Skip to content

Pricing ​

What each plan costs and includes, what a sandbox hour and a million model tokens cost, and the limits that keep spending in check.

Plans ​

The hosted service uses a personal Yi Labs subscription. Each person has one workspace; these plans do not provide shared member seats.

PlanMonthly price per personMonthly creditYearly price
Free$0$5$0
Pro$20$25$192
Max$100$125$960
Ultra$200$250$1,920
EnterpriseBy contractBy contractBy contract

Yearly credit is still issued monthly. Manage top-ups, subscriptions, codes and invoices in your Yi Labs account. Account help explains expiry and eligibility. Free monthly and new-user credit cannot be used for Cloud.

How spending works ​

Review reserves funds from your personal wallet, settles actual usage and releases unused reservations. Insufficient available funds stop new usage. Plans and top-ups do not remove per-run budgets or resource safeguards; see Billing and usage.

Standalone deployments without Account use the organization billing guide; their organization plans are not hosted subscription options.

Unit prices ​

Prices are in US dollars, as billed.

Sandboxes, per hour, by class. Counted per second from when the virtual machine is created until it is deleted, at least 1 minute per virtual machine. x86 and ARM cost the same.

ClassPer hour
Small8.75¢
Standard (default)17.5¢
Large35¢
Extra large67.5¢
2X large$1.50
GPU (coming soon)$1.98

A custom spec costs what its size adds up to, counted the same way.

Models, per million tokens, for the tokens actually used: the model's list price in US dollars plus 30%. So $1 of model bill is billed as $1.30. Each project uses one of these, chosen in its header (see the platform's models).

ModelInputCached inputOutput
Gemini 4 Argon (Strongest · slower, the default)$5.20$0.26$26.00
GPT-6.1 Sol (Strong · faster)$2.60$0.13$13.00
Gemini 3.8 Flash (Balanced)$1.95$0.195$9.75
GPT-6 Luna (Economy)$0.13$0.013$0.65

Usage is billed exactly, token by token. Cached input is the part of a prompt the provider serves from its cache. A model's thinking counts as output tokens.

Storage: 12.5¢ per GB per month, charged by the day.

Traces uploaded: free for now (metered, see what is metered).

The AI Gateway: a call on a platform model is billed like any platform-model usage above (list price plus 30%); a call with your own provider key (x-provider-key) costs nothing on our side.

An example ​

A 30-minute run on a Standard sandbox, whose agent uses 1 million input tokens and 0.2 million output tokens through the platform, on each of the project's models:

Gemini 4 ArgonGPT-6.1 SolGemini 3.8 FlashGPT-6 Luna
Sandbox: 0.5 hours × 17.5¢8.75¢8.75¢8.75¢8.75¢
Model input: 1 million$5.20$2.60$1.9513¢
Model output: 0.2 million$5.20$2.60$1.9513¢
Totalabout $10.49about $5.29about $3.99about $0.35

When the platform judges the run, the judge costs extra, on the same model. Input the provider serves from its cache costs less. On the Free plan, the run on GPT-6 Luna uses about 7% of the month's $5; on Pro, the run on GPT-6.1 Sol about 21% of the month's $25.

What is metered ​

These are measured and appear on your usage page and invoice:

