Since I was really upset about choosing between subscription plans before, why not leave some of my experiences here so others can possibly refer to them?
There’s a certain absurdity in paying $20 a month for something you can’t hold, can’t own, and can’t fully predict. But here we are — the AI subscription economy has arrived, and whether you like it or not, the question is no longer should I subscribe, but which one, and what do I actually get.
I’ve used all of these. Not as a benchmark tester or a tech journalist, but as someone who actually needs to code, write, research, and think with these tools daily. These are my honest field notes.
The $20 Tier
Claude Pro — The one with a soul
$20/month · claude.ai
If you’ve talked to Claude long enough, you’ll understand what I mean when I say it feels like talking to someone. Not a product. Not a search engine wearing a chat interface. Something that actually reads what you wrote and responds to it.
When reasoning is set to high or even higher, the output quality is genuinely remarkable — the kind of writing that makes you pause and re-read. This is where Claude is unmatched.
The problem is quota. The $20 plan runs lean, and it’s shared across everything — claude.ai conversations, Claude Code, Claude Cowork, and Claude Design, all eating from the same pool. You’re looking at roughly 50 minutes to an hour of Claude Code before you hit the wall, assuming you’re not spawning multiple subagents, which burns through usage fast.
The Opus model adds another layer of complexity. It’s more expensive per token, but arguably more token-efficient — it reasons faster and solves harder problems with less back-and-forth. Whether using Opus or Sonnet saves you more quota depends entirely on your task. For quick, focused problems, Opus often wins. For sprawling multi-file refactors, Sonnet is safer on the balance sheet. (Conventional wisdom leans toward Sonnet as your default to conserve quota, reserving Opus for genuinely hard reasoning.)
For writing and light cognitive work on claude.ai, $20 goes much further. The quota crunch hits hardest when you’re running agentic coding sessions.
Best for: Writers, people who value conversation quality, developers doing focused, bounded coding work.
ChatGPT Plus — The generous one
$20/month · chatgpt.com/pricing
ChatGPT’s quota generosity on the chat side is almost suspicious. I’ve had marathon conversations without hitting a wall. The Plus plan gets “unlimited*” messages and interactions with fast response times — a meaningful jump from Go, which is throttled by bandwidth and availability.
Here’s what that $20 actually unlocks beyond the Go tier: GPT-5.5 Thinking (the reasoning model, not just Instant), image generation with Thinking, interactive tables and charts, scheduled Tasks, Developer Mode (beta), Deep Research and agent mode, expanded memory with past chats, and a 256K total context window for reasoning — roughly 320 pages of text input. You also get proper Codex access (Go only gives “Limited”), legacy model access, and early access to new features.
Codex sessions still run into the ~50-minute wall, though. The community approach is to configure lighter mini models as subagents — a reasonable workaround, but still a workaround.
GPT-5.5’s coding personality is worth mentioning: it’s excellent at diagnosing and squashing bugs. It can be overzealous about minor issues — occasionally treating a style preference like an emergency — but for backend logic and debugging, it’s sharp. Frontend UI generation is a weaker suit, where Claude tends to outperform it.
One thing worth knowing: the Go plan at $8 may include ads — something OpenAI buries in the fine print. If you’re weighing Go against Plus, that alone might change the calculus.
Best for: Image-heavy workflows, chat-heavy use cases, backend developers who need a reliable bug-hunter.
Google AI Pro — The ecosystem play
$19.99/month · Google AI Plans
I’ll be honest: I haven’t used Gemini Pro deeply for a while. But what I remember is this — Gemini’s multimodal capability is genuinely the most native of any platform here. Text, images, audio, video — in and out, fluidly, without the duct-tape feeling of bolt-on features. If your work lives across modalities, Gemini is built for that from the ground up.
The Pro plan gives you 4x higher usage limits in Gemini compared to the free tier (Plus only gets 2x), access to Gemini’s Pro-tier model, 5 TB of Google storage, and YouTube Premium Lite thrown in. That last point sounds like marketing — but if you use YouTube at all, it’s a real perk.
