321AI Model MarketplaceLive marketplace →

Token groups explained: default vs stable vs guaranteed

Platform mechanics · 2026-11-12 · 321AI engineering notes · get a key

Quick answer: A token group is a per-key routing tier: same account, same code, but the gateway optimizes each key for a different goal. 321AI ships three groups — default for lowest price, stable for highest stability, guaranteed for merchant-supplied lanes. Choosing per workload is a one-click cost and reliability lever.

Sign up & get trial creditSee live pricing

What a token group changes — and what it does not

It changes one thing: how the routing engine prefers lanes for requests made with that key. It does not change your models, your code, your base_url or your prices’ source — the market floor mechanics stay the same. You create separate keys per group in the console and point different workloads at them; moving a workload between groups is a key swap, not a migration.

A useful mental model: the group is a routing policy attached to a credential. Requests themselves carry no group information, so nothing in your request body, headers or SDK changes — the console binding is the whole mechanism. That is also why audits are easy: a key’s group is visible next to the key, and every log line attributes its traffic to the key that made it.

default: the price-first lane

The default group optimizes for the lowest price among healthy lanes. It is the right choice wherever volume is high, latency tolerance is generous, and an occasional retry is acceptable: batch evaluation, offline processing, cron jobs, data labeling, anything you would happily re-run. Combined with the market floor and merchant-set prices, this is where the cheapest possible bill lives.

One caution: cheap is not always cheap. A price-first lane that fails on 3% of calls inside a ten-step agent loop multiplies retries across every step, and while failures refund, the latency and complexity cost is real. If a default-lane workload starts showing retry patterns in the logs, that flow has outgrown the price-first lane — move it to stable and stop paying in debugging time.

stable: the success-rate-first lane

The stable group prioritizes measured stability over price. It is built for traffic where a failed call is user-visible: production chat, checkout-path features, agent loops whose failure means a broken flow. The premium over the default lane is the cost of not explaining errors to your users — and since upstream self-check records per supplier are published on the marketplace, the trade-off is visible data, not vibes.

guaranteed: merchant-supplied lanes

The guaranteed group runs on merchant-supplied lanes: every listing names its merchant and channel, and merchants publish their own upstream self-check records. This is the lane for workloads where supply transparency matters commercially — check the named merchant and its records before routing production traffic here.

Matching groups to workloads

WorkloadGroupWhy
Batch eval / offline processingdefaultPrice-first; retries are free (failures never billed)
User-facing production chatstableSuccess-rate-first; a failed call is user-visible
Commercial / regulated deliverablesguaranteedNamed merchant + upstream self-check records
Prototypes and experimentsdefaultCheapest possible exploration with trial credit

Switching costs nothing

Groups are selected per token in the console; no code changes, no re-deploys. A common pattern is three keys — one per group — with each workload pointed at its key, and conditional routes still doing live lane selection inside whatever preference the group sets. The FAQ has the short version, and the how-it-works page documents the merchant-supplied lanes behind the guaranteed group.

A rollout plan for teams

A clean rollout: start everything on default for a week and collect logs; move user-facing keys to stable and watch the success-rate delta; put authenticity-critical flows on guaranteed and note which suppliers serve them. Revisit quarterly — workloads drift, and a batch job that became a product feature should not still be riding a price-first lane. The whole migration is key management in the console; nothing ships. The setup guides cover per-key configuration in the usual tools.

FAQ

Do token groups change my code?

No. Groups are a property of the API key, selected in the console. Same base_url, same request shape — the routing preference is the only thing that changes.

Is the guaranteed group slower or more expensive?

It runs on merchant-supplied lanes, which is a supply-transparency choice rather than a speed one. Live upstream self-check records for every lane are public on the marketplace, so judge with those numbers.

Can I use different groups for different services?

Yes — that is the intended pattern. Create one key per group and point each service at its key; production on stable, batch on default, authenticity-critical flows on guaranteed.

What happens if a guaranteed lane serves the wrong model?

Every listing names its merchant and channel, and merchants publish their own upstream self-check records; open a support ticket with the request id and we route it to that merchant.

Do groups affect which models I can call?

No. Every model stays callable from every group — the group changes lane preference within the model you asked for, not the catalog.

Keep reading

What is an AI API aggregation platform? How multi-supplier bidding keeps model prices low 9 checks before you top up an AI API relay Claude API cost guide: the per-1M-token math GPT-5.x family pricing guide for developers Gemini 3.x pricing and long-context cost math

Live prices: /en/models/ · billing: /en/pricing · setup: /en/guides/ · FAQ

All model pricesFeatured · 321AI communityFeatured · Open-source toolsFeatured · Advertise with usFeatured