Local AI or Frontier AI?
A plain guide for business owners: what each one costs, what it takes to keep running, and where your data goes.
Side by side
We use both kinds of AI every day, and we like both. They're good at different things.
Choose your track
Choose a track, then scroll. The train stops at every station.
What we mean
Local AIOpen models running on a computer your business owns, in your office or on a server you control. A model is the AI program itself, the part that reads and writes. Open models are ones anyone can download and run. Examples are Qwen, Llama, Gemma, Mistral and DeepSeek's open models.
In betweenOther cloud models and routers sit in between. GLM, DeepSeek, Kimi, Qwen and others cost little per use, and a router lets one account reach all of them. Your data goes to more companies, some of them overseas.
Frontier AIThe top models from the big AI companies: Claude, ChatGPT and Gemini. They run in the company's data centers, which are large buildings full of its computers. You pay a monthly fee per person, or you pay for what you use.
Quality on hard work
Local AIGood at everyday writing, summaries and sorting. Behind on the hardest work. In a 2025 study, open models that fit on one graphics card trailed the best cloud models by about 6 to 12 months on public tests25.
In betweenThe best of them, such as GLM and Kimi, are among the strongest open models, which trailed the frontier by about 4 months in early 202624. Their low-cost models are weaker.
Frontier AIThe strongest models you can use today. On Epoch AI's combined test score, the best open models trailed them by about 4 months in early 2026, and the biggest open models are far too big for an office machine24,26.
Where the data goes
Local AIStays on machines you control, as long as every connected tool, log and backup stays local too75.
In betweenTo that company, or through a router to whichever company hosts the model. Rules differ: some store data in China, Singapore or the U.S., and some may train on it45,48,50,54.
Frontier AIGoes to the vendor's data centers. Business plans don't use it for training by default. Personal plans may32–38.
Cost to start
Local AIAbout $2,700 to $52,000 for hardware, plus setup time.
In between$0 to try.
Frontier AI$0 to try.
Cost at heavy use
Local AIAbout the same. You pay more only when you outgrow the machine.
In betweenGoes up with use, from a low base. GLM's coding plan caps it at $18 to $168 a month47.
Frontier AIGoes up with use. Premium seats are $100 to $125 per person1,3. Big automated jobs on pay-per-use can cost thousands a month.
Working offline
Local AIYes.
In betweenNo.
Frontier AINo. It needs the internet.
Upkeep
Local AIReal and ongoing: updates, backups, testing new models, fixing breaks. We estimate 2 to 6 hours a month.
In betweenNo hardware to look after. Workflows still need testing when models change. A router you run yourself, like LiteLLM, adds upkeep64.
Frontier AINo hardware to look after. Workflows you build on it still need testing when models change29,30.
Who fixes it when it breaks
Local AIYou, your IT person, or a service you hire.
In betweenThe host or the router company. You wait, or the router moves you to another host.
Frontier AIThe vendor. You wait.
Vendor changes
Local AIYou decide when anything changes. A model you downloaded keeps working, though updates to other software can still break things.
In betweenFast. Prices and model names change often. MiniMax cut its M3 price in half53, and DeepSeek's older model names now point to its newer model44.
Frontier AIVendors retire models and change prices and terms, with notice6,29,30.
These are two ways of using AI: open models you run yourself, and top commercial models the big companies run for you. You can also rent open models in the cloud.
See all three side by side
| Point | Local AI | In between | Frontier AI |
|---|---|---|---|
| Quality on hard work | Good at everyday writing, summaries and sorting. Behind on the hardest work. In a 2025 study, open models that fit on one graphics card trailed the best cloud models by about 6 to 12 months on public tests25. | The best of them, such as GLM and Kimi, are among the strongest open models, which trailed the frontier by about 4 months in early 202624. Their low-cost models are weaker. | |
| Where the data goes | Stays on machines you control, as long as every connected tool, log and backup stays local too75. | ||
| Cost to start | About $2,700 to $52,000 for hardware, plus setup time. | $0 to try. | $0 to try. |
| Monthly cost | Electricity, about $6 to $80 in typical use, plus upkeep time. | Pay for what you use. Our eight-workflow office costs about $5 to $100 a month on these models, depending on the model. | |
| Cost at heavy use | About the same. You pay more only when you outgrow the machine. | Goes up with use, from a low base. GLM's coding plan caps it at $18 to $168 a month47. | |
| Speed | Depends on the machine. A good setup writes faster than you can read. Big models on small machines are slower27. | Varies by host. A router can switch to a backup host when one fails61. | |
| Working offline | Yes. | No. | No. It needs the internet. |
| Upkeep | Real and ongoing: updates, backups, testing new models, fixing breaks. We estimate 2 to 6 hours a month. | No hardware to look after. Workflows still need testing when models change. A router you run yourself, like LiteLLM, adds upkeep64. | |
| Who fixes it when it breaks | You, your IT person, or a service you hire. | The host or the router company. You wait, or the router moves you to another host. | The vendor. You wait. |
| Vendor changes | You decide when anything changes. A model you downloaded keeps working, though updates to other software can still break things. |
The short answer
Frontier AI fits when you want the best answers on hard work, the newest features, and nothing to maintain.
