Is ChatGPT Plus Worth It? I Was Paying $100/Month for AI — Here's What I Built Instead

Is ChatGPT Plus worth it? Honest answer: it depends on whether you're paying for it once — or paying for it plus Claude Pro, plus Perplexity, plus Gemini Advanced. For most people running that kind of stack: no, not really. You're not buying four different products. You're buying the same capability four times in four different interfaces.

I was spending just over $100/month on AI subscriptions by the time I sat down and actually counted. ChatGPT Plus. Claude Pro. Perplexity Pro. A Gemini Advanced trial I forgot to cancel. I wasn't being reckless — each one had a reason. But when I looked at what I was actually doing with each of them, most of it was the same three tasks shuffled across different tabs.

This post does one thing: it breaks down whether ChatGPT Plus is worth keeping, what the real cost of subscription stacking looks like with actual numbers, and what I built instead. I'm not a developer. If I figured this out, you can too.

---

Section 1: The Honest ChatGPT Plus Review — What $20/Month Actually Gets You

ChatGPT Plus is a genuinely good product. GPT-5.6 Luna — the model that runs on Plus as of August 2026 — is fast, capable, and handles most everyday tasks well: drafting, summarizing, brainstorming, explaining things. For a casual user doing one or two AI tasks a day, $20/month is a reasonable price.

That said, here's what Plus actually includes and where it caps out:

What you get: Access to GPT-5.6 Luna, the "Think" button (explicit reasoning mode), image generation, web browsing, and the GPT store. Good feature set. Clean interface.

Where it thins out: The "Think" mode burns through usage faster because the model is doing more work internally. During peak hours, Plus users regularly hit slowdowns — the model responds, just slower. If you're a heavy user, you'll feel the message caps.

What you don't get without upgrading: For extended reasoning with the Sol or Astra modes, you'd need ChatGPT Pro at $200/month. That's a steep ladder. Most people don't need it — but OpenAI is clearly building toward a tiered model where Plus is the floor, not the ceiling.

For someone who wants one interface, doesn't need an API, and uses AI casually: Plus is fine. The problem isn't the product. The problem is what happens when you also have three other subscriptions running alongside it.

---

Section 2: The Subscription Stack Problem (And Why It Happens)

Nobody sets out to spend $80/month on AI. It accumulates.

You start with ChatGPT because it's what everyone's heard of. Then you try Claude for a long document and it's noticeably better at following complex instructions, so you upgrade to Pro. Then someone recommends Perplexity for research because it cites sources, and you realize it's solving a specific problem the other two don't handle cleanly. Then Gemini Advanced shows up in your Google Workspace and you start using it there.

Each decision made sense at the time. But here's what people's actual stacks look like in September 2026 — this isn't fabricated, it's exactly what's showing up in threads where people are comparing notes:

Stack	Tools	Monthly Cost
Stack A	ChatGPT Plus + Claude Pro + Perplexity	$60/mo
Stack B	ChatGPT Plus + Claude Pro + Perplexity + Gemini Advanced	$80/mo
Stack C	Stack B + Copilot Pro	$100+/mo

The uncomfortable truth: these tools have genuinely different strengths. ChatGPT has the widest plugin ecosystem. Claude is better at long-context reasoning and following nuanced instructions. Perplexity is the right tool when you need cited sources. People aren't being irrational — they've correctly identified that models have different capabilities.

But there's a hidden cost nobody talks about: the cognitive overhead. Four interfaces. Four login sessions. Four sets of conversation histories that don't know about each other. You spend real mental energy just remembering where you left things.

The question isn't whether each tool is good. The question is whether paying subscription prices for all of them makes sense when there's a cheaper alternative that gives you most of the same output.

---

Section 3: The Cost Breakdown — Subscriptions vs. Pay-As-You-Go API

Here's where the math gets interesting.

Every one of these AI products also sells direct API access — meaning you can use the same underlying models without a subscription, paying only for what you actually use. The unit is tokens (roughly, a token is about ¾ of a word — "API" is a term for the direct programmatic connection to the model, not a consumer interface).

