List prices with their dates and sources, computed in your browser. Nothing you type leaves it.
What the same AI load costs self-hosted: the cheapest cloud API next to one rented GPU, list prices as of 9 September 2026.
Self-hosted saves
$57,442
a year: $4,787 a month, 89% less than Anthropic
at your volume, against the cheapest cloud API, list prices as of 9 September 2026
| Option | Model or hardware | A month |
|---|---|---|
| Cheapest cloud API Anthropic (Claude API) |
Claude Sonnet 5 | $5,400 |
| Self-hosted, one rented RTX 6000 Ada billed by the month |
Qwen 3.5 35B-A3B MoE, FP8 1 × NVIDIA RTX 6000 Ada 48GB, 730 GPU hours at 70% utilisation |
$613 |
| Your own hardware at scale our fleet, measured |
$0.114 per 1M tokens at 70% utilisation | $185 |
Same class of model for the same job. Hardware and hosting only: setup and running it is our fixed-price work. Top tier: quality parity not claimed.
| Provider | Model | In / out per 1M | A month |
|---|---|---|---|
| Anthropic (Claude API) | Claude Sonnet 5 | $2.00 / $10.00 | $5,400 |
| OpenAI | GPT-5.6 Terra | $2.00 / $12.00 | $5,940 |
| Amazon Bedrock (Claude) | Claude Sonnet 5 | $2.00 / $10.00 | $5,400 |
Bedrock lists Claude at Anthropic prices on global endpoints; single-region endpoints cost about 10% more.
| Boxes needed | 1 × 2 cards |
| Capex | $22,000 |
| Running cost a month | $250 |
| Amortised a month, 36 months | $861 |
| Payback against the cheapest cloud bill | 5 months |
Volume. 1.35B input and 270M output tokens a month (30 days). Cloud API = tokens × list price, no discounts or batch pricing.
Models. Fast: Qwen 3.5 35B-A3B MoE, FP8, 1x GPU. Standard: Qwen 3.5 35B-A3B MoE, FP8. Top: Qwen 3.5, largest open-weight size, FP8, multi-GPU. For Fast and Standard, parity with the cloud model is measured on your golden set before release. For Top we do not claim parity.
Rented GPU. NVIDIA RTX 6000 Ada 48GB, $0.84 per hour, a full month per card at minimum. GPU hours = (output tokens + input tokens / 10) / (320 tokens a second × 3600 × utilisation). We pick the card the way we would for a client: the cheapest class that runs the tier.
Own fleet. $0.084 per 1M tokens all-in at 95% utilisation, measured on NVIDIA L40S, Qwen 3.5 35B-A3B MoE FP8, production content moderation, 12B tokens a week; scaled to your utilisation.
Sources. Anthropic, OpenAI and AWS Bedrock list pricing pages; RunPod Secure Cloud list prices; our self-hosted inference platform case. All checked 9 September 2026. 1 USD = 0.92 EUR.
Unrote Assist for your team: the subscription, the server it runs on and the GPU, a month. Prices as of 2026-10-16.
Unrote Assist for 10 people
€2,013
a month, everything in: €201 per seat
nothing to buy up front
| What | Which | A month |
|---|---|---|
| Unrote Assist €1,500 base for 10 seats, €40 a seat to 100, €25 above |
base with 10 seats, then per seat | €1,500 |
| Company server | Hetzner Cloud CX43 | €16 |
| Backups | Hetzner backups €3 + Storage Box BX11, 1 TB | €6 |
| Qwen 3.8 27B FP8 on a rented RTX 6000 Ada 48 GB hardware only, at the provider's price |
1 card(s), 730 GPU hours | €491 |
Hardware at the providers' list prices with their dates; Unrote Assist on top. 2 of 10 seats active at the peak; 1 card(s) needed.
| Mode | Cards | Hardware a month | All in | Once |
|---|---|---|---|---|
| 24/7 | 1 | €513 | €2,013 | - |
| Working hours | 1 | €156 | €1,656 | - |
| On the first message | 1 | €89 | €1,589 | - |
| Your own card | 1 | €499 | €1,999 | €13,774 |
| Hetzner GPU server | 1 | €1,021 | €2,521 | €999 |
| Gemini 3.8 Flash (to 2026-12-31) (cloud) | - | €134 | €1,634 | - |
Buying 1 × RTX 6000 Ada 48 GB and a server costs €13,774 once; power and upkeep €94 a month. Against renting the same cards 24/7 it pays back in 35 months; against renting in working hours in 347 months. Write-off over 36 months.
Load. 15% of seats write at once at the peak; a card holds a model-specific number of active seats (Qwen 3.8: 40, Qwen 3.5: 120 on an RTX 6000 Ada; other cards scaled by speed). Our estimate from bench v3, 2026-10-15/16.
Cloud models. 20 messages per person a working day, 35,000 tokens in and 700 out per message (tools and context included), list prices per 1M tokens, no cached share by default.
Rent. RunPod list prices 2026-10-16; disk paid while the pod runs; cold start 10 min, idle stop 20 min. Hetzner GEX63 999 EUR a month + 999 setup.
Own card. Card at street price plus a 1U server for 2 cards ($8,000, estimate), 0.20 EUR per kWh, 50% load, 150 W for the box, 50 EUR a month upkeep, 36-month write-off.
Server. Hetzner Cloud CX43 up to 50 seats, CX53 above; backups 20% of the server; Storage Box for database dumps. ECB rate 1 EUR = 1.1253 USD.
We will run the same numbers on your real traffic and tell you honestly whether self-hosted pays off. If a cloud API is the better deal for you, we say so.