Cerebras through Squiid

What Cerebras is

Cerebras builds wafer-scale processors and sells inference on them. The Cerebras Inference API serves open-weight models through an OpenAI-compatible interface, and its selling point is raw output speed: token rates high enough that a long answer finishes while a GPU-hosted copy of the same model is still on its first paragraph.

That matters for a narrow but growing set of features in AI-coded products. Agent loops that make many sequential model calls, voice interfaces where the reply has to start immediately, and interactive tools where the user watches the text appear all feel different at high token rates. Cerebras is typically added as a speed tier next to an existing frontier provider rather than as a replacement for it. The models offered are open-weight families rather than proprietary ones, so the same prompt usually runs unchanged against a slower host if you need a fallback. Speed also changes evaluation economics, because a test suite that takes an hour on a slow endpoint can run on every commit when the endpoint is fast enough.

Why connect it through Squiid

Cerebras is on the Squiid roadmap. Once it is connected, calls will go to https://api.squiid.io/v1/cerebras/... with Authorization: Bearer $SQUIID_API_KEY, OpenAI-compatible clients will work by pointing their base URL at the gateway (gateway convention), and tokens will be metered against the same prepaid credits at Cerebras' list price. Squiid will hold the upstream key and show per-model spend beside your other providers on one invoice. Adding a speed tier will not mean adding a second card, a second invoice or a second key to rotate.

• One account, one key, one bill, one dashboard. Cerebras sits beside every other service your project uses, on the same invoice and the same credit balance.
• Key custody. Squiid holds the Cerebras credentials. Your code carries SQUIID_API_KEY, so CEREBRAS_API_KEY, CEREBRAS_BASE_URL never land in a repository, a preview deployment or an agent’s context.
• Metering you can see. Every call is counted per service and per day, so you know what Cerebras costs before the month ends.
• Pause on zero. When prepaid credits run out the gateway stops authorizing usage instead of running up a bill you did not fund.
• One rotation. Rotating the Squiid key cuts off every agent, worker and deployment at once, with no provider key to hunt down.

What you get

• Open-weight models on wafer-scale hardware
• OpenAI-compatible endpoints through the gateway
• A latency tier next to your existing providers
• One key and one invoice for every model vendor
• Cerebras list pricing, no markup

How to connect Cerebras

One variable replaces the provider credentials this service would normally need. Requests go to the gateway and the official SDKs work by pointing their base URL at it (gateway convention).

# .env
# replaces CEREBRAS_API_KEY, CEREBRAS_BASE_URL
SQUIID_API_KEY=sq_live_xxxxxxxxxxxxxxxxxxxxxxxx

# every request: https://api.squiid.io/v1/cerebras/...
# header:       Authorization: Bearer $SQUIID_API_KEY

Pricing through Squiid

Cerebras charges per input and output token by model, at the provider's listed price. Squiid passes that through at par: one dollar of prepaid credit is one dollar of Cerebras usage, with no markup and no per-request margin.

Cerebras has no subscription attached on Squiid today. If a plan is required for a feature you need, it passes through at the provider’s listed price with no Squiid fee.

Squiid adds no markup to provider prices. There is a free plan at $0 a month with a 15% fee when you top up credits, Solo at $19.97 a month with a top-up fee that starts at 8% and falls to 3.5% as your spend grows, Team from $29.97 a seat a month with the same ladder one point lower, and Custom with a rate of 3% or less. Credits are prepaid, so Cerebras usage stops when the balance reaches zero rather than producing an invoice you did not fund. Auto top-up with a cap keeps production running without handing anyone an open-ended card.

If you leave

Nothing to move: Cerebras is usage-metered, so leaving means creating your own Cerebras key and pointing the client at it. Squiid never owns your code, your schema or your domain, and the gateway speaks Cerebras's own API, so the client change is a base URL. The handover page lists what happens for every provider in the catalogue.

Works with every coding agent

Cerebras through Squiid is just an HTTP call with a bearer token, so every agentic coding tool can use it with the single key already in your environment:

• Claude Code
• Cursor
• ChatGPT
• Codex
• Grok
• Gemini CLI
• GitHub Copilot
• Windsurf
• Perplexity
• Devin
• Replit
• v0
• Lovable
• Cline

Vibe coding a feature on Cerebras usually means asking for it in plain language and letting the agent write the integration. Because the key is Squiid’s rather than the provider’s, it can do that without ever holding a credential that could be leaked, logged or committed.