What you're paying for subscriptions:

Subscription	Monthly Cost	What You Get
ChatGPT Plus (GPT-5.6 Luna)	$20/mo	Capped usage, no API access
Claude Pro (Sonnet 4.x)	$20/mo	Capped, Projects, long context
Perplexity Pro	$20/mo	Web search + AI citations
Gemini Advanced	$20/mo	Google Workspace integration
Total	$80/mo	4 interfaces, 4 logins, capped use

What those same models cost at the API level (per 1 million tokens, approximate September 2026 pricing):

Model	Input Cost	Output Cost	Estimated Heavy Monthly Use
DeepSeek V4.1-Flash	$0.22	$0.66	~$2–5/mo
DeepSeek V4.1-Pro	~$1.50	~$5.00	~$8–15/mo
Mistral Small	$0.10	$0.30	~$1–3/mo
GPT-5.6 Luna (API)	$0.20	$0.80	~$5–12/mo
Claude Sonnet 4.x (API)	~$3.00	~$15.00	~$10–25/mo

DeepSeek V4.1-Flash launched September 10th — nine days ago as I'm writing this. It's 90–95% cheaper than Claude at the API level and performs comparably for the majority of writing and reasoning tasks. That's not a rounding error. That's a structural cost difference.

The smarter approach — and what I actually do now — is task routing. Use a cheap, fast model (DeepSeek Flash, Mistral Small) for 80% of tasks: summarizing, drafting, quick Q&A. Reserve a stronger, more expensive model (Claude Sonnet, GPT-5.6 Luna) for the 20% of tasks that genuinely need deeper reasoning. On a real month of heavy use, that approach costs me $10–20 in API fees. Compared to $80 in subscriptions.

---

Section 4: The "Can I Actually Do This?" Section

The honest answer to "can a non-developer do this?" is yes — with one important caveat: you have to accept a brief period of unfamiliarity before it becomes normal. It's like switching from Gmail to Outlook. Not hard, just different enough to feel awkward for a week.

Here's what it actually takes:

Hardware: A Mac (or PC) that's always on. I run mine on a Mac Mini M4 — about $600 one-time. It's on 24/7, handles 5–10 models depending on RAM, and I've paid nothing in monthly hardware costs since I set it up. That $600 breaks even versus $80/month subscriptions in 7.5 months.

Local models (zero ongoing cost): Ollama is a free tool that runs AI models directly on your Mac. No API key. No monthly bill. You can run Llama 4, Mistral, or Gemma locally for nothing after the hardware cost. For lighter tasks — summarizing an article, drafting a quick email — a local model is completely sufficient.

Cloud models without a subscription: For tasks that genuinely need more horsepower, you grab an API key from DeepSeek or Mistral directly. It takes under 10 minutes and you're paying pennies per task rather than $20/month for a subscription ceiling you'll hit by Tuesday.

What you don't need: A background in programming. An understanding of how large language models work internally. A server. A technical cofounder.

The "hard part" is interface-level, not technical. You're learning where to type and what to expect. Give it a week.

---

Section 5: What I Built Instead

About six months into my subscription stack experiment, I started consolidating. I cancelled Gemini Advanced first — I was using it almost entirely for tasks Claude handled better. Then Perplexity, because I realized I only used it for research citations, and I could get the same result with a web-connected model query. Then I looked at ChatGPT Plus and Claude Pro sitting side by side and made a choice: I kept Claude (the one I actually reached for more) and went API-only for GPT access.

Then I went one step further. I set up a personal AI agent on a Mac Mini at home — connected to Slack, always running, routing tasks to whichever model makes sense for the job. Writing tasks go to a mid-tier model. Complex reasoning goes to Claude. Quick lookups go to a local Mistral instance running on the hardware for free. The whole thing costs me about $12/month in API fees.

I started writing down everything I figured out at MyAIAgentOS.com — it's basically the guide I wish had existed when I started. The model routing piece in particular is what made the cost difference click for me.

