Skip to main content

Inference

Frontier models. Flat rate.
No meter.

Kimi K2.6 and K3 behind OpenAI- and Anthropic-compatible APIs, with Qwen embeddings + TTS included. Up to 100 million tokens a day — your bill never moves.

two lines to switch — either SDK
const client = new OpenAI({
  baseURL: "https://api.hypercli.com/v1",
  apiKey: process.env.HYPER_KEY });

Works with OpenAI and Anthropic SDKs, LangChain, LlamaIndex — anything speaking /v1/chat/completions or /v1/messages.

What would your volume cost elsewhere?

Published list prices, July 2026. Drag to your daily usage.

50M
Claude Fable 5 ($10 / $50 per M)$27,000/mo
GPT-5.6 Sol ($5 / $30 per M)$15,000/mo
Claude Opus 5 ($5 / $25 per M)$13,500/mo
GPT-5.6 Terra ($2 / $12 per M)$6,000/mo
HyperCLI — Pro (Kimi K3), flat$149/mo
HyperCLI — Team (Kimi K2.6), flat$79/mo

Assumes 80% input / 20% output at list prices, 30-day month. Solo includes 25M tokens/day, Team 50M, Pro 100M — pooled across your agents and your API key.

Every model, one key.

Open weights, frontier results. Numbers below are published benchmarks, not our marketing.

Standard · chat · vision · tools

Kimi K2.6

The workhorse. A 1T-parameter MoE built for long-horizon coding and autonomous execution — it beat GPT-5.5 at coding on release. 256K context, agent swarms to 300 sub-agents.

Advanced · chat · vision · tools

Kimi K3

The heavyweight. 2.8T parameters — the largest open model ever released — reading entire repos in a single 1M-token prompt. Beats Fable 5 on Terminal-Bench; #1 in blind frontend testing.

Included · memory · voice

Qwen Embeddings + TTS

The quiet ones. Dense open models running your embeddings, vector search, agent memory, and voice — at no extra cost.

Two ways in.

Same models, same flat rate. Start where you are.

Just the API

You have an app. Point it here. No migration, no new SDK — your OpenAI or Anthropic code runs unmodified.

terminal
$ curl https://api.hypercli.com/v1/chat/completions \
  -H "Authorization: Bearer $HYPER_KEY" \
  -d '{"model":"kimi-k2.6","messages":[...]}'

Get an API key →

The whole agent

Same inference — plus a cloud machine that's always on, with browser, desktop, voice, media, memory, and every channel. The models are the brain; this is the body.

terminal
$ hyper agents create my-agent
✓ live on its own machine

See everything it can do →

Questions you should be asking.

100M tokens a day — what's the catch?
There isn't a hidden one. The allowance resets daily, and we can offer it flat because we run open models on a global fabric of aggregated GPU capacity. Fair-use terms apply to automated abuse (key sharing, resale), not to heavy legitimate use.
What happens if I hit the cap?
Requests throttle until the daily reset — you never get a surprise bill. Need more headroom? Add another key and keep going: each subscription carries its own allowance, so scaling up is a checkout, not a negotiation.
Is it really OpenAI-compatible?
Yes — and Anthropic-compatible too. Standard /v1/chat/completions and /v1/messages surfaces with streaming, tool calling, and vision. OpenAI SDKs, Anthropic SDKs, LangChain, and LlamaIndex all work by changing the base URL and key.
Do you train on my data?
No. Prompts and completions are not used for training.
Why open models instead of a frontier lab API?
Kimi K2.6 beats GPT-5.5 at coding, and Pro gets the frontier K3. Open weights mean no vendor lock-in — the models that run your workload are public, so you can verify claims and always have an exit.
API key vs. full agent — which should I pick?
If you're adding inference to an existing app, take the key. If you want an always-on agent with a browser, voice, memory, and channels, deploy the agent — and you can add the other later on the same account.

Stop counting tokens.

Frontier inference, one flat bill. Live in minutes either way.