Pricing · $5 free on signup
Pay for tokens, not for vibes. Serverless per-token inference, per-second dedicated GPUs, and training priced per 1M tokens. Every price on one page — no hidden multipliers.
Per 1M tokens · OpenAI + Anthropic compatible
| Model | Modality | Input / 1M | Cached / 1M | Output / 1M | Context |
|---|---|---|---|---|---|
| GLM-5.3NEW | LLM | $1.40 | $0.260 | $4.40 | 203K |
| GLM 5.3 FlashNEW | Vision | $0.15 | $0.030 | $0.50 | 66K |
| DeepSeek V4.1 Flash | LLM | $0.22 | $0.022 | $0.66 | 1.048576M |
| Qwen3.8-27B Flash | LLM | $0.40 | $0.040 | $1.60 | 33K |
Swipe the table sideways for the cached, output, and context columns.
Batch API (demo): async jobs, up to 100k requests per job, same gateway. Demo jobs are simulated — nothing is processed or billed yet.
† Listed price under review (catalog id / weights divergence) — confirmed pricing lands with the next billing sweep; you are charged the table rate until then.
Prefer flat-rate over per-token? Unlimited-token founding plans from $12/mo → (₹1,999/mo in India; 45 prepaid seats light one dedicated node.)
Paste one key into your app — billed per token from the table above.
Per-second billing · autoscale + scale-to-zero · your weights, your VPC.
Efficient ×4
$3.60/hr
NVIDIA L4 24GB ×4
FP8 · g2-standard-48 — multi-worker farm on one VM
Deploy this presetWorkstation 1/8
$0.30/hr
NVIDIA RTX PRO 6000 Blackwell (1/8 GPU · 12GB) ×1
FP8 · g4-standard-6 — Blackwell slice for light jobs
Deploy this presetWorkstation 1/4
$0.60/hr
NVIDIA RTX PRO 6000 Blackwell (1/4 GPU · 24GB) ×1
FP8 · g4-standard-12 — quarter GPU
Deploy this presetWorkstation 1/2
$1.20/hr
NVIDIA RTX PRO 6000 Blackwell (1/2 GPU · 48GB) ×1
FP8 · g4-standard-24 — half GPU
Deploy this presetWorkstation
$2.40/hr
NVIDIA RTX PRO 6000 Blackwell 96GB ×1
FP8 · g4-standard-48 — try any app from $2.40/hr (park/wake receipted D27)
Deploy this presetWorkstation ×2
$4.80/hr
NVIDIA RTX PRO 6000 Blackwell 96GB ×2
FP8 · g4-standard-96 — two Blackwell GPUs
Deploy this presetWorkstation ×4
$9.60/hr
NVIDIA RTX PRO 6000 Blackwell 96GB ×4
FP8 · g4-standard-192 — four GPUs
Deploy this presetWorkstation ×8
$19.20/hr
NVIDIA RTX PRO 6000 Blackwell 96GB ×8
FP8 · g4-standard-384 — full-node Blackwell
Deploy this presetFull Precision
Popular$96.00/hr
NVIDIA B300 288GB ×8
FP8 · Max quality for frontier dense models
Deploy this presetThroughput
Popular$80.00/hr
NVIDIA B200 180GB ×8
NVFP4 · Best tokens/$ for large MoE models
Deploy this presetMinimal
Popular$48.00/hr
NVIDIA B300 288GB ×4
NVFP4 · Smallest frontier-capable footprint
Deploy this presetFull 8× H100 (80GB) VMs bill the same way — H100 pricing, Spot and regions →
Managed SFT/DPO priced per 1M training tokens (dataset tokens × epochs). RL jobs bill per GPU-second at on-demand rates. Checkpoints deploy to inference at base-model prices.
| Model size | LoRA SFT | LoRA DPO | Full SFT | Full DPO |
|---|---|---|---|---|
| Up to 16B params e.g. gpt-oss-20b | $0.50 | $1.00 | $1.00 | $2.00 |
| 16–80B params e.g. llama-3.3-70b | $3.00 | $6.00 | $6.00 | $12.00 |
| 80–300B params e.g. qwen3.8-235b-class | $6.00 | $12.00 | $12.00 | $24.00 |
| 300B+ params e.g. glm-5.3 | $10.00 | $20.00 | $20.00 | $40.00 |
Swipe the table sideways for the full rate matrix.
Per 1M training tokens · VLM fine-tuning billed the same way · open training console →
Training is free during beta — published rates apply only when billing goes live.
Early-stage teams get $500 in credits, locked beta pricing, and a direct line to engineering.
Startup program →[ credits ]
One balance across everything we meter — serverless tokens, per-second GPU deployments, and training runs. The ledger splits every call into a line item, so you always know exactly what it cost.
$5 free on signup · no card · balance never expires
Serverless tokens
A growing model catalog, one key. Cached prefixes up to 90% off.
Dedicated GPU-seconds
B200/B300/GB300, autoscale, scale-to-zero. Same balance.
Training runs
SFT/DPO per 1M tokens, RL per GPU-second. Ships to prod in seconds.
Never expires
Sit on your balance for a year — it stays yours.
Usage is always metered separately. Plans buy limits, support, and guarantees.
$0 + usage
For developers getting a first agent into production.
$250/mo + usage
For startups scaling inference with guardrails.
Custom
For regulated scale: residency, VPC, and contracts.
No — your balance never expires. Top up once and spend it across serverless tokens, GPU deployments, and training runs whenever you're ready.
About 3.5M input tokens on GLM-5.3, or ~22M on DeepSeek V4.1 Flash. Enough to prototype a full agent loop before paying.
Repeat prompt prefixes (system prompts, RAG context) are billed at the cached rate — up to 81% off on GLM-5.3 and 90% off on DeepSeek V4.1 Flash. Cache hits are automatic.
Serverless for variable traffic and prototyping. Dedicated (B200/B300/GB300, per-second) when you need guaranteed throughput, custom weights, or VPC isolation.
The Batch API (beta) runs async jobs — evals, backfills, up to 100k requests — through the same gateway and ledger. There's no separate batch discount while it's in beta.
Global routing is cheapest. US-only is the same price; EU/APAC add 15% for data-residency guarantees.
Estimate what moving your monthly inference spend to an open-model mix saves you. Rows above are the real prices.
$900
closed APIs
$17.6
runaii mix
$882.4
you keep / mo
Estimate on DeepSeek V4.1 Flash ($0.22/1M in · $0.66/1M out); exact numbers on this page. Cached prefixes cut it further. Start with $5 free →