Table of Contents

• How you really pay
• 2026 ranges (approx., and they move)

  • ChatGPT (OpenAI)
  • Claude (Anthropic)
  • Gemini (Google)
  • Open and cheap-per-token models
• The cost that doesn't show on the landing page
• Three scenarios

  • 1. You write 10 pieces a week
  • 2. You code every day
  • 3. Team of 4
• When 20 dollars are enough
• How to structure the spend
• Frequently asked questions

  • Can you do serious work on the free plan?
  • Plus, Claude Pro or Gemini at 20?
  • Isn't the API always cheaper?
  • Does Ollama save me the 20 dollars?

How you really pay

There are three rails. Mixing them is normal. Confusing them is expensive.

Chat subscription. You pay a fixed fee and use the app. The ceiling isn't "unlimited": it's a quota that refills every few hours and, sometimes, a weekly one too. If you go over, the model steps down or you get pushed to the plan above.

API (pay per token). You pay for what you consume. A token is a piece of text: roughly 750 words in English is, with slack, a thousand tokens. It comes out cheap with short prompts and expensive if an agent reads the repo ten times a day. It works when you already know what you'll repeat.

Local. Ollama and similar charge no license. The cost is hardware, electricity and your setup time. The data never leaves your machine. In exchange, don't expect the frontier model unless you have a serious GPU.

A 20-dollar Plus isn't "the API with a discount": it's a chat with a ceiling. An API that costs 8 one month and 80 the next isn't "a cheap plan": it's a variable invoice. Ollama isn't free ChatGPT: it's another category.

2026 ranges (approx., and they move)

All in list USD. In LATAM add bank or fintech conversion, possible VAT, and the plan not being equally available in your country.

ChatGPT (OpenAI)

• Free: for trying things and short tasks. You'll hit the ceiling.
• Go: around 8 dollars a month.
• Plus: 20 dollars. The default for anyone working with text every day.
• Pro: 100 to 200 dollars, depending on the usage multiple (in practice, 5x or 20x versus Plus).
• Business: around 25 dollars per seat (sometimes less on annual). Central admin and invoicing.

The OpenAI API is a separate account. Plus doesn't authorize you to plug it into your systems as if it were unlimited. If you automate, quote tokens separately.

Claude (Anthropic)

• Free: comparing tone and light work.
• Pro: 20 dollars. The cultural equivalent of Plus.
• Max: 100 dollars (5x) or 200 (20x). It shows up when you code or paste long contexts every day and Pro runs out mid-week.
• Team: around 25 dollars per seat. There are pricier variants if the team needs more usage or code tools. Not all "team" seats include the same.

Claude feels expensive as an all-day coding agent. It feels cheap if you save it for the work that needs judgment.

Gemini (Google)

• Free: bundled with the Google account. Enough if you already live in Docs and Gmail.
• AI Plus: roughly 5 to 8 dollars, depending on timing and country. It usually includes extra Google One storage.
• AI Pro: around 20 dollars.
• Ultra: it has fluctuated; take it at 200 to 250 dollars (sometimes a tier near 100 appears). Confirm with Google.

If you already pay for One or Workspace, check whether the AI is included before opening another card.

Open and cheap-per-token models

• Ollama local: 0 in licensing. On a 16 GB laptop you run small and mid-size models, not the frontier one.
• APIs from DeepSeek, Qwen, GLM and the like: far cheaper per token than OpenAI or Anthropic. Orders of magnitude, not 10% less. Drafts, classification, routine code, volume.
• OpenCode Go: around 10 dollars a month (the first month sometimes comes out less). It's not a chat for the client: it's a model plan for agent-based coding, with usage ceilings. The product breakdown is in OpenCode Go: price, limits and whether it's worth it.

"Cheap per token" isn't free. It's that the million tokens that hurt on a frontier model, here sometimes you don't even notice.

The cost that doesn't show on the landing page

Limits. The 20-dollar plans run out. They refill every few hours, but if you work in blocks —a morning of code or ten pieces in one day— you'll see the warning. The 100 or 200 one isn't smarter: it's more quota.

Time. Waiting for the reset, rewriting because the free model got lazy, or copying the thread to another tool are hours. At freelance rates, one lost hour already paid the difference between Free and Plus.

Lock-in. Custom GPTs, Projects, Gems and long histories live with one provider. Switching hurts more than the 20 dollars. Don't keep the only copy of the briefs and prompts you work with there.

Quality. The cheap model saves you money and charges you in review. If you publish without editing, the savings are fictional.

Paying in dollars from LATAM. Bank surcharge, VAT where applicable, and the "payment failed" mid-month. A 20-dollar plan can end up closer to 25–30 in local currency.

One account, five people. Sharing Plus breaks terms, mixes client chats and leaves no one accountable when someone pastes too much. It's cheap until the first scare.

Three scenarios

The tallies are order-of-magnitude. They serve to locate the tier, not to quote.

1. You write 10 pieces a week

Forty pieces a month. If each takes 4 to 8 back-and-forths (brief, outline, draft, two trims), you're at 160 to 320 messages, mostly short.

