A chat AI subscription is sold as a single monthly number, but the bill a reader actually pays can sit far from that headline. This page maps subscription tiers to cost ranges and usage intensity, explains why the same tier diverges by region, and shows the thinking behind the point where running a model locally becomes the cheaper option.
The figures on this page are ranges with a verification date of 2026-09-17. Prices and tiers change, and regional adjustments move them further, so the official pricing page of any service is the only authoritative source. No specific vendor price is stated here.
Subscription tiers by cost range and usage
The table below groups hosted chat plans into tiers by their monthly cost range and the usage intensity each tier targets. It is intentionally generic: naming one vendor’s price would be outdated and would not hold across regions. The point is the shape of the market, not a quote.
| Tier | Monthly cost range | Fits which usage intensity | Note on regional spread |
|---|---|---|---|
| Free tier | No monthly charge | Light, occasional, trial use | Same everywhere in money terms, but feature limits vary by region |
| Entry paid tier | Low single-digit to low-tens of currency units per month | Regular personal use, one person | Local taxes and currency push the real number around the list price |
| Heavy personal tier | Roughly tens of currency units per month | Daily power user, long sessions, larger context | Annual billing often lowers the effective monthly rate versus month-to-month |
| Team or business tier | Per-seat, scales with headcount | Small group or company use | Seat count and local business tax make totals vary widely |
| Enterprise or usage-based tier | Custom or metered | High or unpredictable volume | Quoted case by case; no single public number applies |
The ranges are broad on purpose. A reader should treat them as a map, then confirm the exact figure on the vendor’s page for their own country.
Why the same tier costs different amounts by region
A list price is a starting point, not a final bill. Several forces pull the real monthly cost away from it, and most of them are regional.
Local taxes are the first force. Value-added tax, goods and services tax, and similar levies are applied differently by country, so the same plan lands at different totals. Currency exchange is the second: a price set in one currency is converted at a rate that drifts, and some vendors set separate local prices rather than a live conversion, which can be higher or lower than the raw exchange would suggest. Regional pricing strategy is the third: vendors sometimes set lower prices in markets with lower willingness or ability to pay, so two readers see two different numbers for the same tier.
Billing method adds more spread. Annual plans usually cost less per month than month-to-month when spread over twelve months, so the same tier shows two effective prices. Team plans multiply by seat count, and business tax treatment changes the total again. Some tiers carry a minimum spend, so a light team still pays a floor amount. None of these appear in the headline price.
The practical rule: take any quoted monthly number as the number for one specific country and one specific billing choice, then verify for your own situation.
Subscription versus usage-based billing
Hosted AI is sold in two shapes, and they behave differently as usage changes.
A subscription is a capped cost. The reader pays the same whether the month is quiet or busy, up to whatever the plan’s limits allow. The advantage is predictability: the worst-case bill is known on the first of the month. The disadvantage is paying for capacity that goes unused in light months.
Usage-based billing is linear. Cost tracks tokens or requests, so a quiet month is nearly free and a heavy month is proportionally expensive. The advantage is paying only for what is used; the disadvantage is an open-ended ceiling that depends on behavior the reader does not always control, such as an automated job that loops.
These two shapes cross at a break-even point. Below it, a subscription is cheaper because the fixed fee is less than the metered total. Above it, metered billing or a higher tier starts to dominate. The crossing moves with the subscription price, the per-token rate, and the reader’s own volume, which is why a fixed cutoff cannot be quoted.
When a local model becomes the cheaper path
Running a model on your own hardware trades the subscription for electricity and hardware amortization. The question is not “is local free” but “does the subscription I would otherwise buy cost more than my local running cost at my usage.”
The comparison needs three inputs. The first is the monthly token volume the reader actually generates, which the token cost calculator helps estimate from a typical workload. The second is the subscription tier that volume would require on a hosted service, which sets the baseline the local option must beat. The third is the local cost: the power draw of the machine while the model runs, spread over the months of hardware life, plus the fact that the hardware may already be owned for other work.
The LLM API cost comparison page lays the hosted side next to the local side in more detail. The break-even thinking is the same there: at low volume a hosted plan wins on convenience and zero setup, while at high and steady volume the repeated subscription payments can outrun the amortized cost of hardware the reader already runs.
A useful way to frame it: a reader who uses a model a few times a week rarely reaches the volume where local pays off, because the subscription stays cheap and the hardware would sit idle. A reader whose machine generates large token counts every day, or who runs many automated tasks, can cross into the range where the recurring subscription is the larger line item. The crossing is personal, not universal.
Cost is not the only variable
A purely monetary comparison misses the reasons readers pick one side.
Privacy is the first non-price factor. A hosted service sends prompts to a vendor’s systems, and depending on plan and policy those prompts may be used for training or retained. A local model keeps the data on the reader’s machine, which matters for sensitive work regardless of price. This is a qualitative difference, not a number that offsets a bill.
Rate limits are the second. Hosted tiers cap how fast and how much a reader can send, and heavy use can hit those caps. A local model is bounded by the reader’s own hardware, not by a vendor’s policy, so throughput is predictable in a different way. The batch size and throughput VRAM page shows how local concurrency translates into memory, which is the local side of the same limit question.
Availability is the third. A hosted service depends on the vendor’s uptime and on the reader’s connection. A local model depends on the reader’s own machine staying up. Neither is universally better; they fail for different reasons.
These factors do not appear on a pricing page, yet they often decide the choice when the dollar amounts are close.
Estimating your own break-even
- Note the tier you would buy if you stayed hosted, and its real local monthly price after taxes and billing method.
- Estimate your monthly token volume from a normal week of use.
- Compare that subscription cost against the amortized local running cost for the same volume.
- Add the non-price factors: privacy needs, rate limits, and whether the hardware is already owned.
- Re-check whenever the vendor changes price or your usage pattern shifts.
The run LLM locally page is the entry point for the local side of that decision, and the local LLM hardware guide frames what hardware a given workload needs before any cost math makes sense.
A note on verification
Every number in the tier table is a range, not a quote, and carries the verification date 2026-09-17. A reader acting on this page should open the official pricing page for their country and plan before committing, because the tiers and their prices move. The model file size lookup and the VRAM calculator then help size the local option the comparison points toward.
What this page does not tell you
- This page gives no real-time prices. Figures change by vendor, plan, and region, so any stated price would be wrong soon after publication; the official pricing page is the only live source.
- This page does not compare model capability or quality between hosted and local options. Which side produces better answers is a separate question from what each costs.
- This page does not make team or compliance judgments. Whether a company may send data to a hosted service, or must keep it local, depends on policy and law that this page does not assess.