[ public beta ] — $1 free credits on every new account

[ runaii cloud — serverless · dedicated · training ]

Own your
inference.

Serverless tokens, dedicated Blackwell GPUs, and training — one API your SDK already speaks. Cached prefixes up to 97% off.

Start building — $1 free Read the docs

no card · no sales call · how we benchmarked the incumbents

runaii — zshexit 0 · 312ms

curl https://api.runaii.cloud/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $RUNAII_API_KEY" \
  -d '{
    "model": "runaii/glm-5.3",
    "messages": [{"role": "user", "content": "Say hello in Spanish"}]
  }'

live

OpenAI + Anthropic compatibletry the live playground →

<0ms: P50 cold start, B200
0 tok/s: peak MoE throughput
0+: serverless models
$0: free, no card
$0.000/M: cached input · up to 97% off

gateway livebilled to the microstatus →

[ featured — glm-5.3-flash ]

Frontier speed. $0.03 cached.

The flash-class workload everyone benchmarks — we ship it with 1M context, 412 tok/s on our fleet, and cache-hit input at under a nickel. One key, every model.

Try it — $1 free Model card →

input

$0.15

/ 1M tokens

cached

$0.030

/ 1M tokens

output

$0.50

/ 1M tokens

B — fleet speeds

Tokens per second, on our fleet

Median output speed, Standard tier. Same architecture, honest numbers.

GLM 5.3 Flash	412 tok/s	1M ctx	$0.15/1M · ⚡0.030 cached
DeepSeek V4 Flash	385 tok/s	1M ctx	$0.22/1M · ⚡0.022 cached
OpenAI gpt-oss-120b	260 tok/s	131K ctx	$0.15/1M
MiniMax M3	190 tok/s	512K ctx	$0.30/1M
Llama 3.3 70B Instruct	175 tok/s	131K ctx	$0.23/1M
GLM-5.3	118 tok/s	1M ctx	$1.40/1M · ⚡0.260 cached
Qwen3.8-Max	104 tok/s	262K ctx	$2.00/1M
Kimi K3	96 tok/s	1M ctx	$3.00/1M

how we benchmarked the incumbents →

[ open models — one api, one key ]

all 15 →

GLM-5.3$1.40/1M· ⚡$0.260 cached GLM 5.3 Flash$0.15/1M· ⚡$0.030 cached Kimi K3$3.00/1M DeepSeek V4 Pro$1.32/1M· ⚡$0.044 cached DeepSeek V4 Flash$0.22/1M· ⚡$0.022 cached Qwen3.8-Max$2.00/1M Qwen3.8 Flash$0.40/1M OpenAI gpt-oss-120b$0.15/1M OpenAI gpt-oss-20b$0.05/1M MiniMax M3$0.30/1M Llama 3.3 70B Instruct$0.23/1M Qwen3 Embedding 8B$0.10/1M Qwen3 Reranker 8B$0.20/1M Whisper V3 Large$0.25/1M FLUX.1 Kontext Pro$0.04/1M

runaii/glm-5.3 $1.40/$4.40/1M · $0.26 cached/runaii/glm-5.3-flash $0.15/$0.50/1M · $0.03 cached/runaii/kimi-k3 $3.00/$15.00/1M/runaii/deepseek-v4-pro $1.32/$3.96/1M · $0.044 cached/runaii/qwen3.8-max $2.00/$6.00/1M/runaii/gpt-oss-120b $0.15/$0.60/1M/runaii/minimax-m3 $0.30/$1.20/1M/runaii/llama-3.3-70b $0.23/$0.85/1M/runaii/whisper-v3 $0.25/M/runaii/flux-kontext $0.04/img/

A — 01/03

The stack

Three primitives. One control plane. No replatforming between stages.

01 / Serverless

Per-token inference. Zero capacity planning.

Fifteen open models behind one OpenAI-compatible URL. Standard, Priority, and Fast tiers; cached prefixes bill for almost nothing.

1M ctx · streaming · tools →

02 / Deployments

Your weights on B200s, billed per second.

Reserve Blackwell-class GPUs with autoscaling and scale-to-zero. Your VPC, your region pin, your control plane.

B200 · B300 · GB300 →

03 / Training

LoRA to RL, then straight to prod.

Managed SFT/DPO per 1M tokens, dedicated RL per GPU-second. Every checkpoint deploys to inference in seconds.

SFT · DPO · RL →

[ metered in real time ]

Watch every token land.

Per-request receipts: tokens in/out, cache hit rate, latency, and the exact micros charged — streamed to your console the moment a call finishes.

Open the console receipts doc →

[ try it — watch a request happen ]

Type a prompt. See the bill.

A working preview of the request flow — route, stream, meter. The same path your code takes, minus the API key.

