Skip to Content
ZeroLeaksDocs

Models and pricing

Model access, pay-as-you-go pricing, service limits, and how Shield handles submitted content.

Models

Model IDAccessBehavior
shieldFree with an account, API key, and research acknowledgement.The small Shield classifier with deterministic rules.
shield-baseEvery active paid ZeroLeaks plan, or pay as you go.The Base classifier with the same binary attack category.
shield-largeEvery active paid ZeroLeaks plan, or pay as you go.An ensemble that combines the larger classifier and Base.
shield-tieredEvery active paid ZeroLeaks plan, or pay as you go.Runs Shield first and escalates uncertain scores to Large.

Paid personal subscriptions and qualifying paid workspace memberships provide paid Shield access. Existing trial and subscription-status rules apply. A key remains subject to its owner's current access. Selecting a paid model does not purchase a subscription.

Shield uses S15e and Base uses SB1, unchanged since the 2 October 2026 release. From Thursday, 29 October 2026, Large uses Shield 1.1: the same ensemble of 0.6 × L7a (Qwen3-1.7B) plus 0.4 × SB1, with L7a retrained and served unquantized. All tiers include rules v2 and use a model threshold of 0.5. Tiered escalates to Large, so the texts it escalates are scored by Shield 1.1.

On the 35-dataset benchmark from the model card, Large's mean balanced accuracy is 91.5% with Shield 1.1, against 90.4% for Shield 1 Large, and it flags 4.1% of benign texts instead of 5.4%. That benchmark was also used to choose between training rounds; a fresh held-out test follows. The Shield 1.1 model card addendum reports every dataset, the regressions and the limitations. The Shield 1 model card covers the other tiers and the comparison with other detectors.

Tiered escalates small-model scores in the interval [0.01, 0.97) to Large. These models classify prompt injection and jailbreaks under one binary prompt_injection label. They do not provide independent probabilities for individual attack types.

Pricing

Pay as you go

Accounts without a paid plan can turn on pay as you go at /dashboard/shield to use shield-base, shield-large and shield-tiered with their existing API keys. You add a card once and are billed monthly for what you used. Each model has its own price per 1 million billable tokens, shown on that page. A text's billable tokens are its input tokens, with a minimum of 256 per text. Only successful classifications are billed; requests rejected by authentication, validation, rate limits or errors are not. shield stays free. Paid plans include every model and are never metered. Pay-as-you-go requests are handled like paid traffic: submitted content is not retained or used for training. Without a paid plan or pay as you go, a paid model returns 403 model_not_available.

Contributor pricing

Add :contribute to a paid model ID, for example shield-large:contribute, to pay half price on pay as you go. These IDs use the same models and return the same results, and responses echo the ID you sent. In exchange, Shield keeps exactly what it keeps for the free shield model: a redacted excerpt of each flagged attack, for security research and to train Shield models, under the same acknowledgement. Without that acknowledgement on your key, the request returns 403 research_consent_required. Paid plans can use :contribute too: excerpts are kept the same way and billing does not change. Model IDs without the suffix keep no submitted content and bill at the full price. GET /v1/models lists every ID with its pay-as-you-go price.

List models

curl https://api.zeroleaks.ai/v1/models \
  -H "Authorization: Bearer $ZEROLEAKS_API_KEY"

The response follows the OpenAI model-list format. Each entry includes:

{
  "id": "shield-large",
  "object": "model",
  "owned_by": "zeroleaks",
  "shield": {
    "requires_paid_plan": true,
    "available": false
  }
}

The example omits the created timestamp. shield.available reflects your account's model entitlement. Free inference also requires the current research acknowledgement on the key; listing a model does not grant that acknowledgement.

Limits

Shield has no monthly call meter and no routine per-minute quota. Service concurrency and queue length are bounded. A full queue returns a retryable 429; an unavailable inference service returns 503. Retry with backoff and respect Retry-After when supplied.

For input and request body limits, see moderations limits. These limits do not guarantee complete token coverage. Read each result's shield.coverage before accepting a long document.

Data handling

Request activity

The Shield activity dashboard shows request counts, flagged inputs, errors, token usage, and average successful-request latency per API key. Filters use UTC calendar days. A request record does not contain the text you sent. It records the date, the time the request finished, and the classification score. The history view also shows the model, endpoint, HTTP status, input counts, coverage truncation, latency, and request ID. The log does not store submitted text, attack excerpts, authorization headers, or secret keys. The only stored text is a redacted excerpt from the free Shield model without a paid plan or from a :contribute model ID, described under submitted content below.

For new requests, select a latency value to see authentication, input reading, capacity admission, inference, and background cleanup times. Response latency measures the API handler until its response is ready; it excludes network travel between your application and Shield, and excludes background cleanup. Earlier records retain their original latency measurement and do not have a breakdown.

Classification responses also include a Server-Timing header with shield (total handler time), auth, body, admission, and inference durations in milliseconds. Only stages that ran are included. Cleanup starts immediately when inference finishes and is tracked in the background, so its timing is available in request activity rather than in the response header. Capacity remains reserved until release completes or its existing lease expires.

Request metadata and daily aggregates expire within 30 days and are removed by the next hourly cleanup. Recording happens after the API response and is best effort; it is not a billing ledger or a complete audit trail. History begins when activity tracking is enabled. Requests rejected before authentication cannot be attributed to a key and do not appear.

A key owner can share read-only activity with specific verified email addresses. This exposes all retained and future activity for that key, so use a dedicated key for each customer. Viewers cannot manage the key or access other account resources. The owner can remove access from the activity page. Sharing does not send an email; send the viewer the activity link separately.

Submitted content

Content handling follows the account's paid access and whether the model ID ends in :contribute.

RequestSubmitted content retained?
Paid plan, or pay as you go on shield-base, shield-large or shield-tiered; model ID without :contribute; any verdictNo.
Any unflagged resultNo.
No paid plan and using shield, or any account using a :contribute ID; flagged result; acknowledged keyA redacted excerpt of up to 2,000 characters may be retained for up to 12 months, for security research and to train Shield models.

Free use and :contribute IDs require an explicit acknowledgement when creating the key. A key without it receives research_consent_required. Paid accounts using model IDs without :contribute do not need to opt in.

Keys created from 31 October 2026 accept the current acknowledgement, which covers security research and model training. Keys that accepted the earlier research-only acknowledgement keep working: their excerpts are used for research and never to train models. To contribute to training from such an account, create a new key.

The collector preserves actual flagged input after best-effort redaction of detectable email addresses, phone numbers, credentials, credential-bearing URLs, and names. This is not a guarantee that every personal or proprietary detail is removed. Do not submit sensitive content to the free service if it cannot be retained after this redaction.

Each excerpt is capped at 2,000 characters, with at most eight excerpts per request. Where rule evidence identifies an attack, the excerpt includes nearby context. For model-only flags, the excerpt uses the beginning of the redacted input and may not contain the text that caused the flag. Redaction runs before clipping and again at the database write boundary. The collector uses hashes of the redacted excerpts to deduplicate contributions within the account.

Excerpts enter a separate research quarantine, associated with the contributing account and key. They expire at the UTC 12-month anniversary of collection and are removed by the next hourly cleanup. Each excerpt records which acknowledgement its key accepted. Excerpts collected under the current acknowledgement may be used to train Shield models after review; a classifier flag alone does not make an excerpt training data. Excerpts collected under the earlier acknowledgement are never used for training. Models trained on excerpts may be kept after the excerpts expire.

On this page