inferaft

Find a page

Where do you want to go?

Open models for your coding agent.A reserved speed lane. Hosted in the US and EU. No prompts stored.

Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash, served from data centers in the US and Finland. DeepSeek V4.1 Flash also runs on a reserved speed lane: about 620 tokens a second when the lane serves you (measured median). Every model lists its quantization and limits, and prompt bodies are never stored. Works in Claude Code, Codex, OpenCode and Cline.

curl -fsSL https://inferaft.com/setup.sh | bash
Tokens served, last 30 days
89M+
Tokens served, last 30 days
Input served from cache
95%
Input served from cache
Prompt bodies stored
0
Prompt bodies stored
  • kimi-k3
  • glm-5.3
  • kimi-k3.1-pro
  • qwen-3.8-27b
  • deepseek-v4-flash
  • deepseek-v4-pro
  • All models

Flash bonus — 20 of 20 spots left

Subscribe to any plan and get $100 of DeepSeek V4.1 Flash at list prices on top of it. Until 2026-11-15 06:00 UTC, or earlier if the reserved capacity runs out.

Get the Flash bonus

Founding 100 — 100 of 100 spots left

The first 100 paid subscribers get +50% tokens on every plan, for as long as they stay subscribed.

Claim a founding spot

Unlimited tickets

500 tickets → 72 hours unlimited. It starts the moment ticket #500 is sold.

0 / 500 sold

  • One $5 ticket covers the whole 72-hour window.
  • Up to 2 requests at a time per account, fair use.
  • If a round misses its sale deadline, your ticket moves to the next round.

Billions of tokens a month, not a few requests.

Each plan is a dollar pool spent at the posted per-token rates. Coding agents resend the same context, so most input is served from cache at a tenth of the price. This is how far each plan goes on one model.

Tokens per monthDeepSeek V4.1 FlashGLM 5.3 FlashGLM 5.3Kimi K3
Lite$10/mo
1.1B194M with no cacheFounding: 1.6B
731M195M with no cacheFounding: 1.1B
117M20M with no cacheFounding: 175M
51M10M with no cacheFounding: 76M
Pro$20/mo
2.1B387M with no cacheFounding: 3.2B
1.5B390M with no cacheFounding: 2.2B
234M40M with no cacheFounding: 350M
101M19M with no cacheFounding: 152M
Max$50/mo
5.3B968M with no cacheFounding: 8.0B
3.7B975M with no cacheFounding: 5.5B
584M101M with no cacheFounding: 876M
253M48M with no cacheFounding: 379M
Ultra$100/mo
11B1.9B with no cacheFounding: 16B
7.3B2.0B with no cacheFounding: 11B
1.2B202M with no cacheFounding: 1.8B
505M96M with no cacheFounding: 758M

Up to, per 30-day term, if one model uses the whole plan. Big figures assume typical coding-agent traffic: 95% of input served from cache, output about 1.1% of input. The smaller figure assumes no cache. Rolling 5-hour and 7-day windows still pace usage. +50% for Founding members.

Keep your agent, or use ours.

Both run on the same key, the same catalog and the same plan.

Inferaft API

Your agent, our models.

Point the agent you already run, or any OpenAI-compatible client, at one base URL. Your prompts, tools and habits stay as they are.

  • Claude Code
  • Codex
  • OpenCode
  • Cline
  • Cursor
See the setup

Inferaft Code

Our agent, made for them.

A terminal agent with sub-agents, a macOS sandbox, rewind and memory. Included in every plan.

brew install inferaft/tap/inferaft-code
Explore Inferaft Code

A gateway shaped around agent work.

Long contexts, heavy caching and many small turns. Inferaft is built for that traffic, and shows you what each turn cost.

Switch models without rewiring

Keep the same endpoint and client configuration when you move between models. Plan with a flagship, iterate with a fast one.

Three wire formats

Whatever your agent speaks, it reaches the same catalog.

  • POST/v1/chat/completions
  • POST/v1/messages
  • POST/v1/responses

No surprise charges

One shared allowance, paced by rolling windows. At the limit you get a clear plan_usage_limit error, never an overage bill.

Last 5 hours
Last 7 days
This term

Your prompts stay yours

Request history stores metadata and token counts. Prompt and response bodies are not retained.

More than you pay for

Every plan includes about 2.4x its price in usage at the published model rates, with the limits shown before you buy.

Connected in a minute.

No SDK to adopt, no proxy to run, no config format to learn.

  1. 01Create a key

    One key per project or tool, with its own usage history.

  2. 02Paste the base URL

    Point your agent at inferaft.com. Claude Code, Codex, OpenCode and Cline each have a ready-made block.

  3. 03Pick a model and ship

    Change models by name whenever you like. The endpoint and key stay the same.

All client guides
export ANTHROPIC_BASE_URL="https://inferaft.com"
export ANTHROPIC_API_KEY="$INFERAFT_API_KEY"
unset ANTHROPIC_AUTH_TOKEN
export ANTHROPIC_MODEL="custom/kimi-k3"
export ANTHROPIC_CUSTOM_MODEL_OPTION="$ANTHROPIC_MODEL"
export CLAUDE_CODE_MAX_OUTPUT_TOKENS=8192
export ANTHROPIC_DEFAULT_SONNET_MODEL="$ANTHROPIC_MODEL"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="$ANTHROPIC_MODEL"
export ANTHROPIC_DEFAULT_OPUS_MODEL="$ANTHROPIC_MODEL"
export ANTHROPIC_SMALL_FAST_MODEL="custom/glm-5.3-flash"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

claude --model "$ANTHROPIC_MODEL"
OpenAI Chat Completions · Anthropic Messages · OpenAI ResponsesClient setup guide

Pick a size. Every model is included.

Every plan uses the same model catalog and agent setup. Choose the amount of included usage that fits your workflow.

Compare plans

Lite

$10/ 30 days

≈1,095 Kimi K3 requests / 30 days

5-hour pace: ≈219 Kimi · ≈737 GLM

Choose Lite

Pro

$20/ 30 days

≈2,189 Kimi K3 requests / 30 days

5-hour pace: ≈438 Kimi · ≈1,473 GLM

Choose Pro

Max

$50/ 30 days

≈5,473 Kimi K3 requests / 30 days

5-hour pace: ≈1,642 Kimi · ≈5,525 GLM

Choose Max

Ultra

$100/ 30 days

≈10,946 Kimi K3 requests / 30 days

5-hour pace: ≈3,284 Kimi · ≈11,049 GLM

Choose Ultra

Every plan includes one shared usage allowance across the catalog. Published token rates are safety ceilings. Included usage cannot be withdrawn, and purchases are final. Refund policy

Questions, answered plainly.

Billing questions are answered on the pricing page. Anything else: support@inferaft.com.

Can I keep my current coding agent?

Yes. Inferaft speaks OpenAI Chat Completions, Anthropic Messages and OpenAI Responses, so most agents need one base URL and one key. The docs carry a copy-paste block for each supported client.

Which models can I use today?

The live catalog is the source of truth for what is available now. You can change models without changing the endpoint or API key.

Do I need Inferaft Code?

No. It is one more client, included in every plan and tuned for the catalog. Any agent that speaks one of the three API formats works just as well.

What do you keep from my requests?

Request history stores metadata and token counts so you can see what each turn cost. Prompt and response bodies are not retained.

Where do I get help?

Email support@inferaft.com with the request ID from the console. Never send your API key or password.

Bring your next idea. Keep your tools.

Create a key, paste one base URL, and your agent is running on open models in under a minute.

See pricing

Card handled by Paddle