> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sailresearch.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Service tiers

> Trade off latency for lower token prices

Service tiers allow you to balance processing speed, latency, and cost for inference requests.

Sail serves low-latency inference by default for [core models](https://docs.sailresearch.com/models.md), at lower cost than traditional inference providers. You can opt in to the Balanced or Flex service tiers to cut token costs drastically for latency-tolerant work.

## Service tier comparison

| Tier | Queueing time | Target TPS (Experienced TPS varies by model and workload.) | Price vs. competitor model APIs |
| - | - | - | - |
| Default (**ASAP**) | None | 50-100 TPS | 30-70% less |
| **Balanced** | Up to 5 min for first turn, much faster for subsequent turns (seconds) | 30-50 TPS | 45-80% less |
| **Flex** | Up to 1 hour for first turn, much faster for subsequent turns (seconds) | 10-20 TPS | 55-85% less |

Check [Pricing](https://docs.sailresearch.com/pricing.md) for current per-model availability.

## Service tier details

### Default (ASAP)

Default, aka ASAP, is the standard low-latency tier for Sail’s core models. It is characterized by:

1. **No queueing delay** - TTFT is dependent on prompt size and model.
2. **High TPS** - targeting 50-100 TPS. Your experienced TPS will vary by workload and model.
3. **Foreground-only** - this tier is not compatible with background mode or batch requests.

Developers use ASAP for:

* Chat interfaces
* Realtime interactivity

Jump to:

* [Default (ASAP) notes and troubleshooting](#asap-requests)

### Flex

Flex is characterized by:

1. **A flexible start** - the first turn in a run is guaranteed to start within 1 hour.
2. **Much faster subsequent turns** - follow-up turns of an in-flight run see minimal (seconds) queueing delay.
3. **Low TPS** - targeting 10-20 TPS. Your experienced TPS will vary by workload and model.
4. Support for **extreme scale by design** - Flex offers the most generous capacity and highest rate limits.
5. **Background-only** - Flex is not compatible with streaming or foreground APIs.

> [!NOTE]
>
> Sail uses longest-prefix matching to automatically recognize subsequent turns.

Developers use Flex for:

* Proactive long-horizon agents
* Background agents and tasks
* Rollouts
* Automations and routines
* Large-scale evals
* High parallelism work
* Offline processing

Jump to:

* [How to use Flex](#how-to-use-flex)
* [Flex notes and troubleshooting](#flex-requests)

### Balanced

Balanced is the middle ground between Flex and Default, characterized by:

1. **A flexible start with a tighter bound than Flex** - the first turn in a run should start within 5 minutes.
2. **Much faster subsequent turns (Sail uses longest-prefix matching to automatically recognize subsequent turns.)** - like Flex, follow-up turns of an in-flight run see minimal (seconds) queueing delay.
3. **Medium TPS** - targeting 30-50 TPS. Your experienced TPS will vary by workload and model.
4. **Streaming, foreground, and background-compatible.**

Developers use Balanced for:

* Subagents
* Frequent automations and routines
* RL rollouts
* Evals
* Coding agents

Jump to:

* [How to use Balanced](#how-to-use-service-tiers)
* [Balanced notes and troubleshooting](#balanced-requests)

> [card] Multi-turn agents and rollouts with Balanced and Flex
>
> Balanced and Flex are designed to support multi-turn agents and rollouts. For
> the first turn in a run, Balanced and Flex have a “warm-up” queueing delay of
> up to 5 minutes (for Balanced) and 1 hour (for Flex), respectively.
> Afterwards, Sail automatically uses longest-prefix matching to recognize
> subsequent turns of an in-flight run. These follow-up turns are prioritized
> and incur minimal to no queueing time.

## How to use service tiers

### Via the `service_tier` field

Set `service_tier` to `asap`, `balanced`, or `flex`. Reserved TPM
reservations use the same field and values.

```bash Responses theme={null}
curl https://api.sailresearch.com/v1/responses \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "input": "Explain the key ideas behind transformers.",
    "service_tier": "balanced"
  }'
```

```bash Chat theme={null}
curl https://api.sailresearch.com/v1/chat/completions \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "messages": [
      {"role": "user", "content": "Explain the key ideas behind transformers."}
    ],
    "service_tier": "balanced"
  }'
```

```bash Messages theme={null}
curl https://api.sailresearch.com/v1/messages \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Explain the key ideas behind transformers."}
    ],
    "service_tier": "balanced"
  }'
```

To use Flex, you must use the [Responses API](https://docs.sailresearch.com/support.md#responses-api)
with `background=True` or submit requests through the [Batch API](https://docs.sailresearch.com/requests_at_scale.md#batch-api).

