Migrating from Hyperbolic
Hyperbolic retired its serverless inference API. If you built on it, your integration stopped working through no fault of yours. Here is a working home — one base-URL swap away.
Written 2026-10-05. Hyperbolic's status and all prices were checked that day; links go to the sources.
What changed
Hyperbolic's own FAQ says it plainly: “The serverless inference API and the playground have been retired… We are building a new inference product. We cannot share a release date yet.” (hyperbolic.ai FAQ, accessed 2026-10-05). Existing API keys no longer work for inference, and remaining credits apply to GPU rentals rather than tokens. They were good at this, and they may be again — but if you need tokens today, waiting on an unannounced date is not a plan.
The switch
Synap speaks the same protocol Hyperbolic did — the OpenAI API. Migration is a key and a base URL; your request and response handling code does not change.
1. Sign in with Google — new accounts start
with $2 of free credit, good for 14 days, no card.
2. Copy your key (sk-pool-…) from the
account page.
3. Point your SDK at us:
# before
base_url = "https://api.hyperbolic.xyz/v1"
# after
base_url = "https://pool.linkrra.com/v1"curl https://pool.linkrra.com/v1/chat/completions \
-H "Authorization: Bearer $LINKRRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-coder-30b",
"messages": [{"role": "user", "content": "Say hello"}]
}'Cline, Continue, Cursor, Aider, LiteLLM and anything else OpenAI-compatible work the same way — setup snippets for each are on Docs.
Where your models land
We serve open-weight models only, and ten of them: five we run on our own GPUs and the five most asked-for models we do not. If you were calling a Llama, Qwen, DeepSeek or GLM family model, the closest homes here are below — prices read live from /v1/models on 2026-10-14 against /v1/models, which is what the gateway bills.
| If you ran | Closest on Synap | In /1M | Out /1M |
|---|---|---|---|
| A small Llama / general chat | gpt-oss-20b | $0.017 | $0.085 |
| A Qwen coder model | qwen3-coder-30b | $0.0675 | $0.27 |
| A fast general or Gemma model | gemma-4-26b-a4b-it | $0.040 | $0.210 |
| DeepSeek, fast tier | deepseek/deepseek-v4-flash | $0.165 | $0.331 |
| DeepSeek, reasoning | deepseek/deepseek-v4-pro | $2.279 | $4.051 |
| A GLM model | z-ai/glm-5.3-flash | $0.380 | $0.712 |
| A frontier GLM | z-ai/glm-5.3 | $1.695 | $4.735 |
| Long-context work | moonshotai/kimi-k3 | $3.845 | $17.09 |
The full catalog, with vision and voice models, is on Models. If a model you depended on is missing, tell us: weare@linkrra.com — catalog requests from migrating users get read first.
What to know before you commit
The numbers are published. Our flagship coder runs on consumer RTX 5090s — that is how it costs $0.0675/$0.27 per 1M. The measured throughput is on the benchmark page, not hidden behind a product name.
Zero data retention, enforced per request. Every forwarded call
carries zdr: true; a model with no compliant endpoint fails the request
rather than quietly routing somewhere that keeps your data.
Prepaid, like you are used to. Pay-as-you-go credits that do not expire, or plans with carry-over — details on Pricing. Crypto works too (USDC on Base via x402, TON).
Two dollars of credit is enough to run your real prompts against the models above and decide with evidence. If it does not fit, you have lost an afternoon, not a migration.