POST /v1/chat/completions · glm-5.3-flashdemo

$ Explain vector search in one line
— hit Run to watch tokens stream —

Making request—

Visual demo of the request flow — route, stream, meter. Run it against your own key in the playground →

Silicon, priced per second

all gpus →

8× NVIDIA B300 288GB	FP8	Max quality for frontier dense models	$96/hr $0.0267/s
8× NVIDIA B200 180GB	NVFP4	Best tokens/$ for large MoE models	$80/hr $0.0222/s
4× NVIDIA B300 288GB	NVFP4	Smallest frontier-capable footprint	$48/hr $0.0133/s
4× NVIDIA B200 180GB	NVFP4	Great for ≤70B dense models	$40/hr $0.0111/s
16× NVIDIA GB300	FP8	For 1T+ parameter models	$184/hr $0.0511/s

B — receipts

Us vs everybody

full 7-way table →

Serverless APIs

15+ open models, one URL

Modal: SDK-first · Baseten: curated set

Dedicated GPUs

B200/B300/GB300, per-second

RunPod: 30+ SKUs · Koyeb: A100/L40S

Training loop

Checkpoint → endpoint in seconds

Fireworks: the same loop, closed walls

Start cost

$1 free, no card, no sales call

Baseten: sales-led · Koyeb: credits via program

Model library

all 15 →

GLM-5.3NEWFrontier agentic model. Top-tier coding, tool use, and long-context reas…$1.40/$4.40⚡ $0.260/M cached1M ctx GLM 5.3 FlashNEWUltra-fast multimodal workhorse for chat, extraction, and routing.…$0.15/$0.50⚡ $0.030/M cached1M ctx Kimi K3NEWFrontier open-weights model for coding, reasoning, and long context.…$3.00/$15.001M ctx DeepSeek V4 ProSWE-bench leader. Strong math, code, and agentic tool use.…$1.32/$3.96⚡ $0.044/M cached1M ctx DeepSeek V4 FlashSpeed-tuned V4 for high-volume pipelines.…$0.22/$0.66⚡ $0.022/M cached1M ctx Qwen3.8-MaxQwen flagship. Multilingual strength and dense world knowledge.…$2.00/$6.00262K ctx

Ship tonight

cookbook →

example_01.py

Streaming chat in 9 lines

OpenAI SDK, new base URL, token-by-token SSE.

model="runaii/glm-5.3-flash"

example_02.py

RAG that cites sources

Embed → retrieve → rerank → answer, cached.

model="runaii/qwen3-embedding-8b"

example_03.py

Tool-calling agent loop

Declare tools, loop 8 deep, cap the spend.

model="runaii/kimi-k3"

[ works with — your stack, unchanged ]

OpenAI SDK Anthropic SDK LangChain Vercel AI SDK LlamaIndex Cursor Claude Code curl

-60%

model bill after switching

We moved our main agent to an open model and nobody noticed — except the bill dropped 60%.

Platform team — beta partner

<1s

cold start on B200s

Cold starts under a second on B200s. Evals went from overnight to over lunch.

ML lead — inference beta

1 day

base-URL migration to prod

One base URL change and we were compatible. The playground sold the team in a day.

Founding engineer — early access

customer stories →

Start with $1.
Stay for the bill.

Get started →Pricing

[ public beta ] — $1 free credits on every new account

runaiicloud

[ runaii cloud — serverless · dedicated · training ]

Own your
inference.

Serverless tokens, dedicated Blackwell GPUs, and training — one API your SDK already speaks. Cached prefixes up to 97% off.

Start building — $1 free Read the docs

no card · no sales call · how we benchmarked the incumbents

runaii — zshexit 0 · 312ms

curl https://api.runaii.cloud/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $RUNAII_API_KEY" \
  -d '{
    "model": "runaii/glm-5.3",
    "messages": [{"role": "user", "content": "Say hello in Spanish"}]
  }'

live

OpenAI + Anthropic compatibletry the live playground →

<0ms: P50 cold start, B200
0 tok/s: peak MoE throughput
0+: serverless models
$0: free, no card
$0.000/M: cached input · up to 97% off

gateway livebilled to the microstatus →

[ featured — glm-5.3-flash ]

Frontier speed. $0.03 cached.

The flash-class workload everyone benchmarks — we ship it with 1M context, 412 tok/s on our fleet, and cache-hit input at under a nickel. One key, every model.

Try it — $1 free Model card →

input

$0.15

/ 1M tokens

cached

$0.030

/ 1M tokens

output

$0.50

/ 1M tokens

B — fleet speeds

Tokens per second, on our fleet

Median output speed, Standard tier. Same architecture, honest numbers.

