification; disclosure of the driving model; Inference Hook availability beyond enterprise-certified users; plugin-catalog expansion. Any of them moves and this page gets an update entry.
Access from outside China — and the usual disclaimer
This site tracks Chinese model APIs for international builders: official endpoints, pricing and open-weights status. Hosted Agents is currently a ¥-denominated product on the domestic platform (platform.kimi.com); our Nov 7 check did not surface a hosted-agents pricing page on the international platform (platform.kimi.ai). The model APIs behind it are the ones in the comparison table and the cheapest-API guide.
Independence disclaimer: ChinaModelAPI is an independent information site, not affiliated with Moonshot AI, Kimi or any vendor. Rate-card and API facts above are from direct reads of the official pages; plugin price points and beta scope are attributed to their sources; nothing here is an availability, routing or pricing promise.
Primary sources
- Kimi open platform (WeChat, Nov 5, 2026) — the GA announcement this brief verifies; read directly
- platform.kimi.com — Hosted Agents pricing (billing formula, ¥0.072/sandbox-hour; read directly Nov 7)
- platform.kimi.com — Hosted Agents quickstart (REST API, SSE events, kha CLI; read directly Nov 7)
- platform.kimi.com — chat pricing (K3 context 1,048,576, batch 60%, K2 sunset; read directly)
- Netease Tech (via Sina, Nov 5, 2026) — plugin price points ¥0.001-0.45, September beta scope, Wind/iFinD/S&P Global, Microsoft/Salesforce context
- 36Kr newsflash (Nov 5, 2026, 12:28) + Shanghai Securities News via Eastmoney (14:13) — launch-date confirmations
- 站内对照：Kimi K3.1 Pro launch brief · DeepSeek superagent prototype · DeepSeek Harness v0.1 · Kimi API guide
FAQ (2026)
What is Kimi Hosted Agents?
Moonshot's hosted agent runtime, generally available Nov 5, 2026: 7×24 cloud-sandbox sessions, long-running tasks and multi-agent collaboration, consumed via REST API, the kha CLI or a web console. Kimi pitches it as "Harness as a Service".
How much does Kimi Hosted Agents cost?
Three parts: sandbox runtime hours × ¥0.072 + model inference at standard token rates + plugin calls × unit price (¥0.001-0.45 per call). A sandbox running 30 days costs ¥51.84 in runtime fees alone (our arithmetic from the official rate).
Can I use Kimi Hosted Agents outside China?
The published rate card is RMB-denominated on platform.kimi.com; our Nov 7, 2026 check surfaced no hosted-agents pricing on the international platform (platform.kimi.ai). The GA quickstart needs only a platform account and API key; the September beta targeted domestic enterprise-certified users.
Which model powers the hosted agents?
Not stated. Docs bill inference at "the chosen model's rate" without naming a default. Kimi's API lineup is led by kimi-k3.1-pro ($2/$10 per MTok, 1,048,576-token context, verified Nov 4, 2026), so expect K3-series rates to apply.
How do I call the Kimi Hosted Agents API?
REST at api.moonshot.cn/v1 (agents, environments, sessions, artifacts) with header kimi-api-version: 2026-09-01-beta, SSE event streams per session, plus the kha CLI and a web console that needs no API key.
How does it compare with other agent-harness platforms?
Microsoft's Foundry Agent Service runs Copilot harness in public preview with usage billing from Dec 1, 2026; Salesforce Agentforce started at $2 per conversation (Oct 2024); DeepSeek publishes harness research and an open prototype instead of a product. Moonshot is the first Chinese lab to publish a hosted-harness rate card.