> [!WARNING]
>
> Flex is **not** compatible with Chat Completions, Messages, or **foreground**
> (`background=False`) Responses API requests.

```bash Responses highlight={7} theme={null}
curl https://api.sailresearch.com/v1/responses \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "input": "Explain the key ideas behind transformers.",
    "background": true,
    "service_tier": "flex"
  }'
```

```bash Batch theme={null}
# Batch always uses flex, so no service tier is specified.
curl https://api.sailresearch.com/v1/batches \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "endpoint": "/v1/responses",
    "requests": [
      {
        "custom_id": "request-1",
        "params": {
          "model": "zai-org/GLM-5.3",
          "input": "Explain the key ideas behind transformers."
        }
      }
    ]
  }'
```

### Via request header

Alternatively, if your client cannot set `service_tier` in the request body,
you can set the service tier using the `X-Sail-Completion-Window` request
header.

```bash Responses highlight={4} theme={null}
curl https://api.sailresearch.com/v1/responses \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Sail-Completion-Window: balanced" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "input": "Explain the key ideas behind transformers."
  }'
```

```bash Chat highlight={4} theme={null}
curl https://api.sailresearch.com/v1/chat/completions \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Sail-Completion-Window: balanced" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "messages": [
      {"role": "user", "content": "Explain the key ideas behind transformers."}
    ]
  }'
```

```bash Messages highlight={4} theme={null}
curl https://api.sailresearch.com/v1/messages \
  -H "Authorization: Bearer $SAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Sail-Completion-Window: balanced" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Explain the key ideas behind transformers."}
    ]
  }'
```

### Default behavior

For inference requests for [core models](https://docs.sailresearch.com/models.md), the default behavior (when no service tier is specified) is low-latency inference (`asap`).
The exceptions to this default behavior are:

* [Batch inference requests](https://docs.sailresearch.com/requests_at_scale.md#batch-api), which default to Flex and do not support other service tiers.
* [LoRA requests](https://docs.sailresearch.com/loras.md), which are Flex-only.
* Requests that use custom tools, which default to Balanced for core models because ASAP does not support them.
* Responses API requests with `background=True` use Balanced or Flex by default, since they do not support ASAP.

For [Flex-only models](https://docs.sailresearch.com/models.md#flex-only-models), the default (and only) service tier is Flex.

## Troubleshooting

### ASAP requests

* ASAP does not support custom tools (`tools` with `type="custom"`) on the
  Responses or Chat Completions APIs. Use Balanced, or omit the service
  tier to default to Balanced for core models.
* ASAP does not support Responses requests with `background=True` or the
  Batch API. Use Balanced or Flex for Responses requests with `background=True`.
* Setting `service_tier="asap"` with custom tools or
  `background=True` returns HTTP `400` with error code `unsupported_asap_request`.

### Balanced requests

* You can send Balanced requests synchronously or as [Responses requests with `background=True`](https://docs.sailresearch.com/support#responses-api).
* A synchronous streaming Balanced request opens its stream right away and keeps the
  connection alive while it waits. Output starts when the model begins generating.

### Flex requests

* Chat Completions and Messages reject Flex with HTTP `400`, with or without
  streaming. Use ASAP or Balanced on those endpoints.
* Responses requests using Flex reject an omitted `background` value, `background=None`, or
  `background=False` with HTTP `400`. Set `background=True` explicitly.

### Request settings

* If you set both `X-Sail-Completion-Window` and `service_tier`, the value in
  the request body takes precedence.
* Type checkers can reject Sail's tiers in `service_tier`, because the OpenAI
  and Anthropic SDKs type the field with their own values. The SDKs still send
  the value as written. When your type checker rejects it, silence only that
  line: add `// @ts-expect-error` on the line above `service_tier` in
  TypeScript, or end the line with `# type: ignore` in Python. OpenAI's SDKs
  already allow `flex`, so a `flex` request there needs neither.
* The model must support the tier you choose. Check [Pricing](https://docs.sailresearch.com/pricing.md) for
  availability.

### Batch and Background

* Responses requests with `background=True` cannot use `stream=True`.
  Retrieve results by polling or using [webhooks](https://docs.sailresearch.com/webhooks.md).
* A polling timeout does not cancel a Responses request with `background=True`.
  Keep polling the same response ID. See [Retrying requests](https://docs.sailresearch.com/retries.md) if you
  lose the connection before receiving an ID.
* The Batch API does not use `X-Sail-Completion-Window`.
* Batch always uses Flex and runs in the background. Setting a subrequest's
  service tier to `asap` or `balanced` returns HTTP `400`.
* Batch rejects models that do not support Flex.

### Capacity errors

* For ASAP requests, Sail may wait briefly for capacity. If none becomes
  available, it returns `503` with code `model_capacity_unavailable`
  without queueing the request.

* If the service is overloaded, Sail can return `503` with code
  `service_overloaded` before accepting a request in any service tier.

* If the response includes `Retry-After`, wait at least that many seconds.
  See [Retrying requests](https://docs.sailresearch.com/retries.md) before resubmitting after a timeout or another
  server error, since the request may already have started.
