llm.priority asks the model provider to run this agent’s turns in its premium service tier. It is aimed at one number — the pause between the caller finishing a sentence and the agent starting to speak — and it is the only setting on an agent that changes what a token costs: 1.75x to 2.5x, for as long as it is on. It exists on config.llm and, separately, on config.llm.fallback. It does not exist on a realtime speech-to-speech model, and it is refused on the post-call analysis model.

What it buys

Every provider in the catalog schedules priority traffic ahead of standard traffic for a per-token premium, and all three take the request the same way — a service_tier parameter on the call Talqing makes for each turn. What that buys is narrower than it sounds: the model starts producing tokens sooner when the provider is busy. It does not make the model think faster, does not change the answer, and does nothing at all to speech-to-text, text-to-speech or the avatar. It steadies the worst case rather than speeding up the average. On a phone call that is the right target, and it is the case Talqing measures a call against: a call whose p95 time to first word crosses 3.5 seconds is flagged with a slow-response issue on its record, however good the average was. To whoever is on the other end, a long enough silence reads as a dropped call rather than as thinking. The flag exists on a text agent too — the same language models serve all three channels — but nobody is listening to silence in a chat thread, so the same premium buys far less there.
The lane is best-effort, and no provider guarantees it. Under load a turn can be served at standard speed anyway, and you still pay the premium for it.Measured on xAI’s grok-4.5 in August 2026, the lane moved the median time to first token from 1.57s to 1.51s — roughly nothing, because there was no queue to jump that day. That is what best-effort looks like when capacity is not scarce. Measure it on your own traffic before you leave it on.

The price matrix

USD per 1M tokens, from the provider catalog. These are the rates your provider bills your own key — Talqing charges only the per-minute platform fee, which the lane does not change. GET /v1/catalog is authoritative and carries the same priority block per model. The premium is not one multiplier the platform applies. It is transcribed per model from each vendor’s own premium price list, which is why OpenAI charges 2x across the 5.6 family, 2.5x on gpt-5.5 and 1.75x on the 4.1 pair. One rate breaks its own pattern: gemini-3.5-flash-lite prices cached input at 0.05ratherthanthe0.05 rather than the 0.054 an exact 1.8x would give, because $0.05 is what Google publishes.

Models with no lane

  • openai/gpt-5.4-nano is absent from OpenAI’s premium price list, so the lane is not sold for it. Asking for one is a save error, and the dashboard shows no control on that model.
  • Every realtime speech-to-speech model. Setting config.realtime clears stt, llm and tts together, so there is no llm.priority to set. See models.
  • Speech-to-text, text-to-speech and avatar models. No provider in the catalog sells a premium tier for those, and the field does not exist on those slots.

What the multiplier applies to

Only the language model’s own token spend, on the model the flag is set on. A voice call also pays for speech-to-text, text-to-speech, possibly an avatar, and the platform fee, and none of those move. Whether that makes the lane cheap or ruinous depends entirely on what share of your call the language model already is — read it off the cost breakdown of a real call before you decide. The premium is also not something Talqing bills you. Under bring-your-own-key the tokens go on your provider invoice; the only thing Talqing debits is the per-minute platform fee, which is identical either way. What Talqing does is meter the tokens and report the spend, at the priority rates above, so the number you see reflects what your provider will charge.

Talqing prices the tier you asked for, not the tier you got

The tier that actually served a turn is not on the wire by the time anything is metered, so a usage row is tagged priority because the agent requested the lane for that model. On OpenAI and xAI the response does echo service_tier back, so their own invoices settle the difference on their side.
Gemini never reports which tier served a turn. Google validates the service_tier parameter — a bad value fails the request outright — but sends nothing back on the response, while documenting an automatic downgrade to standard under load.So a Gemini turn that Google quietly served at standard speed is invisible here and is still reported at the 1.8x priority rate. If you want the premium you pay for to be verifiable, the lane on OpenAI and xAI models is; on Gemini it is not.

Where it is refused

All of these are save errors, checked on every write to the draft rather than waiting for a publish. Per-call configuration is checked by the same rules, so an agent_override that switches the lane on for a model without one is rejected when the call is requested. See validation. The analysis rule is the one people trip over. Post-call analysis reads a finished transcript after everyone has hung up, so a premium lane there would double that spend and shorten nobody’s wait; the model that does it is configured separately and always runs standard. See call analysis. Agent tasks carry the same llm.priority field and are checked by the same rules. Nobody waits on a task run in real time either, so it costs the same there and buys correspondingly less.

The fallback carries its own

llm.fallback.priority is a separate flag, validated against the fallback model’s own catalog entry. Nothing is copied down from the primary, and that is deliberate: support is not uniform even inside one vendor, so a failover that inherited the primary’s flag could ask for a lane its model does not sell. The practical version: set the lane on the model you expect to run, and leave it off on the one that only runs when the first has failed. A failover that is a little slower is a call that survived. The same rule applies to reasoning_effort and builtin_tools — see model fallbacks.

Priority or thinking

reasoning_effort and priority both act on the pause before the agent speaks, from opposite ends, and they are not interchangeable.
  • reasoning_effort decides how many tokens the model spends before its first visible one. Raising it always makes the pause longer.
  • priority decides how quickly the provider gets round to you. It does not reduce the thinking.
So the order matters. If first-token time is high on nearly every turn, that is thinking, and the lane will not fix it — lower the effort instead; unset already means the fastest value the model offers. If most turns are quick and a handful are slow, that is a queue, and the lane is the lever pointed at it. Raising the effort while the lane is on compounds: more tokens on the same turn, every one of them at the premium rate. Get the answer acceptable at the lowest effort the model will accept, then decide about the lane. See thinking and provider tools.

Deciding whether it is worth it

1

Baseline it with the lane off

Take a batch of real calls on the agent as it stands and read the LLM first token stage — on a call, in the health band above the transcript and per turn beside each agent message; across a date range, in the latency section of the observability dashboard. Note the average and, more importantly, how many turns sit far above it. See calls.
2

Turn it on and publish

The flag lands on the draft. Live traffic keeps running the published version until you publish again — see versions.
3

Run comparable traffic

The same agent, the same hours, enough calls that a handful of slow ones is a pattern rather than an accident. Provider queues vary by time of day, so a morning against an evening tells you about the hour, not the lane.
4

Compare the tail, then the bill

Expect the average to move a little and the slow turns to thin out. If neither happens, the lane bought nothing on your traffic and you are paying up to 2.5x for it.
Every LLM line that ran in the lane is tagged priority in the cost breakdown, on the call and in the workspace cost view, and the observability response totals it separately as priority_cost — so what the lane cost you over a range is one figure rather than a subtraction you have to do yourself. See costs. The agent editor’s per-minute cost estimate also prices the switch: it recalculates against the priority rates the moment you flip it, before you save anything.

Setting it

In the dashboard it is a Priority processing toggle in the Models card, directly under the model it applies to, with that exact model’s multiplier in its help text. A model with no lane shows no toggle at all. The fallback model has its own toggle inside the fallback block. Over the API, the whole config is written at once — whatever the config omits reverts to that field’s default — so read the agent, change the one field and write it back.
The fallback needs its own line — config.llm.fallback.priority — and only when the agent has a fallback model at all. To try the lane on a single call without touching the agent, set it in agent_override on the request that starts the call; see per-call configuration.

Next

Thinking and provider tools

The other lever on the pause before the agent speaks, and the one to reach for first.

Models

Every model, its flags and its standard price.

Costs

Reading provider spend against the platform fee, line by line.