One endpoint in front of many models and providers. Call a model by name, or say auto and let the layer read the request and choose. Your integration stays still while the model world keeps moving.
A new frontier model ships every few weeks, and the best tool for a job keeps changing. Hard-wiring one vendor into your stack means re-integrating every time the leaderboard flips. 0api sits in between: an interface that stays still for your code, and stays current underneath.
Not a menu of forty models to maintain, and not a black box either. Call a model directly, or hand the choice to auto. Either way: one endpoint, one key, one bill.
An OpenAI-compatible endpoint in front of multiple providers and model families. Pin kimi-k3 by name, switch models by changing one string — never an SDK.
Send model=auto and the router sizes the job — task intent, context length, in five languages — and picks a lane on the way through. Explicit names always bypass it. Both live behind the same key.
Every routed response carries three headers you can audit: the tier the router chose, the model that actually served the request, and whether reasoning context was restored. Not a marketing claim — response headers. X-Manifest-Tier · X-TokenHub-Served-Model · X-TokenHub-Reasoning-Restore
Kimi, GLM and DeepSeek expect their reasoning context back on multi-turn tool calls — most clients drop it. 0api restores it at the proxy, so agent sessions stay coherent across clients. Exact session match; on a miss it passes the request through safely, and never invents hidden reasoning.
Smart routing is the first capability shipped on this layer — live today, not a roadmap slide. Grey dots are incoming requests. The moment one lights up, the router has already sized the job and picked a lane.
Most routing layers read English and translate everything else — intent gets lost on the way. 0api classifies task intent natively in five languages, live in production today. We haven't found another routing layer that ships this.
Five scripts, five workloads, five different lanes — intent read natively, no translation hop. Today this is a high-confidence rule baseline verified on real public traffic; full natural-language understanding is on the roadmap below.
Translations, formatting, quick answers. Instant, and nearly free.
Everyday coding, docs, summaries. The workhorse lane.
Large refactors, multi-file generation, long context. Code flagships.
Gnarly debugging, math, architecture. Slow thinking, on purpose.
Every text tier runs on a 1M-token context window. Image and multimodal requests share one dedicated media lane, audio gets its own — same endpoint, same key.
The endpoint you integrate today is the first layer of an adaptive system. Here is the honest state of each part — what's live, what's being built, and what's a direction.
One OpenAI-compatible endpoint, direct model calls, auto routing, response headers you can audit, reasoning continuity for multi-turn agents. Everything above this line is what you saw on this page — in production today.
English, Japanese, Korean, Thai, Chinese. The router reads task intent natively in each — "explain a concept" versus "prove, compare, review" — verified on real public traffic. Today this is a high-confidence rule baseline, not full natural-language understanding; that's the next line.
Every routed request produces a signal: what the router predicted, what actually happened. The plan is a classifier trained on that real usage, replacing hand-written rules where the data says it's better.
Usage signals feeding model adaptation, so the layer doesn't just pick models — it improves them for the work you actually send. This is the direction the system is designed toward, not a shipping date.
Pay as you go: auto is $1.40 per 1M input tokens — $0.42 cached — and $4.40 out. Subscriptions add a monthly auto credit at those list prices. One rule to know: credit covers auto only — pinned models always bill your wallet.
Beta pricing — capped at 50 subscribers, personally onboarded. Windows, rollover, first-month bonus and fair use: see full pricing →
curl https://api.0api.dev/v1/chat/completions \
-H "Authorization: Bearer $ZAPI_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"auto","messages":[{"role":"user","content":"hey"}]}'
# or pin one by name: "model":"kimi-k3"from openai import OpenAI
client = OpenAI(base_url="https://api.0api.dev/v1", api_key=ZAPI_KEY)
resp = client.chat.completions.create(model="auto",
messages=[{"role": "user", "content": "hey"}]) # or model="kimi-k3"
# which model served it: X-TokenHub-Served-Model response header# ~/.claude/settings.json
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.0api.dev",
"ANTHROPIC_AUTH_TOKEN": "your 0api key",
"ANTHROPIC_MODEL": "auto"
}
}One key, one endpoint — and a layer behind it that keeps up with the models so you don't have to.
Request beta access