The decision flowchart below is roughly how I'd help someone else think through it:

graph TD
    A[Do you currently pay for AI subscriptions?] -->|Yes| B[How many?]
    A -->|No| Z[ChatGPT Plus alone: probably fine to start]
    B -->|Just one| C[Do you use it daily?]
    B -->|Two or more| D[You're almost certainly paying for overlap]
    C -->|Yes| E[Keep it — $20/mo for daily use is reasonable]
    C -->|No, maybe weekly| F[Consider API-only: cheaper for light use]
    D --> G[Log what you actually use each tool for this week]
    G --> H{Is 80% going to one model?}
    H -->|Yes| I[Cancel the others — route that 20% to cheap API]
    H -->|No, it's genuinely split| J[Consider a single agent that routes across models]
    I --> K[Estimated savings: $40–60/mo]
    J --> K

---

FAQ

Is ChatGPT Plus worth it in 2026?

For casual users doing one or two AI tasks a day who want a single, polished interface: yes. ChatGPT Plus at $20/month is a reasonable price for what GPT-5.6 Luna delivers. For power users already paying for Claude Pro and Perplexity, probably not — you're paying for capability overlap, not genuine additions to your toolkit. The question to ask yourself: if you could only keep one subscription, which one would you keep? That's the answer.

How much are people spending on AI subscriptions per month?

Based on what's showing up in threads and discussions in September 2026, common stacks run $60–80/month. The most frequent combination is ChatGPT Plus + Claude Pro + Perplexity at $60/month. Power users who've added Gemini Advanced and Copilot Pro are regularly crossing $100/month. Most didn't plan to spend that much — it accumulated one "I'll just try it" upgrade at a time.

What's cheaper than ChatGPT Plus that actually works?

DeepSeek V4.1-Flash — launched September 10, 2026 — is 90–95% cheaper at the API level and performs comparably to much more expensive models for the majority of writing, summarizing, and reasoning tasks. Mistral Small is also a strong budget pick, particularly for European users who prefer a non-US provider. Neither requires a subscription; you pay only for what you use.

Can a non-developer run their own AI agent?

Yes. Tools like Ollama for local models and any agent interface that accepts API keys require no coding knowledge and no technical background. The typical setup takes under an hour. The learning curve is interface familiarity — learning where to type and how to phrase requests for a different tool — not technical skill. It's comparable to switching between email clients: unfamiliar at first, then completely normal.

Should I cancel my ChatGPT subscription?

If ChatGPT is your only AI tool and you use it daily, keep it — $20/month is reasonable for consistent daily use. If you're stacking it with Claude Pro and Perplexity, you're almost certainly paying for redundant capability across three interfaces. In that case: identify which model you actually reach for first, cancel the others, and route the tasks you were using the cancelled tools for through a single agent with API access to cheap models. The overlap you'll lose is minimal. The savings are $40–60/month.

What is the cheapest way to use GPT-level AI in 2026?

Use DeepSeek V4.1-Flash or Mistral Small via their direct APIs for 80–90% of tasks — these cost fractions of a cent per query. Reserve stronger models like GPT-5.6 Luna or Claude Sonnet 4.x for the tasks that genuinely require their reasoning depth. If you have a Mac or PC that stays on, run Ollama locally for zero-cost inference on light tasks. Total monthly API cost for a heavy user doing this right: $5–20, versus $60–100 in subscriptions.

---

The Decision You Can Make This Week

Here's a practical framework, not a pitch. If you're paying for more than one AI subscription, spend 15 minutes this week logging what you actually used each one for over the past month. Be specific. You'll almost certainly find that 80% of your usage went to one model — and the remaining 20% could be covered by a $2/month API bill from DeepSeek or Mistral.

If you want a concrete path for setting that up — hardware, model selection, routing logic, all of it — I've documented the whole process at MyAIAgentOS.com. Not because it's complicated, but because having a map when you're doing something unfamiliar makes you 10x more likely to actually finish it.

Cancel one subscription this month. See if you miss it. You probably won't.

---

[INTERNAL LINK: DeepSeek V4.1 deep-dive post TK — link here once Harriet's guide is published] [INTERNAL LINK: Local AI setup / Ollama guide TK — link here once available]