BETAService is in beta test. Use with caution.

Cut your token bills by 67%

Top open-source models through one OpenAI-compatible API — same intelligence, 3.1× cheaper. Change one line, keep your code, pay only for what you use.

Made for agents and high-volume LLM usage — the more you run, the more you save.

Eterial cost
$0.0000154
OpenRouter cost
$0.0000486
Saved / 1M requests
$33
Connect nowView docs
Scroll To Learn About Eterial

How it works

  • Decentralized host network

    Your requests run across a network of independent compute hosts — not one cloud's markup. That's what makes inference this cheap.

  • Backup failover

    If the network is ever unavailable, we switch to Tier-1 providers automatically. Uptime stays intact. Same performance. 3.1× less on tokens.

  • Frontier open models

    GLM 5.2 (1m context) and Kimi K2.6 match OpenAI and Anthropic on quality for everyday tasks. You cut the bill, not the output.

Build like tokens are free

The coding agent that reads your entire codebase.
The research agent that reads 200 sources before it answers.
The sales agent that writes a personalized email for every lead.

  • Orchestrate without counting

    Chain, verify, retry — the full stack, meter off.

  • Run agents at full depth

    Let loops run to completion. Depth costs pennies now.

  • Keep the smart model on

    Frontier GLM and Kimi, now the cheap option. Downgrading's over.

  • Ship what you shelved

    Too expensive to run at scale? Build it now.

  • Outgrow your spend caps

    The caps you set to survive frontier prices stop getting in the way.

  • Orchestrate without counting

    Chain, verify, retry — the full stack, meter off.

  • Run agents at full depth

    Let loops run to completion. Depth costs pennies now.

  • Keep the smart model on

    Frontier GLM and Kimi, now the cheap option. Downgrading's over.

  • Ship what you shelved

    Too expensive to run at scale? Build it now.

  • Outgrow your spend caps

    The caps you set to survive frontier prices stop getting in the way.

Smarter than Gemini. Almost free.

Top-tier intelligence 3.1× cheaper.
See the source.

Artificial Analysis Intelligence Index
60
Claude Fable 5
56
Claude Opus 4.3
55
GPT-5.5 xhigh
53
Claude Sonnet 5
51
GLM-5.2 max Coming soon
50
Gemini 3.5 Flash
46
Gemini 3.1 Pro
46
Qwen 3.7 Max
44
MiniMax-M3
44
DeepSeek-V4 Pro
44
Kimi K2.6
43
Muse Spark
42
MiMo-V2.5-Pro
40
DeepSeek-V4 Flash
40
GPT-5.4 mini
38
Nemotron 3 Ultra
38
Grok 4.3
34
Qwen 3.5 397B
30
Mistral Medium 3.5
30
Claude 4.5 Haiku
29
Gemma 4 31B
24
gpt-oss 120b
22
Nova 2.0 Pro
17
K2 Think V2
15
gpt-oss 20b
14
Solar Pro 3
Claude Fable 5
60
Claude Opus 4.3
56
GPT-5.5 xhigh
55
Claude Sonnet 5
53
GLM-5.2 max Coming soon
51
Gemini 3.5 Flash
50
Gemini 3.1 Pro
46
Qwen 3.7 Max
46
MiniMax-M3
44
DeepSeek-V4 Pro
44

Get started. Set up in minutes.

One balance. Chat and API.
The whole switch is one line.

Select Amount
$5$10$25$50
Amount$25
Min. deposit $10.00
Add $25.00 Credit

Add credit

Top up once. You’re paying for tokens you use, nothing else. A registered EU company — proper invoices and VAT receipts for your accounting.

Chat with frontier open source models

GLM-5.2, Kimi 2.6 in one interface. That’s Eterial Chat.

API Secret Key
et-sk-8f2jK9LmPq3sT5vW4xY2zCopy
Base URL Endpoint
https://chat.eterial.ai/v1
Quick Python Integration
import openai

client = openai.OpenAI(
    api_key="et-sk-8f2jK9LmPq3sT5vW4xY2z",
    base_url="https://chat.eterial.ai/v1"
)

response = client.chat.completions.create(
    model="minimax-m2.7",
    messages=[{"role": "user", "content": "Hello Eterial"}]
)

Turn on the API

Your key and OpenAI-compatible endpoint sit right in settings.

Swap one base_url

Point your existing code at eterial. Same SDK, 3.1× cheaper.

Pay as you go. No subscriptions.

Our pricing is dynamic and may
vary based on a number of factors. See details.

eterial.aiDeepSeek-V4 FlashOpenRouterDeepSeek-V4 FlashOpenAIGPT-5.5
Price / 1M input$0.04739$0.1477$5.00
Price / 1M output$0.09354$0.2954$30.00
Decentralized GPU network
Automatic failover
OpenAI-compatible

Our figure is an effective rate — the GPU network and its fallback blended. Run the same model on OpenRouter alone and you pay 3.1× more per token.

Still have questions?

Check the docs.

How is inference here so much cheaper?

Because the compute isn’t ours to mark up. Requests are served by Gonka, a decentralised network where independent operators compete to supply GPU capacity — so a token costs a fraction of what a conventional provider charges for the same open-weight model. On top of that, the network charges one rate for input and output. Most gateways charge several times more for the tokens a model generates than for the ones you send it.

So where do you make money?

Once, when you add credit — not on every request. From then on your balance is drawn down at the real price of the inference you spend: no per-request commission, no subscription, no minimum.

A decentralised network sounds fragile. What happens when it drops my request?

It does drop requests — between 10% and 35% of them, depending on conditions. We’d rather say that than let you discover it. Every one of those failures is retried against the equivalent model on a major provider inside the same request, before anything reaches you. You don’t write the fallback, watch the network’s health, or keep a second provider wired up for the days it’s struggling. A request the network drops is not a request that failed.

Do you support tool calling, web search, image and file input?

Yes — all four. This is where a decentralised network normally stops: on the network alone these are either missing outright or served by only some of the machines. Our API is not just the network. A request that needs a capability is routed to a backend that has it — so function calling and tool_choice, web search, image parts and PDF input all work in the same OpenAI-shaped request body you already send.

Are the prices on this page what I’ll actually pay?

They’re the blended figure, not the best case. Tokens the network serves are charged at the network’s rate; tokens that fell back are charged at the fallback’s. The number in the table weights the two by the share of traffic each backend actually carried over the last seven days — measured across every request our users sent, not modelled. Your own mix won’t be exactly the average, so treat it as a number to plan a budget around rather than a quote.

What does a failed request cost me?

Nothing, in almost every case. Each attempt reserves an estimate against your balance and releases it the moment that attempt fails, so a request that tried two backends before one worked is charged once — for the one that worked. A request nothing could serve costs nothing at all. The single exception is a stream that broke partway through: those tokens were generated and delivered to you, so those are billed.

Create your accountand top up in minutes

Start routing for free.
No credit card required.

Start trial session for free