ItemUnitNotes
Sandbox timeHours, per class or custom specCounted per second from when the virtual machine is created until it is deleted, with a minimum of one minute per virtual machine. Pre-runs ("Check once") use sandbox time too. One price per class, whether it runs on x86 or ARM; a custom spec is priced by its size. A virtual machine that never reached the platform (it did not start up, or was stopped before it did) is not charged.
Judge modelUsageThe judge model grading replies and outcomes: in runs, in scorers that ask a judge (an SDK experiment's platform scorers, scorers added to an experiment, "Try it", calibration), in the Playground's scorers, and in online scoring (see the platform's models).
Platform models (AI under test)UsageCalls to one of the platform's models through the AI Gateway without your own key, and the Playground's prompt columns; a chat AI you test on a platform model instead of your own API and key; and the model calls of an agent in a repository or image project with no OPENAI_API_KEY in its secrets (see your agent's model), or on a local test runner with no OPENAI_API_KEY (and no OPENAI_BASE_URL) on that machine.
AssistantUsageThe in-app assistant and the platform AI's other steps for a project (drafting tests, "Generate rows with AI", "Let AI finish and check", automatic approval, diagnosis, the issues it finds and the like), on the project's model.
StorageGB-daysRun bundles and evidence kept in the platform.
Traces uploadedGB, with the number of spansTraces sent through the SDK, OpenTelemetry or the AI Gateway, measured as sent. Metered and shown on the usage page ("Traces uploaded"), but not charged for now.

Not metered by us:

  • Your own model calls. The agent under test, your app calling its provider directly or through the AI Gateway with x-provider-key, a task in an SDK Eval, and a coding AI you connect use your own model keys; you pay your provider directly. The exceptions are metered as above: a call on a platform model through the Gateway or the Playground, a chat AI on a platform model, and an agent in a repository or image project with no OPENAI_API_KEY in its secrets, or on a local test runner with none on that machine, whose model calls the platform answers.
  • Scorers that don't ask a judge. Rules, code scorers and trajectory matching cost nothing.
  • Local test runners. Tests on your own machine use no sandbox time (the judge model, and a platform model answering the agent, are metered as above).
  • Embeddings for similarity checks. The platform's embedding model, used by similar and answer_relevance, isn't billed.

What the platform runs again because of its own problem (re-runs caused by a platform error, the fill-in run after a sandbox was lost part way, the lost sandbox's time that left no result) is metered and shown on the usage page as usual, but not charged. See the platform's problem isn't charged.

The platform's models ​

These are models the platform calls with its own accounts on Microsoft Azure and Google Cloud, so you need no API URL or key of your own. Each project uses one, chosen in its header next to its name ("Strongest · slower", "Strong · faster", "Balanced" or "Economy", or "Follow the organization default"). That model judges the project's runs, does the platform AI's work for the project and answers the model calls of an agent with no OPENAI_API_KEY (your agent's model). It is also the starting choice for a new chat AI on a platform model. See Models.

ModelTierInputCached inputOutput
Gemini 4 ArgonStrongest · slower (the default)$4$0.20$20
GPT-6.1 SolStrong · faster$2$0.10$10
Gemini 3.8 FlashBalanced$1.50$0.15$7.50
GPT-6 LunaEconomy$0.10$0.01$0.50

These are the models' list prices in US dollars per million tokens, for reference. What you are billed is in the unit prices above: the list price plus 30%. Azure charges more above a long-context threshold, but the platform's calls stay far below it and are billed at the short-context price. The estimate before a run is in US dollars as billed and includes what a platform-model chat AI will cost. A run stops at its spending limit (below).

A chat AI built earlier on gpt-5-mini or DeepSeek-V4-Pro keeps that model and is billed at its price: $0.325 / $2.60 and $2.262 / $4.524 per million input / output tokens as billed. These two are no longer offered for new chat AIs.

Sandbox classes ​

The class sets the size of the virtual machine for a cloud run. Standard is the default; a project can choose another under "Settings" → "Sandbox". Every class, and a custom spec, is available on every plan, Free included.