The real value proposition might actually be the Google ecosystem itself. Gmail, Drive, Docs, Meet — if that’s your daily environment, having Gemini woven into it amplifies the experience in ways that a standalone chat app simply cannot. NotebookLM, included with the subscription, deserves its own mention: for research, lecture processing, and learning from uploaded sources, it’s phenomenal. Nothing quite like it. (Worth noting it’s also bundled into the cheaper $4.99 Plus tier, just with lower limits.)
For coding? I wouldn’t choose Gemini. Its strength is lateral and multimodal, not in grinding through a codebase. But for Frontend UI generation, there’s a surprisingly authentic visual sense to its outputs — distinctive in a way that feels less template-like than GPT.
Best for: People living in Google’s ecosystem, multimodal workflows, serious research with NotebookLM.
Ollama Cloud — The open-model gamble
$20/month (or $200/year) · ollama.com/pricing
If you’re willing to bet on open models, Ollama Cloud Pro is the most quota-generous option in this tier. I’ve coded for hours with Kimi K2.6 without once thinking about limits.
Here’s what makes Ollama’s metering fundamentally different from everyone else: usage is measured in GPU time, not tokens. That means the same plan stretches further with lighter models and shorter prompts. Models are tiered by compute cost — level 1 (light, like gpt-oss:20b) to level 4 (heavy, like deepseek-v4-pro). Your usage resets on two cycles: session limits every 5 hours and weekly limits every 7 days. Pro gives you 50x more cloud usage than the free tier and lets you run 3 cloud models concurrently. Prompts and responses are never logged or trained on — a policy Ollama requires from its hosting partners too.
What makes it genuinely different from the others is where you can spend that quota. Ollama Cloud uses an OpenAI- and Anthropic-compatible API, which means your subscription works across a huge range of tools — not just the Ollama CLI. From the official integrations docs:
Coding Agents — Claude Code, Codex, OpenCode, Droid, Goose, Pi Assistants — OpenClaw (Hermes Agent works too via the OpenAI-compatible endpoint) IDEs & Editors — VS Code (Copilot-compatible), Cline, Roo Code, JetBrains, Xcode, Zed Chat & RAG — Onyx Automation — n8n Notebooks — marimo
The magic is the ollama launch command (v0.15+). Running ollama launch claude or ollama launch codex sets up environment variables, discovers available local and cloud models, and starts the agent — no manual configuration needed. For Claude Code specifically, Ollama’s Anthropic-compatible API layer (shipped in v0.14) means it just works: the agent thinks it’s talking to Anthropic’s API, but the inference runs on your Ollama Cloud quota. The models the docs recommend for coding are kimi-k2.5:cloud, glm-5:cloud, minimax-m2.7:cloud, and qwen3.5:cloud — though newer entries like minimax-m3:cloud and kimi-k2.7-code:cloud have since landed in the catalog (check ollama.com/search?c=cloud for the exact current tags). Claude Code even supports web search through Ollama’s web search API when configured.
As for the current cloud model lineup — it’s grown fast. The catalog now spans GLM (5, 5.1), Kimi (K2.5, K2.6), MiniMax (M2.5, M2.7, M3), DeepSeek (V4 Flash, V4 Pro), Qwen (3.5, 3.6), and NVIDIA Nemotron 3 (Nano/Super/Ultra), alongside Western open models like Gemma 4, Devstral, Ministral, and gpt-oss. Cloud models behave exactly like local ones — you can ls, run, pull, and cp them normally. They always run at their full context length too, unlike local models that default to 4K. Several support vision input (Kimi K2.5/K2.6, Qwen3.5, MiniMax M3 — which ships a 1M-token context and native vision), so the multimodal limitation is only partial. The gap is on the output side: image and audio generation isn’t available through Ollama Cloud yet.
One important note: Ollama may deprecate and retire older cloud models as better ones come out. If you’re building something that depends on a specific model, know that you might need to switch. They’ll notify you in advance, but it’s something to plan for — local models are never affected by this.
The honest caveat remains: open models show more variance than proprietary ones. They can solve hard problems, but tend to consume more tokens doing it. When a proprietary model solves something in two passes, an open model might take five — and that compounds across a long session. Check Artificial Analysis before committing to a model; the benchmarks there are the most honest scorecards available.
Best for: Developers who need long uninterrupted agentic sessions across multiple tools, those comfortable navigating model trade-offs, open-source advocates.