Local AI fits when the data has to stay in your building, when you run the same kind of job all day, or when it has to work without the internet.
Most businesses do best with a mix. Private, everyday work goes to local. The hardest jobs go to frontier.
Other cloud models and routers sit in between. GLM, DeepSeek, Kimi, Qwen and others cost little per use, and a router lets one account reach all of them. Your data goes to more companies, some of them overseas.
What it costs
Prices change
Prices change often, sometimes from one month to the next. We checked every price in this guide on November 17, 2026. The small numbers point to the sources at the end.
Drag the sliders. Each train stops at what that way of using AI might cost a month.
Frontier AI business seats
$100–$125 a month
Frontier AI pay per use
$18.90 a month
In between pay per use
$2.52–$10.50 a month
Local AI your own machine
$252 a month
For scale: handling one request at a time, NVIDIA's DGX Spark wrote about 55 tokens a second27. Through 8-hour workdays, this office would need about 2 a second. We haven't tested whether one machine keeps up.
Our math
5 people × 20 uses × 21 workdays = 2,100 requests. Each sends about 2,000 tokens and gets back about 500. That's 4.2 million tokens in and 1.05 million out.
Business seats: $20–$25 per person a month, Standard seats1,3.
Frontier pay per use: an everyday model (Sonnet 5.5 or GPT-6.1 Sol), $2 per million tokens in and $10 out1,66.
In between: DeepSeek V4.1 Flash at $0.30 / $1.2044, up to GLM-5.3 at $1.40 / $4.4046.
Local: a $3,449 small box spread over 3 years is $96 a month. Add $6 of electricity and 3 hours of upkeep at $50 ($150)15,18,20,22.
Our estimate, from prices checked November 17, 2026. A real bill depends on your tools and how you work.
Using both
We split work by task, not by tool. For each job, we ask four questions:
Pick one job you have in mind, then answer each question.
Is it private? Client records, health information, finances, HR files.
Use local, or a business plan with the right agreement.
Is it hard? Tricky reasoning, long multi-step work, careful analysis.
Use frontier. Take out names and account numbers first if you can.
Is it the same job over and over? Sorting email, tagging documents, first-draft summaries.
A local model or a low-cost model handles this well.
Does it need the newest features or the internet? Live web search, voice, making images.
Use frontier. If it has to work when the internet is down, use local.
Local AI
Local AI fits when the data has to stay in your building, when you run the same kind of job all day, or when it has to work without the internet.
Frontier AI
Frontier AI fits when you want the best answers on hard work, the newest features, and nothing to maintain.
A mix
Most businesses do best with a mix. Private, everyday work goes to local. The hardest jobs go to frontier.
Frontier AI
If you're starting from zero, a frontier business plan is the cheaper first step. It costs about $20 to $25 per person a month and needs no upkeep1,3. Local makes sense when privacy rules or steady, heavy use make it worth the work.
Where we'd run it
| Task | Where we'd run it |
|---|---|
| Summarizing your own client files | Local |
| Drafting replies based on your past emails | Local |
| Sorting and tagging incoming documents | Local, or a low-cost model |
| Thinking through a hard business problem, with names removed | Frontier |
| Writing website or marketing copy | Frontier |
| Research that needs current web results | Frontier |
| Anything with patient or financial records | Local, or a business plan with a signed agreement |
Where we fit
Our Managed AI service sets up open-source AI on equipment and accounts you own, and takes care of the upkeep described above. We also help teams use frontier tools like Claude, ChatGPT and Gemini well. If you'd like a second opinion on the mix that fits your business, we're glad to talk it through.
Sources
All checked November 17, 2026.
- 1
- 3
- 6
Google, Gemini Developer API pricing (page updated October 15, 2026)
- 15
- 18
- 20
- 22
- 24
Epoch AI, "Open models lag state-of-the-art closed models by 4 months" (May 29, 2026)
- 25
- 26
- 27
- 29
- 30
- 32
Anthropic, Updates to Consumer Terms and Privacy Policy (August 28, 2025)
- 33
Anthropic Privacy Center, Is my data used for model training? (commercial products)
- 34
Anthropic Privacy Center, How long do you store my organization's data?
- 35
- 36
OpenAI Help Center, How your data is used to improve model performance
- 37
Google Workspace, Generative AI in Google Workspace Privacy Hub
- 38
- 44
- 45
- 46
- 47
- 48
Z.ai, Privacy policy and API data processing terms (updated September 29, 2025)
- 50
- 53
- 54
- 61
- 64
- 66
- 75