Frequently asked questions

Model usage is billed at the provider's list price, passed through untouched. Our only fee is 4.9% (minimum $0.50) when you top up credits. There are no seats on Free and Pro, no per-request fees, and no markup hidden inside token prices.

Never. Top up $50 today and it's worth $50 in five years. Team seats ($19/seat/month) and enterprise agreements are the only recurring charges on the platform.

Every response carries an x-sg-cost header with the precise amount, computed from the provider's list price and the real token counts. The ledger stores the same number, so the sum of your traces equals your bill to the cent.

Exactly 0 credits. If the gateway answered from cache, no provider was called and nothing is billed — not a discount, zero. Hits show on the ledger as $0.000 with a cache: hit tag.

A velocity breaker in the gateway's hot path. It watches request rate, spend rate and repetition per key; when thresholds are blown, further calls get a clean 429 breaker_tripped and the key owner is notified. It's on by default with conservative limits you can tune.

Keys are assigned to named users. Each user (or team, or environment) gets a spend cap enforced before tokens are bought — when the budget is gone, calls stop, rather than a report telling you afterwards.

Keys created with a fixed credit balance and optional expiry. Perfect for contractors, classrooms, hackathons and demos: the key is the budget, so no invoice surprise is possible. Unused balance returns to your org if you burn the key early.

No, and it never will. Pure routing is the product: prompts and completions transit to the provider you chose; the gateway keeps metadata (model, tokens, latency, cost) for the ledger and nothing else. Details on the security page.

The catalog covers Anthropic, OpenAI, Google, Mistral, Groq, AWS Bedrock, Meta, DeepSeek, xAI, Alibaba and Cohere — 25 models today, each with its own page listing price and context — plus your own models via the device agent.

A meta-model that routes each request to a strong current default, with French-aware routing built in: francophone traffic is detected sub-millisecond and sent to models that score well on our French rubric. Explicit model IDs are never rerouted.

Configurable cross-provider retry: the gateway fails over to sibling models under one policy, and the status page reports the resulting effective success rate — the number you actually experience.

Not an AI — a small piece of software (binary or container) you install on your own server next to vLLM, Ollama or any OpenAI-compatible runtime. It opens an outbound mTLS tunnel to the gateway and your model appears in your catalog as local/, callable through the same API, billed at 0 credits.

No. Inference happens inside your network; the tunnel is outbound-only with no inbound ports, and only routing metadata reaches the control plane.

That's on the roadmap. When the marketplace opens, any tunneled model can be published with per-token prices you set; the gateway meters buyers, bills their credits and pays you out. Tunnels set up today will be marketplace-ready.

Change base_url to https://api.switchgate.ai/v1 and swap the key. The API is OpenAI-compatible, so existing SDK code — Python, JavaScript, Node.js, Go, cURL — keeps working; model IDs use the provider/model format.

Yes: trial credits, the full catalog, and every governance feature. Pro is pay-as-you-go with no subscription; Team adds shared governance at $19/seat; Enterprise adds SSO, DPAs, private models and an SLA.

The control plane and ledger run in Montréal, Canada, designed for Canadian privacy law and the GDPR, with additional residency options on enterprise agreements.