Decide by the numbers

See every number
before you sign up

Every LLM gateway comparison page says "fast" and "hundreds of models." Almost none let you check, before signing up, what the TTFT is right now, exactly how many models there are, and how much has actually been paid out to creators. teai.io moved all of that to a public API with no authentication. This post shows what one curl tells you — and, based on those numbers, what is and isn't a reason to use us yet.

Context for readers outside Japan: teai.io is a Tokyo-based LLM gateway (OpenAI- and Anthropic-Messages-compatible) with JPY billing. Everything below is reproducible with curl; no account needed.

Published 2026-09-10

Where to check /api/v1/stats/public ・ /api/v1/latency

Refresh every 60 seconds, no auth required

394

models, 11 providers, one API key

420ms

fastest TTFT (DeepSeek v4 flash, measured)

MCP tools (voice, image, legal RAG…) on the same balance

80%

to you when your tool or voice is used

1. Check it yourself in 30 seconds

No account, no key. Run this and you get the current values.

curl -s https://teai.io/api/v1/stats/public | jq '{models, mcp, credit_rate}'
curl -s https://teai.io/api/v1/latency | jq '.measured[] | {model, ttft_median_ms, median_ms}'

What comes back is not the value when this post was written — it is the value when you ran it. Better or worse, it shows as is. Few gateways let you see this much before signing up.

2. What you can see (as of 2026-09-10 12:00 JST)

Visible	Current value	What it means for you
Model count	394 (11 providers)	Claude / GPT / Gemini / DeepSeek / Kimi / Qwen / Grok on the same code and the same invoice. This model today, another next month, no rewrite
Inference latency (TTFT/E2E)	14 models measured continuously. TTFT from 420ms (deepseek-v4-flash) to a 1.9s median	For chat UIs, time-to-first-token is what users feel. Pick models on numbers for your own use case
MCP tools	86 (voice-clone TTS, image/video generation, Japanese legal RAG, invoicing…)	Called with the same credits as your LLM balance. No per-tool contracts or key management
Credit-to-yen rate	¥1 = 6 credits	What a "credit" is worth is stated. Pro ¥4,350/mo = 26,100cr. Unit prices track public list rates at −2%
Creator payout ledger	22 paid calls, 197cr, 2 creators, all time	List an MCP tool or a voice and 80% of each payment is credited → withdraw from /voices. The ledger is public. The numbers are small for now (see below)

3. Why we published — because 10 personas said "not yet"

We recently benchmarked teai.io and KOE (voice cloning) against the field and asked 10 personas — a SaaS CTO, enterprise procurement, an MCP author, a VTuber, a VC, a US ML engineer, and more — to judge us harshly. 10 out of 10 said "not yet." The reason was essentially one thing: "fast" and "380+" were written down, but nothing could be verified before signing up.

The same day we shipped the public API, TTFT measurement, and a withdrawal UI to production, and removed "under 100ms" from the landing page (we couldn't trace the evidence). We switched to: look at the numbers, then decide.

4. Who has a reason to start today

Developers who want to switch between models — 394 models on one key, both OpenAI-compatible and Anthropic Messages-compatible. Append :fastest to a model name to auto-route to the fastest measured path
Companies that need JPY billing and qualified invoices — no more finance blocking on USD gateways. Tokyo region
People who choose by latency — pick models on live measurements, not claims. The method (prompt, sequential, empty = failure) ships in the same JSON
People building MCP tools or voices — list it and 80% of every payment is credited and withdrawable. The marketplace is nearly empty today. Whoever lists first is listed first
Anyone who wants a voice-driven CLI agent — sente. Your own cloned voice, 3-layer memory, local PII scrubbing. 100cr on signup, no card

5. Who does not have a reason yet (honestly)

Companies that require SOC2 — not certified (targeting 2027). DPA/APPI evidence is still being assembled. It will not pass procurement
Creators who want proof of earnings first — total payouts are ≈¥32 and external creators are zero. If you are waiting for traction, now is not it
People who decide on p95 — latency is one sample per model so far. Wait for it to accumulate (probes run every 30 minutes)
People happy with Cursor / Claude Code flat plans — credits are pay-as-you-go. If you want flat-rate peace of mind, staying put is simpler

We are not hiding that the numbers are small. 22 paid MCP calls, zero external creators, ≈¥32 paid out. Nobody would call that "busy." But we decided that growing it with the ledger public is more honest to the people reading than writing "80% revenue share" and keeping the ledger private. As it grows, the same URL shows it.

6. If you start, start in this order

Look at the numbers (no account) — the curl above or /api/v1/stats/public. If you don't like them, stop here. That's fine

Sign up free — 100 credits, no card. /register

Swap the base_url in existing code — https://api.teai.io/v1 for the OpenAI SDK, or the Messages-compatible endpoint for the Anthropic SDK. Change nothing but the model name

If you have a tool or a voice, list it — link voices at /voices, MCP via Docs. 80% of payments are credited to you

Start with the numbers

No auth. Sign up if you like them (100cr, no card). If not, that's fine too.

See the live numbers Start free

See every numberbefore you sign up

1. Check it yourself in 30 seconds

2. What you can see (as of 2026-09-10 12:00 JST)

3. Why we published — because 10 personas said "not yet"

4. Who has a reason to start today

5. Who does not have a reason yet (honestly)

6. If you start, start in this order

Start with the numbers

Related articles

See every number
before you sign up