API keys & routing

A key is a routing rule: it decides which vendors and models a request may reach and how the gateway chooses between pricing groups. Keys are created under API Keys in the console and look like sk-xxxxxxxx.

1. Three routing modes

Choose a routing mode when creating the key:

Mode	Reachable	Good for
By vendordefault	Every model of the chosen vendor (Anthropic = all Claude models); smart routing inside that vendor's groups	Single-vendor clients such as Claude Code / Codex; one key per team or project
By model	Only the ticked models, across vendors; still routed to the cheapest available group	Restricting downstream systems or contractors to specific models
All models and vendors	Everything on the platform	Personal use, multi-model clients like Cherry Studio / Chatbox

Vendor vs. model name

The model field says what to call; the key's routing mode says whether that model is allowed and which group serves it. Calling gpt-5.6-terra with an Anthropic-only key returns 403 — switch to an "all models" key or tick the model in "by model".

2. How smart routing picks a group

The same model often lives in several pricing groups (for example claude-opus-5-5 in both Claude Enterprise and the coding-plan group). For every request the gateway:

Sorts candidates by effective price
Candidates = groups available to your account that contain the model, ordered by discounted price ascending (ties by group order).
Tries the cheapest first
If that group has no healthy channel for the model it moves up to the next one.
Fails over automatically
Upstream errors are retried on another channel in the same group, then on the next group; you only see an error when everything failed.
Falls back to special groups last
Models that only exist in Unlimited / Unreviewed groups are reached only when no regular group has them.

You normally never pick a group by hand; each usage-log entry shows the group and unit price that actually served the call.

3. The group system

Groups are the unit of pricing and upstream pools. Badges on the Models page mean:

Group	Meaning	Examples
Enterprise	Official full-strength models, no client restrictions, long context, images and search; the main production tier	Claude / GPT / Gemini Enterprise, GLM / Kimi / DeepSeek / Qwen Enterprise
Economy	Discounted pools; GPT Economy is limited to Codex clients and cannot generate images, some pools do not guarantee cache hit rates	Claude / GPT / Gemini / GLM Economy
Edu	Promotional group excluded from automatic fallback; select explicitly	Claude Edu
Workflow	Official list price with the vendor's full catalogue (incl. legacy models); for non-coding toolchains and batch jobs	GPT / Claude / GLM / Gemini / Cohere / Mistral Workflow
Plan	Coding-plan subscriptions with their own concurrency and quota; falls back to the wallet when used up	Claude Code / Codex plans
Free	Free or promotional pools without availability guarantees; a positive balance is still required	DeepSeek Free, TypeSafe Free tier
Unreviewed	Third-party unreviewed image / video pools billed per image or per second	Seedream / Seedance / Qwen Image / Wan / MiniMax H3 / FLUX / LTX
Unlimited	Uncensored and abliterated models	GLM-5.3 / DeepSeek V4 Pro unlimited, Roleplay

Discounts, per-model prices and tiered pricing live in the model square, which is the source of truth.

4. Quota modes

Mode	Behaviour
Unlimited	Bounded only by the account balance (default)
Fixed total	The key stops (`429`) once its lifetime spend reaches the cap; raise it any time in the console
Periodic	Resets daily / weekly / monthly; the period is anchored to the moment you saved the setting and resets once per full period

Keys can also carry an expiry date and an IP allow-list; deleting a key revokes it immediately.

5. Concurrency and rate limits

Concurrency is limited per account tier: untiered (T0) accounts get 30 in-flight requests, T1 and above 300. Exceeding it returns 429 with an explanatory message.
Coding-plan groups use the plan's own concurrency pool and do not stack with the above.
Image generation has a separate concurrency gate so bursts cannot starve other traffic.
Genuine high-concurrency needs (batch jobs, evaluations) can be unlocked per account — contact support with the use case. Pointless load testing is treated as an attack.

6. Account tiers and discounts

Accounts are tiered automatically by cumulative spend (T0 → T1 → T2 → V1 → V2); higher tiers get lower prices in each group, as shown in the model square's tier matrix. Tiers apply to every key on the account — no need to recreate keys.

7. Recommendations

One key per purpose: Claude Code (by vendor: Anthropic), Codex (by vendor: OpenAI), chat client (all models) — isolated usage and spend.
Give downstream systems a by-model key with a fixed or periodic quota.
Keep keys server-side / in environment variables; proxy browser and mobile traffic through your own backend.
GET /v1/models lists exactly what the current key can reach — the fastest way to debug a 403.