MarsAI · PRICING
Transparent pricing · No markup · Pay as you go
No minimums, no lock-in, no charge for failed requests — what you see is what you pay. Models by the token, GPUs by the hour.
01·MODEL PRICING
Model API (per token)
| Model | Context | Input $/1M | Output $/1M | Cache | Tier |
|---|---|---|---|---|---|
| gpt-5.1 | 256K | $1.25 | $10.00 | $0.13 | Official price, passed through |
| claude-opus-4-8 | 200K | $5.00 | $25.00 | $0.50 | Official price, passed through |
| gemini-3-pro | 1M | $2.00 | $12.00 | — | Official price, passed through |
| deepseek-v3.2 | 164K | $0.27 | $0.40 | $0.03 | Self-hosted |
| qwen3-max | 256K | $1.25 | $3.75 | — | Self-hosted |
| llama-4-maverick | 1M | $0.20 | $0.60 | — | Self-hosted |
| kimi-k2 | 256K | $0.80 | $3.40 | — | Self-hosted |
| glm-5 | 1M | $1.40 | $4.40 | — | Self-hosted |
Premium passes through at the official price with zero markup, verifiable line by line. Open-source rows show reference unit prices for our self-hosted, optimized inference (in line with the market) — see the console for live pricing ().
02·GPU PRICING
GPU compute (per GPU-hour)
- 24GB GDDR6X/GPU · 192GB total
- PCIe 4.0
- Dedicated node · 0 egress
- 32GB GDDR7/GPU · 256GB total
- PCIe 5.0
- Dedicated node · 0 egress
- 48GB GDDR6/GPU · 384GB total
- PCIe 4.0
- Dedicated node · 0 egress
- 80GB HBM2e/GPU · 640GB total
- NVLink 600 GB/s + RDMA
- Dedicated node · 0 egress
- 80GB HBM3/GPU · 640GB total
- NVLink 900 GB/s + RDMA
- Dedicated node · 0 egress
- 141GB HBM3e/GPU · 1,128GB total
- NVLink 900 GB/s + RDMA
- Dedicated node · 0 egress
- 192GB HBM3e/GPU · 1,536GB total
- NVLink 1.8 TB/s + RDMA
- Dedicated node · 0 egress
- 192GB HBM3e/GPU · 1,536GB total
- NVLink 1.8 TB/s
- Dedicated node · 0 egress
Dedicated bare metal, zero egress fees. Monthly reserved runs ~30% below hourly; top-end SKUs are starting prices. Figures are indicative ().
03·RATE LIMITS
Rate limits
Sensible defaults that scale with your usage — every limit can be raised on request.
Pro
Default- 600Requests / min
- 2MTokens / min
- 60Concurrent
Scale
- 6,000Requests / min
- 20MTokens / min
- 400Concurrent
Enterprise
- CustomRequests / min
- CustomTokens / min
- DedicatedConcurrent
Limits are enforced per account at the gateway (RPM = requests/min, TPM = tokens/min) and rise automatically as your spend grows. Premium pass-through models also honor the upstream provider's own limits.
04·BILLING FAQ
Billing FAQ
Q1How are tokens billed?
Priced per token. Input and output are billed separately — see the table above (per 1M). Premium matches the official price; open-source uses our self-hosted, optimized pricing.
Q2Is there a cache discount?
Yes. Cached input tokens get a discount (e.g. 0.5×, ), cutting repeated-context cost significantly.
Q3How are GPUs billed?
By the GPU-hour, metered per minute underneath — start and stop anytime; monthly reserved ~30% off, zero egress fees.
Q4Do failed requests cost money?
No. Failed requests don't count toward usage; clearly anomalous requests are waived automatically.
Q5Any minimums or lock-in?
None. No minimum spend, no annual commitment — stay or leave whenever you like.
Q6Is my data used for training?
Never. Your requests and data are encrypted end to end and never used to train any model.
Start on credit. Scale when you're ready
Pay as you go, no minimums — the models and compute your agents need, all in a transparent price list.