A post on V2EX, the Chinese developer forum, argues you can run a serious AI tool stack on $20 a month by sending only about 10% of your work to a paid frontier model. The author, who dropped Claude Pro, published it on 24 October 2026, and the replies were far less kind than the idea deserves.
The core rule is sound. Some of the details are stale or unverifiable. Here is what to keep and what to check.
Key takeaways
- The post’s “90/10” rule: hard architecture and debugging work goes to one paid model, routine edits go to free ones.
- The author says they quit Claude Pro over rate limits, peak-hour quality and account-ban worries. These are their own claims.
- V2EX readers pointed out that the post cites Claude 3.5 as top-tier, which looks out of date in October 2026, and some suspected it was AI-written.
- The tiering idea works regardless. Check your own usage before you cancel anything.
What does the 90/10 rule actually say?
The author, posting as hashtome, says many indie developers suffer “subscription anxiety”: ChatGPT Plus, Claude Pro, Cursor Pro and Midjourney together can cost $100 to $200 a month, while most of that capacity sits idle. Using a top model to translate a few words or rename a variable is, in their words, using a cleaver to kill a chicken.
So they split work in two. About 10% is the hard stuff: multi-system design, tricky algorithm debugging, cross-file refactors and key technology choices. The other 90% is small scripts, single-function edits, JSON or CSV conversion and proofreading.
What is in the $20 stack?
According to the post, the single paid item is a $19.99 monthly plan in Google’s ecosystem. The author claims it bundles Gemini and access to Claude models with a long context window. We could not verify that bundle, so treat it as a claim.
The free tier is a mix of Chinese models (Doubao, Tongyi Qianwen, Kimi), Tencent’s WorkBuddy international edition, and local models through Ollama, such as Qwen2.5-Coder at 7B or 14B parameters. The author also promotes their own open-source browser automation plugin, AgentLimb. That is a self-interest flag worth noting.
Why did they drop Claude Pro?
Three reasons: limits hit after roughly two to three hours of heavy coding, quality allegedly dips at peak hours, and users in China risk bans with slow appeals and no refunds.
The limit complaint is real in structure. Anthropic’s pricing page lists Pro at $20 a month, or $17 a month billed annually. It says usage is measured per 5-hour session, with weekly limits on top, and that other caps may apply. Heavy coders can hit those walls. The peak-hour quality claim has no evidence in the post.
Why did V2EX readers push back?
Mostly on credibility. One reader asked what year the author’s AI was from, since the post calls Claude 3.5 Sonnet and Gemini 1.5 top-tier. Others said the long text read as AI-generated, and one asked if it was a bot. Another suspected a ploy to farm coins or sell invite codes. A separate reader disputed the model ranking itself.
Those are readers’ opinions, not findings. But a cost-saving guide that recommends outdated models undermines its own advice. Read the thread at V2EX and judge for yourself.
How do the common plans compare?
| Plan | Monthly price | Notes |
|---|---|---|
| Claude Pro | $20 ($17 annual billing) | 5-hour session limits plus weekly caps |
| Cursor Pro | About $20 | Third-party reports; check Cursor’s site |
| GitHub Copilot Pro | $10 | $15 in AI credits, per GitHub |
| GitHub Copilot Free | $0 | 2,000 inline suggestions a month |
Notice that one $10 plan already covers a lot of the “90%” tier. You may not need five tools, and you may not need to leave your favourite one.
How do you build your own tiered setup?
- Pull your last few months of billing and usage. Our subscription audit guide explains why 13 months of statements is the right window.
- Tag each task as hard (design, deep debugging) or routine (renames, conversions, drafts).
- Keep one paid model for the hard tier. Move routine work to a cheaper plan, a free tier or a local model.
- Test for a month. If you hit limits constantly, upgrade; if you never do, downgrade.
Limits are the real cost, as we found in ChatGPT Business limits: why chat feels unlimited but Codex isn’t. Also keep context lean: one repo cut 390,200 tokens in our Claude Code context window piece.
My opinion: the rule is right and the specifics will age within months. Free-tier limits changed repeatedly in 2026, and sources disagree on what Google’s free API tier includes. Check official pages before you commit.
What happens next
Expect more vendors to move to credit-based billing, as GitHub already has, which makes tiering by task even more valuable. Limits and model lineups keep shifting, so any fixed $20 recipe needs a refresh every quarter. The V2EX thread is still open, so more replies may add context.
FAQ
Is Claude Pro worth $20 a month?
If you code with it for hours a day, yes, but you may hit session and weekly limits. Light users can often manage on cheaper or free tools.
Can free models really handle 90% of coding work?
For small edits and data chores, often. The 90% figure is the author’s rule of thumb, not a measured result, so test it on your own tasks.
Should I trust the post’s model claims?
Not blindly. Readers flagged outdated model names, and we could not verify the $19.99 bundle. Use the tiering idea, not the shopping list.
Next step: open your last three billing statements and mark each AI charge as hard-tier or routine. Cancel the one you did not touch.
Leave a comment