• Free or Go (~0–8 USD) works if the pieces are short, you don't paste huge PDFs and you accept waiting.
• Plus / Pro at 20 (ChatGPT, Claude or Gemini) is the sweet spot: better model, less friction, files. For a solo writer or marketer, this is the reasonable ceiling.
• Cheap API (DeepSeek/Qwen/GLM): if you already have a template in / draft out flow, the variable drops to a few dollars —sometimes cents— and you keep the 20 chat for fine editing.
• Local: rewriting and summarizing without sending the client's brief to the cloud. Don't expect it to build the piece better than the 20 one.

Here 20 dollars are enough. Going up to 100 only makes sense with very long documents every day, heavy research or image/video at volume (that last one lives in other quotas).

2. You code every day

This is where the 20 plan breaks. An agent that reads the repo, runs tests and iterates doesn't chat: it eats context. An afternoon of refactoring can burn what a writer burns in two weeks.

• A single Plus/Pro at 20 works if you code with the AI, not through it: one question, one file, one test. If the agent is on all day, you come up short mid-month. It's not lack of discipline; it's the plan's design.
• Max / Pro at 100–200 makes sense if your billable hour is high and the bottleneck is quota, not your judgment. Before jumping, count how many days a month the limit actually stopped you.
• Cheap mix + 20: open API or OpenCode Go (~10 USD) for routine code, and a Claude or ChatGPT at 20 for architecture and PRs you can't send blind. It's the logic of diversifying AI providers: don't burn the expensive model on autocomplete.
• Frontier API all day: it can cost more than Max. Set a spend alert on day one, not when the 180-dollar email arrives.
• Local: good for completion and for code that can't leave. Weak as the only brain on an office laptop with a big repo.

20 dollars often aren't enough if the AI is your work environment. They are if it's a consultant you call, not the IDE.

3. Team of 4

Four Plus accounts on personal emails are 80 dollars plus chaos: nobody knows what was pasted, there's no invoice in the company's name, and the day someone leaves they take the history with them.

• 4 × 20 personal: the number looks fine. The risk doesn't.
• Business / Team ~25 USD per seat: around 100 dollars a month. You pay for admin, access offboarding and, generally, different rules about training on your data. Confirm it in the plan's contract, not in a thread.
• 2 seats at 20–25 + cheap APIs + one local: works when not all four do the same job.
• One shared Pro/Max at 100–200: bad idea. Quota, confidentiality and accountability step on each other.

The jump that matters isn't from 20 to 25 per head. It's moving from personal accounts to a space the company can turn off.

When 20 dollars are enough

They're enough if you're one person, the work fits in human-sized chats (text, email, research, file-by-file code) and you don't run out of quota more than a couple of afternoons a month. The upgrade to 100, in that case, is vanity or fear.

They're not enough if you code with an agent several hours a day, paste repos or databases as if context were free, four people share a login, the 20 model forces you to redo the work, or client data can't go into a consumer plan.

You're throwing money away if you pay 100–200 and 80% of your prompts are "rewrite this in a formal tone". A 10-dollar model or a local one does that. The expensive plan is justified by expensive tasks, not by the logo on the tab.

You're also throwing money away if you stack Plus + Claude Pro + Gemini Pro "just in case". One main chat. The second, when the first runs out or is clearly better at your craft. The third, almost never.

How to structure the spend

1. Measure for two weeks on free or on 20. Note the days the limit stopped you and the tasks that genuinely came out better with the expensive model.
2. Separate chat from API. Chat is for thinking. API is for repeating. If you copy the same prompt 40 times, it's already a flow, and flows are paid per cheap token.
3. Cap the API the day you create the key. Without an alert it's not flexible: it's an open credit card.
4. Don't buy hardware to "save the subscription" unless you need it for privacy or volume.
5. Pay with a method you can cut off.

If you code and you're between a Max at 100 and a cheap combo, the layering strategy is in why you shouldn't depend on a single provider.

Frequently asked questions

Can you do serious work on the free plan?

You can start. You can't depend on it. If the AI saves you an hour a week, the 20 plan has already paid for itself at any reasonable LATAM freelance rate.

Plus, Claude Pro or Gemini at 20?

The same cost tier. Choose the chat you already use, the one that makes you rewrite least, and where your files are. If in two weeks you notice no difference, stay with the one you pay for most easily.

Isn't the API always cheaper?

No. It's cheaper on repeated, short tasks. It's more expensive when an agent loops with long context. The subscription is a buffet with a bouncer. The API is à la carte.

Does Ollama save me the 20 dollars?

Yes, if the hardware is enough and the data can't leave. No, if your job is "get it right the first time" and your machine has 8 GB. The time to set it up counts too.

---

The real price isn't the number on the home page. It's subscription + API + time + staying tied in, in your currency and at your hourly rate. Start at 20 if the free tier holds you back. Move up to 100 when you can point to the days the limit cost you money. If a single provider runs out on you, sometimes the fix is spreading the load or a cheap plan of open models for code.