CLF AI Gateway is live — Kimi & GLM, 40% off for launch

CLF AI Gateway is open to everyone today. One OpenAI-compatible endpoint in front of the Kimi and GLM model families, prepaid USD credits, and per-token billing you can audit line by line. No waitlist, no sales call — sign up, top up, change base_url, ship.

What you get

• Four models, public upstream IDs — kimi-k2.6, kimi-k2.7-code, glm-5.2 (262K context) and glm-4.7-flash (131K). We publish the exact @cf/... upstream IDs we run on Cloudflare Workers AI, so you always know what is answering.
• The OpenAI wire format, unchanged — official Python/Node SDKs, LangChain, and the Vercel AI SDK all work by swapping base_url and the key. Five-minute guide.
• Prepaid credits, no subscription — top up by VietQR bank transfer (VND) or card (USD), spend per real token. Balance runs out → requests stop. Nothing to cancel, no surprise invoice.
• Billing built to be checked — the final usage chunk of every stream carries the token counts you are billed on, full cache hits bill at 10% of price, and a request that fails before producing a single completion token costs nothing. How billing works.
• Boring reliability, in public — status page, docs covering errors, rate limits and idempotency, and measured numbers instead of adjectives.

40% off everything, for launch

For the launch period, every model is billed at 40% off list price. There is no code to enter and nothing to enable: the discount is applied automatically to every request, and the pricing page shows both the list price and the price you actually pay, synced live from the billing engine — what you see there is what the meter runs on.

The numbers we publish

• 13.6 ms gateway overhead p95 — measured over 6,033 requests under sustained load, August 2026.
• 0.65 s median time to first token on glm-4.7-flash, streamed through the edge. We publish the median rather than a p95 because the tail moves with upstream queueing — across two runs on the same day it ranged from 1.5 s to 6.4 s, and a number that unstable is not a promise.
• 10% of list price for a full prompt-cache hit — automatic, no configuration.

These come from the same load and chaos drills we run before every launch-critical change; when the numbers move, the site moves with them.

Start in three steps

1. Create an account at app.clfaigateway.dev — about 30 seconds, no card required.
2. Top up from $5 (VietQR or card) and create your sk-gw-... key.
3. Point your existing OpenAI client at https://api.clfaigateway.dev/v1.

curl https://api.clfaigateway.dev/v1/chat/completions \
  -H "Authorization: Bearer $CLF_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "glm-4.7-flash", "messages": [{"role": "user", "content": "Hello!"}], "max_tokens": 64}'

Questions, edge cases, or something in the docs that reads wrong? Write to [email protected] — during launch, mail goes straight to the people who built this.