ship with open-source ai: one api, the best open models of 2026

Two years ago, "we use an open model" meant "we accepted a quality tradeoff to save money." In 2026, that sentence is out of date. Qwen3.8, MiniMax M3, GLM-5.2, DeepSeek V4-Flash and Gemma 4 match or beat closed models on most production tasks — coding, extraction, RAG, agents, multilingual chat — while costing a fraction per token and being fully inspectable. The interesting question is no longer whether to build on open weights, but where to run them.

ElephantPool's answer: one endpoint, every model, served by a verified pool of GPUs across Europe at roughly 2–3× less than the big US APIs.

if you've used the openai sdk, you already know our api

We didn't invent a new client library, and you shouldn't have to learn one. The gateway is OpenAI-compatible: point your existing SDK at our base URL and change the model name.

from openai import OpenAI

client = OpenAI(
    base_url="https://app.elephantpool.ai/api/gateway/v1",
    api_key="epk_…",  # from your dashboard
)

resp = client.chat.completions.create(
    model="qwen3.8",
    messages=[{"role": "user", "content": "Summarize this contract in 3 bullets."}],
)
print(resp.choices[0].message.content)

That's the whole migration. LangChain, LlamaIndex, Vercel AI SDK, litellm — anything that speaks the OpenAI protocol works unchanged. Swapping between Qwen3.8 and DeepSeek V4-Flash is a one-string change, which makes A/B-testing models across your eval suite almost embarrassingly easy.

streaming that behaves

Pass stream=True and you get standard server-sent events, token by token, exactly as your existing streaming code expects — same chunk format, same [DONE] sentinel. Time-to-first-token matters more than raw throughput for anything user-facing, and the scheduler routes interactive requests to nodes that keep it low; batch-friendly jobs go to high-throughput nodes instead.

prepaid credits, no surprise invoices

Cloud bills have a horror-story genre of their own: a retry loop runs all weekend and Monday brings a four-digit invoice. We removed the genre. ElephantPool is prepaid: you buy credits, usage draws them down, and when they're gone your requests stop — they don't silently start billing a card.

• Spend caps per key and per project: cap a staging key at €10/month and let it die there.
• Live usage in the dashboard, per key, per model, per day.
• No seats, no platform fee, no minimum. You pay for tokens; zero tokens costs zero.

pricing you can put in a spreadsheet

Every model has one public per-token price — input and output, per million tokens, on the pricing page. No "contact sales" tiers, no regional multipliers, no discount theater. The distributed pool is what makes the economics work: 80% of what you pay goes to the machine that served your tokens, 20% runs the network, and there is no datacenter margin stacked on top. Same pricing across the pool; the EU-Sovereign tier, for regulated workloads, is priced separately and still undercuts sovereign clouds.

what this is good for, concretely

Teams on the network today are running customer-support agents on MiniMax M3, code review on GLM-5.2, high-volume extraction and classification on DeepSeek V4-Flash (its speed-to-price ratio is absurd for pipeline work), and multilingual products on Qwen3.8 and Gemma 4. The pattern is the same everywhere: start with your current eval set, swap the base URL, and let the numbers decide.

Open weights mean nobody can deprecate your model out from under you, repriced mid-quarter or gated behind a new tier. The model your evals blessed is the model you keep.