GLM 5.3 Flash	412 tok/s	1M ctx	$0.15/1M · ⚡0.030 cached
DeepSeek V4 Flash	385 tok/s	1M ctx	$0.22/1M · ⚡0.022 cached
OpenAI gpt-oss-120b	260 tok/s	131K ctx	$0.15/1M
MiniMax M3	190 tok/s	512K ctx	$0.30/1M
Llama 3.3 70B Instruct	175 tok/s	131K ctx	$0.23/1M
GLM-5.3	118 tok/s	1M ctx	$1.40/1M · ⚡0.260 cached
Qwen3.8-Max	104 tok/s	262K ctx	$2.00/1M
Kimi K3	96 tok/s	1M ctx	$3.00/1M

how we benchmarked the incumbents →

[ open models — one api, one key ]

all 15 →

A — 01/03

The stack

Three primitives. One control plane. No replatforming between stages.

01 / Serverless

Per-token inference. Zero capacity planning.

Fifteen open models behind one OpenAI-compatible URL. Standard, Priority, and Fast tiers; cached prefixes bill for almost nothing.

1M ctx · streaming · tools →

02 / Deployments

Your weights on B200s, billed per second.

Reserve Blackwell-class GPUs with autoscaling and scale-to-zero. Your VPC, your region pin, your control plane.

B200 · B300 · GB300 →

03 / Training

LoRA to RL, then straight to prod.

Managed SFT/DPO per 1M tokens, dedicated RL per GPU-second. Every checkpoint deploys to inference in seconds.

SFT · DPO · RL →

[ metered in real time ]

Watch every token land.

Per-request receipts: tokens in/out, cache hit rate, latency, and the exact micros charged — streamed to your console the moment a call finishes.

Open the console receipts doc →

[ try it — watch a request happen ]

Type a prompt. See the bill.

A working preview of the request flow — route, stream, meter. The same path your code takes, minus the API key.

POST /v1/chat/completions · glm-5.3-flashdemo

$ Explain vector search in one line
— hit Run to watch tokens stream —

Making request—

Visual demo of the request flow — route, stream, meter. Run it against your own key in the playground →

Silicon, priced per second

all gpus →

8× NVIDIA B300 288GB	FP8	Max quality for frontier dense models	$96/hr $0.0267/s
8× NVIDIA B200 180GB	NVFP4	Best tokens/$ for large MoE models	$80/hr $0.0222/s
4× NVIDIA B300 288GB	NVFP4	Smallest frontier-capable footprint	$48/hr $0.0133/s
4× NVIDIA B200 180GB	NVFP4	Great for ≤70B dense models	$40/hr $0.0111/s
16× NVIDIA GB300	FP8	For 1T+ parameter models	$184/hr $0.0511/s

B — receipts

Us vs everybody

full 7-way table →

Serverless APIs

15+ open models, one URL

Modal: SDK-first · Baseten: curated set

Dedicated GPUs

B200/B300/GB300, per-second

RunPod: 30+ SKUs · Koyeb: A100/L40S

Training loop

Checkpoint → endpoint in seconds

Fireworks: the same loop, closed walls

Start cost

$1 free, no card, no sales call

Baseten: sales-led · Koyeb: credits via program

Model library

all 15 →

Ship tonight

cookbook →

example_01.py

Streaming chat in 9 lines

OpenAI SDK, new base URL, token-by-token SSE.

model="runaii/glm-5.3-flash"

example_02.py

RAG that cites sources

Embed → retrieve → rerank → answer, cached.

model="runaii/qwen3-embedding-8b"

example_03.py

Tool-calling agent loop

Declare tools, loop 8 deep, cap the spend.

model="runaii/kimi-k3"

[ works with — your stack, unchanged ]

OpenAI SDK Anthropic SDK LangChain Vercel AI SDK LlamaIndex Cursor Claude Code curl

-60%

model bill after switching

We moved our main agent to an open model and nobody noticed — except the bill dropped 60%.

Platform team — beta partner

<1s

cold start on B200s

Cold starts under a second on B200s. Evals went from overnight to over lunch.

ML lead — inference beta

1 day

base-URL migration to prod

One base URL change and we were compatible. The playground sold the team in a day.

Founding engineer — early access

customer stories →

Start with $1.
Stay for the bill.

Get started →Pricing

Own yourinference.

Frontier speed. $0.03 cached.

Tokens per second, on our fleet

The stack

Per-token inference. Zero capacity planning.

Your weights on B200s, billed per second.

LoRA to RL, then straight to prod.

Watch every token land.

Type a prompt. See the bill.

Silicon, priced per second

Us vs everybody

Model library

Ship tonight

Streaming chat in 9 lines

RAG that cites sources

Tool-calling agent loop

Start with $1.Stay for the bill.

Own yourinference.

Frontier speed. $0.03 cached.

Tokens per second, on our fleet

The stack

Per-token inference. Zero capacity planning.

Your weights on B200s, billed per second.

LoRA to RL, then straight to prod.

Watch every token land.

Type a prompt. See the bill.

Silicon, priced per second

Us vs everybody

Model library

Ship tonight

Streaming chat in 9 lines

RAG that cites sources

Tool-calling agent loop

Start with $1.Stay for the bill.

Own your
inference.

Start with $1.
Stay for the bill.

Own your
inference.

Start with $1.
Stay for the bill.