Pricing · preview定价 · 预览期
We bill for the players, not the notes. 按演奏席位收费,不按音符。
The engine is free and always will be. The cluster charges for concurrent execution — how many turns you can have in flight at once. Model tokens stay between you and your provider; bring your own key. 引擎永远免费。集群按并发执行收费——也就是你能同时跑多少个 turn。模型 token 的账仍然在你和你的模型厂商之间;自带 key 即可。
Solo独奏
The engine, in your own process. 引擎,跑在你自己的进程里。
- Full tool loop, approvals, steering, guard rails完整工具循环、审批门、轮间注入、护栏
- Any Anthropic-compatible endpoint任意 Anthropic 兼容端点
- MCP tools over your own HTTP endpoint经你自己的 HTTP 端点接 MCP 工具
- Self-host the cluster too, if you want all of it想要整套的话,集群也可以自建
Ensemble合奏
most teams多数团队The hosted cluster. You send turns; we run them. 托管集群。你发 turn,我们跑。
- Session serialisation, leases and epoch fencing included — not an add-onsession 串行化、租约、epoch fencing 都在里面——不是加购项
- Live stream, control channel, orphan reaping实时流、控制通道、孤儿回收
- Bring your own model key — we never mark up tokens自带模型 key——我们不在 token 上加价
- Postponed turns are free: yielding isn't billable work让路的 turn 不计费:让路不是工作量
- Your database stays yours; we hold no authoritative state数据库仍然只归你;我们不持有任何权威状态
Orchestra交响
Your cloud, your compliance boundary. 你的云、你的合规边界。
- Dedicated brokers and Redis, or deployment inside your VPC独享 broker 与 Redis,或直接部署进你的 VPC
- Private model endpoints and on-prem inference私有模型端点与本地推理
- Capacity planning against your real turn mix按你真实的 turn 结构做容量规划
- Support with a named engineer and an SLA指名工程师支持 + SLA
What counts怎么算
What a "concurrent run" means「并发 run」到底指什么
One run is one turn being executed: from the moment a worker takes the session lease to the moment it publishes a terminal fact. Ten concurrent runs means ten turns can be mid-flight at the same instant — not ten users, not ten sessions, and not ten thousand messages a day. 一个 run 就是一个正在执行的 turn:从 worker 拿到 session 租约的那一刻,到它发出终态事实为止。十个并发 run 意味着同一瞬间最多十个 turn 在飞——不是十个用户、不是十个 session,更不是每天一万条消息。
Idle sessions cost nothing空闲 session 不花钱
A session with no turn in flight occupies no slot. Products where users think for a minute between messages need far fewer slots than their user count suggests. 没有 turn 在跑的 session 不占额度。用户消息之间要想上一分钟的产品,需要的额度远少于用户数给人的直觉。
Waiting on a human is still running等人点头也算在跑
A turn parked at an approval gate holds its slot — it holds the lease, the context and the stream. If your workflow leaves approvals open for hours, size for it, or turn the gate off for the tools that don't need it. 停在审批门前的 turn 仍占额度——它握着租约、上下文和流。如果你的流程会让审批挂几个小时,按这个规模买,或者给不需要审批的工具关掉这道门。
Retries you didn't cause are free不是你造成的重试免费
Duplicate deliveries, rebalance handoffs and postponements are our plumbing, not your workload. You are billed for turns that actually execute. 重复投递、rebalance 交接、让路都是我们的管道问题,不是你的工作量。计费只算真正执行的 turn。
Tokens are not ours to selltoken 不是我们的生意
You bring the model key; the bill for inference comes from whoever serves your model, at their price. We report usage per run so you can attribute it, and that's the whole of our involvement. 模型 key 由你提供,推理账单来自给你服务的模型厂商,按他们的价格。我们只按 run 汇报用量供你归因——我们的参与到此为止。
FAQ
Questions people actually ask大家真的会问的问题
Can I move off the cloud later?以后能从云上搬走吗?
Yes, and the exit is the point of the architecture. The cluster is a pure executor with no database of its own — your history, your runs and your facts already live in your system. Self-hosting is the same open-source binary and the same five channels. 能,而且「能走」正是这套架构的用意。集群是纯执行器,自己没有数据库——历史、run、事实本来就在你的系统里。自建用的是同一个开源二进制、同样的五个通道。
Do you see my prompts?你们能看到我的 prompt 吗?
Run context passes through workers in memory in order to be executed — there is no way around that — but it is not persisted by us. Nothing is written to storage we own beyond operational metrics and the facts we hand straight back to you. 运行上下文为了被执行,必然在 worker 内存里经过一遍——这绕不开——但我们不做持久化。除了运维指标,以及原样交还给你的事实,不会有任何东西写进我们的存储。
What happens when a worker dies mid-turn?worker 跑到一半死了会怎样?
Its lease expires, the broker redelivers, and another worker takes the turn from the top. If the dead one comes back and tries to publish, the session epoch fence rejects it — a superseded run cannot overwrite fresh state. The reaper reports the orphan either way. 它的租约过期,broker 重投,另一个 worker 从头接过这个 turn。如果那台死机复活并试图发布产出,session epoch 围栏会拒收它——被取代的 run 盖不掉新状态。无论如何 reaper 都会上报这个孤儿。
Can I use a model that isn't Anthropic-shaped?能用非 Anthropic 形状的模型吗?
Write a Provider. It is one method plus a delta callback, and the loop stays untouched. That is the same seam we use for endpoints, and the same one planned for subscription-backed CLIs.
写一个 Provider 就行。一个方法加一个增量回调,循环本身不动。这条缝我们自己接端点用的是它,规划中的订阅型 CLI 用的也是它。
Is there a free tier on the cloud?云上有免费额度吗?
During preview, one concurrent run is free while you evaluate. If one is genuinely enough, keep it — the engine is free anyway, and we would rather you self-host than pay us for nothing. 预览期评估阶段送一个并发 run。如果一个就真的够用,那就留着——引擎本来就免费,与其让你花钱买不需要的东西,不如你自建。