Classx86 machineARM machinevCPUsMemoryDiskPer hourSuits
Smalle2-mediumn4a-standard-12 shared (x86) / 1 (ARM)4 GB30 GB8.75¢Light agents that only call external model APIs
Standard (default)e2-standard-2n4a-standard-228 GB30 GB17.5¢Most agents
Largee2-standard-4n4a-standard-4416 GB50 GB35¢Agents with a database and several tool backends
Extra largee2-standard-8n4a-standard-8832 GB100 GB67.5¢Small local models or heavy dependencies
2X largee2-standard-16n4a-standard-161664 GB200 GB$1.50Large builds, several heavy services, or local models that need more memory
GPU (coming soon)g2-standard-8None8, and one GPU with 24 GB of GPU memory32 GB100 GB$1.98Local models that need a GPU

On ARM a vCPU is a full core; on x86 it is a hyperthread. The GPU class is shown greyed out ("coming soon") under "Settings" → "Sandbox" until the platform's GPU image is ready and has passed its security checks; then the platform switches it on.

Custom spec ​

When no class fits, choose "Custom…", the last option under "Settings" → "Sandbox" → "Class", and set the size yourself: cores (an even number from 2 to 80), memory (2 to 640 GB), disk (30 to 2,000 GB) and GPUs (none, 1, 2 or 4, with 24 GB of GPU memory each). It is there for sizes no class has, such as 200 GB of memory, and works on every plan, Free included. A custom spec runs in London.

The platform gives you the cheapest machine at least that big. When that is more than you asked for, a line under the fields says what you get, for example "Actually allocated: 26 cores · 200 GB memory · 30 GB disk", beside the price, "Price: about $3.64/hour".

  • Without a GPU: the cores are the larger of the cores you asked for and the memory ÷ 8, rounded up to an even number (a core comes with at most 8 GB, and more cores cost less than extra memory); the memory is the larger of the memory you asked for and 2 GB per core. ARM follows the same rule up to 64 cores and 512 GB; for more, choose x86.
  • With GPUs (x86 only): the smallest machine below with that many GPUs that covers the cores and memory you asked for. Anything bigger is refused.

The price per hour is the machine plus the disk, billed like a class: per second, at least 1 minute per virtual machine.

Per hourx86ARM
Per core7.467¢6.3346¢
Per GB of memory0.8475¢0.7208¢
Per GB of disk0.025¢0.025¢
GPUsCoresMemoryMachine per hour
1416 GB$1.61
1832 GB$1.944
11248 GB$2.278
11664 GB$2.612
132128 GB$3.95
22496 GB$4.556
448192 GB$9.112

The disk of a machine with GPUs is 0.0328¢ per GB per hour. GPUs in a custom spec open together with the GPU class; until then a spec with a GPU is refused with the same "coming soon" note.

Some examples, each with a 30 GB disk:

Asked forAllocatedPer hour
x86, 4 cores, 200 GB26 cores · 200 GB · 30 GBabout $3.64
x86, 8 cores, 64 GB8 cores · 64 GB · 30 GBabout $1.15
x86, 16 cores, 32 GB16 cores · 32 GB · 30 GBabout $1.47
1 GPU, 16 cores, 64 GB16 cores · 64 GB · 1 GPU (24 GB) · 30 GBabout $2.62

The estimate before a run uses the custom spec's price.

Spending controls ​

ControlWith the Yi Labs wallet
Monthly and platform-AI daily spending capsNone; usage is funded by available wallet credit
Judge model spending per runNone; the platform's judge goes on while the wallet funds it, and stops with the usual low-balance message when it can't
Platform-model spending of the AI under test per runNone; a sandbox agent's or a chat AI's platform-model calls go on while the wallet funds them, and stop with the usual low-balance message when it can't
Each runStep and model budgets remain; the launch page shows this run's limits and estimate
Sandbox resourcesConcurrency, runtime and resource safeguards remain; see Organization Policy and the launch page for current values. Account plans are not the old organization resource tiers
Online scoringNo daily budget by default; a budget you explicitly set still applies
Usage and paymentsProduct usage is under Organization → Usage; shared wallet, subscriptions, top-ups and invoices are managed in Account

See Billing and usage. Standalone deployments without Account retain their own organization credit and spending limits; see the standalone guide.

The legal texts on this site are drafts pending legal review and are not yet in force.