Service Tier

service_tier is an optional request field accepted by the three AI generation endpoints. It selects the provider’s priority or flex processing tier and is the literal field name on three of seven upstream provider APIs (OpenAI, Anthropic, Gemini API), making PickLayer drop-in compatible with standard clients.

Per-endpoint enum semantics

The two API shapes share the field name but use partially overlapping enums with one direct collision:

Value	`/v1/chat/completions` and `/v1/responses` (OpenAI shape)	`/v1/messages` (Anthropic shape)
`'priority'`	FAST	(not in enum)
`'auto'`	DEFAULT (equivalent to omitting)	FAST (uses Priority capacity)
`'flex'`	CHEAP (best-effort, lower cost)	(not in enum)
`'standard_only'`	(not in enum)	DEFAULT

'auto' collision warning: service_tier: 'auto' means standard/default on /v1/chat/completions and /v1/responses but means fast/priority on /v1/messages. The value is bound to the endpoint shape you are using.

Per-vendor wire translation

When a request routes to a vendor, the gateway translates the resolved tier into that vendor’s native mechanism:

Provider	`priority`	`standard` (default)	`flex`
OpenAI	`service_tier: 'priority'`	`service_tier: 'auto'`	`service_tier: 'flex'`
Azure OpenAI	`service_tier: 'priority'`	`service_tier: 'auto'`	`service_tier: 'flex'`
Anthropic	`service_tier: 'auto'`	`service_tier: 'standard_only'`	`service_tier: 'standard_only'` (no flex tier)
Bedrock	`serviceTier: { type: 'priority' }` on Converse/ConverseStream	`serviceTier: { type: 'default' }`	`serviceTier: { type: 'flex' }`
OpenRouter	`provider.sort: 'latency'`	omitted (load-balanced)	`provider.sort: 'price'`
Gemini API	`service_tier: 'priority'`	omitted	omitted
xAI	No-op (server logs info — use a model variant suffix like `grok-4-1-fast` instead)	No-op	No-op

Endpoints

service_tier is accepted on:

POST /v1/chat/completions — see Chat Completions
POST /v1/responses — see Responses API
POST /v1/messages — see Messages API

Cost implications

Priority tiers are typically priced higher than standard:

OpenAI Priority: ~25–100% premium (Enterprise-gated, model-dependent)
Anthropic Priority: pulls from committed capacity; falls back to standard automatically
Bedrock Priority: ~25% better throughput latency at a premium
OpenRouter latency-sort: routes to faster but potentially more expensive endpoints (1.5–4× floor price)

The flex tier (service_tier: 'flex' on the OpenAI shape) trades latency for lower cost — useful for background or batch workloads.

Example


{
  "model": "openai/gpt-4o",
  "messages": [{"role": "user", "content": "..."}],
  "service_tier": "priority"
}


{
  "model": "anthropic/claude-sonnet-4-5-20250929",
  "max_tokens": 1024,
  "messages": [{"role": "user", "content": "..."}],
  "service_tier": "auto"
}

Note: in the second example (/v1/messages, Anthropic shape), service_tier: 'auto' selects the fast/priority tier — not the standard default.