Low-cost inference · EU-owned hardware
Cheap usually costs you something else: your data, your region, or an unlabeled quantization. Our LLM API is priced by a public rule against the market — and the EU residency, zero retention and named hardware come with it.
The claim
For Qwen3.8 27B, we checked every provider listed for the same model on 2026-08-26: our €0.29/€1.77 per 1M is below every one of them.
One model, one date, checkable — and a standing rule that keeps it true: 80% of the lowest market price, tracked continuously against the market.
The mechanism
We price inference aggressively low: our target is 80% of the lowest current price from a proper inference supplier for the same model, tracked continuously. Found better pricing at a proper inference supplier? Tell us and we’ll look into lowering ours.
A proper inference supplier is a provider operating and standing behind managed inference for the model — decentralized compute marketplaces don't count. Because the target tracks the market instead of a promotional calendar, the low price is a standing rule, not a launch stunt. The full price list is at /pricing/.
Prices
| Capability | Model | Price |
|---|---|---|
| Chat — input tokens | Qwen3.8 27B | €0.29 / 1M tokens |
| Chat — output tokens | Qwen3.8 27B | €1.77 / 1M tokens |
| Embeddings | Qwen3 Embedding | €0.015 / 1M tokens |
| Image generation | ERNIE Image Turbo | €0.02 / image |
| Image editing · speech-to-text · text-to-speech · music | See the price list | On request |
EUR, exclusive of VAT, pay-as-you-go — no subscription, no minimum. Every API response reports usage, so the bill is reconcilable to the token.
No trade-off
Cheap and sovereign at the same time is the whole point. The low price does not buy down any of this:
| EU residency | Served from Stockholm (eu-se-1) and Málaga (eu-es-1), TLS terminated in the EU, region pinned on your key and echoed on every response. |
|---|---|
| Zero prompt retention | Prompts and completions are processed in memory — not written to disk, not logged, not retained, never used to train. See the privacy policy. |
| Named hardware | Qwen3.8 27B runs on NVIDIA DGX Spark (GB10, 128 GB unified memory) systems AxForge owns and operates — not anonymous capacity resold from somewhere. |
| Full API surface | OpenAI-compatible: streaming, tool calls and usage reporting all included at the listed price. |
Buyer's guide
Per-token headline prices hide most of the story. Whoever you compare — including us — check these four things:
| Check | Why it matters | Where we stand |
|---|---|---|
| Input vs output price | Output tokens usually cost several times more than input. A cheap input rate means little if your workload is generation-heavy — weight both rates by your real traffic mix. | €0.29 in / €1.77 out per 1M, both published |
| Context window | A low price on a small context can cost more overall once you're chunking and re-sending. Compare the context you'll actually use. | 65,536 tokens |
| Quantization | The same model name can be served at very different precisions, and quality follows. A provider should tell you what is actually running. | Disclosed on every model page — Qwen3.8 27B is served as NVFP4 |
| Data terms | A low price funded by retaining or training on your prompts is not a low price. Read the retention policy, not the pricing page. | Zero retention, no training — policy |
FAQ
We make one precise claim: for Qwen3.8 27B, we checked every provider listed for the same model on 2026-08-26, and our €0.29/€1.77 per 1M tokens was below every one of them — and the 80% rule keeps it that way.
We own the hardware the models run on, we serve efficient open-weight models on machines sized for the job, and the price reflects real serving cost plus a thin margin — no cloud markup, no VC-subsidized pricing that has to snap back later. The mechanics are on the pricing page.
No. Prompts and completions are processed in memory in Sweden — not written to disk, not logged, not retained, and never used to train anything. Only request metadata (token counts, timestamps, status) is kept for billing and operations. See the privacy policy.
Prices follow our pricing rule rather than a promotional calendar: the target is 80% of the lowest current price from a proper inference supplier for the same model, tracked continuously. They can move in both directions with the market — the rule is what keeps them low, not a launch discount.
API keys are allocated from the waiting
list — there is no public free tier today. Once you have a key, the API is
OpenAI-compatible: point your existing SDK at
https://api.axforge.ai/v1 and pay per token. Setup details are on
the Qwen3.8 27B page.
Related