The Sub-$10 Tier
OpenCode Go — The best taste for the least money
$5 first month, then $10/month · opencode.ai/go
Five dollars for the first month is an honest offer. You get access to a curated, evolving roster of strong open-source models with rolling limits: roughly $12 worth of usage every 5 hours, $30 weekly, $60 monthly (confirm the live numbers, since they shift). Current models include GLM-5.2, GLM-5.1, Kimi K2.7 Code, Kimi K2.6, MiMo-V2.5-Pro, MiMo-V2.5, Qwen3.7 Max, Qwen3.7 Plus, Qwen3.6 Plus, MiniMax M2.7, MiniMax M3, DeepSeek V4 Pro, and DeepSeek V4 Flash. There’s also Big Pickle, a free model (likely GLM-4.6-based, 200K context) available even without subscribing.
An important detail: since models have different per-request costs, your actual request count varies wildly. MiniMax M3 currently offers a 3x usage multiplier, meaning you can fire off an enormous number of requests. DeepSeek V4 Flash is the cheapest per-request, while GLM-5.2 is the most expensive. The practical result: lean on lighter models for routine tasks and save the heavy hitters for hard problems.
What I didn’t fully appreciate until reading the docs: Go isn’t locked to the OpenCode interface. The endpoint is OpenAI-compatible, so any tool that accepts a custom API base URL can use it — Claude Code, Codex, Hermes Agent, OpenClaw, Pi, your own scripts. The model access is the product, not the UI. Models are hosted across US, EU, and Singapore with zero-retention policies from all providers.
Use /connect in the OpenCode CLI to link your subscription, or just set the base URL and API key manually in whatever agent you prefer. If you exhaust your Go limits mid-session, you can fall back to Zen credits or switch to a backup API provider — OpenCode handles the failover cleanly.
For $10/month as a secondary coding tool, or as your primary if you’re comfortable with open models, it’s a remarkably clean deal. The OpenCode team has been transparent that they roughly break even on the $10 plan — it’s a growth play, not a margin play.
Best for: CLI-native developers, people who want to test agentic coding cheaply, anyone who uses multiple coding agents and wants one subscription to cover them all.
ChatGPT Go ($8) & Google AI Plus ($4.99) — A note
ChatGPT Go at $8/month gives you expanded access to GPT-5.5 Instant, more messages, more uploads, more image creation, and longer memory. It also bumps your context window: 54K for Instant (vs 27K on Free) and 256K for Reasoning. But there are meaningful gaps: no GPT-5.5 Thinking, only limited Deep Research, no Tasks, limited Codex, and response times are still throttled by bandwidth. The ad-supported model is also worth remembering — OpenAI says Go “may include ads.”
Google AI Plus at $4.99/month is the cheapest paid AI subscription out there. You get 2x higher usage limits, Gemini Omni, Daily Brief in Gemini, AI Inbox in Gmail (rolling out), and 400 GB of Google storage. It doesn’t give you Gemini’s Pro model or YouTube Premium — those are reserved for the $19.99 Pro tier. But at five dollars a month with Google ecosystem integration (NotebookLM included), it’s a remarkably low bar to clear for casual use.
The Honest Summary
No single plan wins across all categories. The real question is: what do you actually do with it?
- You write and think for a living → Claude. Accept the quota and work within it.
- You generate images or chat heavily → ChatGPT Plus. The unlimited messages and 256K reasoning context are real.
- You live in Google’s world → Google AI Pro, especially with NotebookLM and the ecosystem integrations.
- You need to code across multiple tools for hours → Ollama Cloud Pro. One subscription powers Claude Code, Codex, OpenCode, and a dozen IDE plugins.
- You want to try agentic coding cheap → OpenCode Go. $5 first month, a dozen-plus models, works with any OpenAI-compatible agent.
- You want to dip a toe in for almost nothing → Google AI Plus at $4.99, or ChatGPT Go at $8 with the ad caveat.
The underlying truth of all of them: there is no free lunch, only different shapes of scarcity. Quota, intelligence, modality, ecosystem — each plan forces a trade. Knowing which trade you’re willing to make is the whole game.
Leave a Reply