License the mind. Keep your machine.

The two planned hosts

The hardware is yours, bought from Apple or NVIDIA at list price; Prismatic does not mark it up. The Jetson AGX Orin Industrial 64 GB (ECC memory, US$3,199 at 1,000+ units) is an option for harsh machines; nothing has run on Jetson yet. Today Prismatic runs on one 16 GB Mac mini M4; neither host in this table has been measured by Prismatic yet.

The licences (proposed list prices)

One flat price per machine, unlimited turns: US$49 per machine a month, unlimited turns: no meter, no usage reports, the same bill however much the mind thinks. There is nothing to meter: a local turn costs Prismatic nothing, so a busy machine pays the same as a quiet one, and the lasting memory a turn reads never costs more.

Sector prices: 27 markets, always unlimited

Each sector is priced in the unit it already buys in, always with unlimited turns, below what that sector already pays per unit for comparable software (for retrofit sites, low enough that at 10 units Prismatic's five-year total, Mac mini included, is still below the sector's own spend), never above a GPT-4.1 cloud mind at 200 turns a day (about US$130 a month), and for sectors priced per person, below US$30 cloud AI seats even with the Mac mini included at 10 people. Proposed list prices, to be validated with the first design partner in each sector.

What a sector pays today is often a whole product (a fleet platform, a robot lease, a field-service suite); Prismatic adds the lasting local memory beside it, so the comparison shows how small Prismatic's line is inside the budget the sector already spends, not a like-for-like replacement. Retrofit markets would use one Mac mini per site; greenfield markets add an activation per unit (Jetson planned, not tested). Each comparison page has a calculator at this price.

Per turn, side by side (200 turns a day, 9,500 tokens in, 300 out)

Cloud prices are list prices per 1M tokens turned into one turn of 9,500 tokens in and 300 out; Copilot Studio and Agentforce bill per interaction or action. Prismatic's figure is its flat licence divided by 200 turns a day; at 1,000 turns a day it is US$0.0016. Its hardware is bought once (below). A very small cloud model can cost less per turn; Prismatic leads there on lasting memory, privacy and working offline, not on price.

Five years, side by side

Estimates. Electricity at 14.2¢/kWh around the clock (30 W Orin, 40 W Mac mini). A smaller cloud model or a high cache share can cost less than Prismatic; the cost-of-a-mind calculator shows when.

Why these prices

• A local mind costs Prismatic nothing per turn, because the customer's own 64 GB host does the thinking, so there is nothing to meter: one flat price per machine, unlimited turns, no usage reports leaving the site, and a bill finance can budget.
• On the comparison charts the flat price is shown as a per-turn equivalent at the same assumptions as the cloud rows (200 turns a day, 9,500 tokens in, 300 out): about US$0.008 a turn, against about US$0.021 for GPT-4.1, US$0.023 for GPT-5.6 Terra and US$0.028 for GPT-5.4 (published list prices).
• Unlimited means heavy use makes Prismatic look better: at 1,000 turns a day the same US$49 is about US$0.0016 a turn, while a GPT-4.1 bill grows to about US$7,800 a year. The cloud charges for every token of memory; Prismatic charges for the machine.
• Robot operations software sells for US$50-500 per robot per month and cobot software for US$500-3,000 a year (reported): US$49 a month sits at the bottom of that range.
• Agent platforms charge per interaction: Microsoft Copilot Studio US$0.01 (about US$730 a year at 200 a day), Salesforce Agentforce US$0.10 per action (about US$7,300). Prismatic is US$588 a year, unlimited.
• Prismatic pays no inference cost, so the licence is close to pure margin: 1,000 machines is about US$588,000 a year of recurring revenue.
• Honestly: a very small cloud model (GPT-5.6 Luna, about US$165 a year at 200 turns a day) is cheaper on price alone; Prismatic leads there on lasting memory, privacy and working offline.
• Hardware is the customer's own, bought from Apple or NVIDIA at list price; Prismatic does not mark it up.

What the market charges today

• OpenAI GPT-5.4: US$2.50 in / US$15 out per 1M tokens (about US$0.028 a turn) Published source
• OpenAI GPT-5.6 Terra: US$2 in / US$12 out per 1M tokens (about US$0.023 a turn) Published source
• OpenAI GPT-5.6 Luna: US$0.20 in / US$1.20 out per 1M tokens (about US$0.002 a turn) Published source
• Microsoft Copilot Studio: US$0.01 per Copilot Credit (one agent interaction), pay as you go Published source
• Salesforce Agentforce: US$0.10 per action (Flex Credits) or US$2 per conversation Published source
• Intercom Fin: US$0.99 per resolution Published source
• Robot fleet and operations software (SaaS): US$50-500 per robot per month Reported source
• Mac mini M5 Pro power: 6 W idle, 145 W CPU maximum (wall) Published source
• Mac mini M5 Pro price: From US$1,699; 64 GB option US$1,000 Reported source
• Jetson AGX Orin prices: 64 GB module US$2,999 (1,000+); developer kit US$3,499; Industrial US$3,199 Published source

Talk to us

Design partners start with the 90-day pilot. Write to info@prismaticagents.com.