FAQ

The most common questions we get. If yours isn't here, write to [email protected] and we'll add it.

Pricing & billing

What's the catch with $5 to start?

There isn't one. It's a refundable card hold, not a charge — we use it to keep account abuse down (no signup-for-free-credits farming) and as a floor for the first few hundred requests so the per-call ledger has somewhere to live. When you run out, you top up.

Why don't you charge per token like everyone else?

Because tokens don't track what your call actually costs us. A short request that triggers a long chain of internal reasoning eats more energy than a long simple completion. We measure the energy and bill that. See What is a joule, here.

How does my bill compare to OpenAI / Anthropic / together / Bedrock?

For mid-size L2-ish workloads we typically come out 30-60% cheaper than the published per-token rates because we route to the cheapest capable silicon rather than charging at the most-expensive-tier rate for everything. For tiny L0 calls, much cheaper still. For huge L3 reasoning calls, the savings are smaller (the expensive silicon dominates). Run a week in parallel and read both bills.

How are refunds handled?

Stripe / SEPA / ACH top-ups are refundable to source until the balance is consumed. USDC top-ups refund to the originating wallet on Base L2. Reward-point top-ups can't be returned to the program of origin (point programs don't permit that) but balance can be re-applied to any future workload.

Routing & models

Which models are available?

Open weights at v1: Llama 3.3 (70B, 405B), Mixtral 8x22B, Qwen 2.5 (72B), DeepSeek-V3, gpt-oss-120b. Hosted closed-weights where licensed: Claude Haiku 4.5. Run GET /v1/models for the live list and each model's typical joule cost.

Can I pin a specific model?

Yes — replace model: "auto" with the model name (e.g. "llama-3.3-70b-instruct"). The router still does region + load balancing within that model, but doesn't substitute.

Can I pin a region?

Yes — pass an X-Region header (eu-fi, eu-de, us-east, ap-northeast, etc.) or set region in invisible.hcl. Defaults to "cheapest capable" worldwide.

What happens when a node fails mid-request?

The gateway fails over to the next-ranked node automatically and your client gets a retried response. Failover events appear on the receipt as execution.retries.

Reliability & data

What's the SLA?

99.5% for v1; targeting 99.9% within a quarter of GA. SLA credits are issued automatically as joule balance when we miss. The status page is status.greenjoules.cloud.

Do you log my prompts and responses?

No. The middleware records operational metadata (joules, silicon, region, duration) but never payload. If you want a customer tag on the receipt for your own attribution, pass X-Customer-Tag; nothing else lands.

Where does my data physically live?

Inference requests stay in memory on the node that serves them — nothing is written to disk unless you write it. Object Store data sits at rest where you put it (jurisdiction is your pick at bucket creation). The portal shows you exactly which legal entity holds each byte.

Is this EU-data-residency-OK for CSRD / GDPR / AI Act?

Yes, when pinned to an EU region. Each request's receipt lists the operator's legal entity, country, PUE, and grid carbon — primary evidence for E1 Scope 2 cloud emissions and the energy-disclosure parts of the AI Act.

Migration & integration

How long does migration from OpenAI take?

Two lines of code (base_url + api_key). Most teams roll it in behind a feature flag, run both endpoints in parallel for a week, then retire the OpenAI client. See Migrate from OpenAI.

From Anthropic? AWS? GCP?

Anthropic and AWS S3 are drop-in; AWS-broader and GCP are by-service. See the guides.

Do you have an SDK?

For inference, use the OpenAI SDK in your language of choice — we're OpenAI-compatible. For workloads / Object Store / receipts, use the jc CLI (installed via curl -fsSL https://invisible.openie.sh/install.sh | sh) or the REST endpoints directly.

Company

Who's behind this?

openIE — the team behind the Joule programming language, JouleDB, and pattern-lang. We're building the energy-priced cloud because we needed one for our own work and couldn't find one. openie.sh.

Is this open source?

The infrastructure substrate (Invisible) is source-available on Forgejo at git.openie.sh/openie/invisible-infrastructure. The signing/receipt verifier is OSI-licensed. The router and billing layer are proprietary.

How do I contact you?

Product / sales: [email protected]. Security: [email protected] (PGP key at /.well-known/security.txt). Founder: [email protected].