
Something is shifting in the AI coding-subscription market, and you can see it happening in real time on China's developer forums. A recent thread on LINUX DO asked a simple question: "Why is everyone subscribing to Claude Pro all of a sudden?" The 50-plus replies that followed form one of the most honest community verdicts on the current coding-assistant landscape — covering quotas, bans, model routing, and real benchmark results on legacy code.
The short answer, per the forum: Anthropic's newest flagship model is finally usable on the $20/month Pro tier, Claude's quota feels generous, and developers are frustrated with what they see as opaque model routing and quality degradation on OpenAI's side. One of the most-liked replies in the thread summed up the mood: switch with your wallet, and "use a language capitalists understand."
A note on sources before we dive in: everything below that comes from the forum thread is attributed as such — it's one community's self-reported experience, not a controlled study. Where a claim is independently verifiable, we've checked it.
The forum wave coincided with a real product event. Anthropic released Claude Opus 5.5 on September 22, 2026, per TechCrunch's reporting: output tokens at $20 per million (down from $25), performance described as comparable to the larger Fable-class models, and Sonnet 5.5 and Haiku 5.5 slated to follow in the coming weeks. The release was reportedly evaluated before launch by outside groups including Frontier Design and METR — unusual extra care, and Anthropic's first launch since the company publicly called for pacing frontier progress.
That much is independently confirmed. What the forum added is the human layer: members nicknamed the new model "pelican" (鹈鹕) after the community's own unofficial benchmark — a one-shot 3D pelican-on-a-bicycle generation test that had gone viral in a previous thread — and several members reported that demos built from single-sentence prompts were impressive enough to convince skeptics the model could handle real coding work. Those demos and the nickname are forum lore, not Anthropic claims; treat them as community color.
Several developers reported that the new flagship is accessible on the standard $20/month Pro plan, and that Claude Pro's quota feels generous ("比较馋人" — tantalizingly ample, in the community's words). For a market where international credit cards are uncommon among students and young developers, the value-per-dollar calculation matters enormously — and Claude is winning that calculation right now, at least in this thread's telling.
This is worth keeping in perspective: quota "feel" is subjective, and several members in the same thread cautioned that heavy agentic-coding sessions can still burn through allowances. But the perception of generosity is itself doing marketing work for Anthropic — nobody rage-quits a subscription they think is abundant.
The other side of the story is discontent with ChatGPT. Members described opaque model routing — not knowing which model is actually answering — and a perceived quality degradation ("降智", roughly "getting dumber") on complex tasks. When you're paying for a top-tier model and suspect you're being served a cheaper one, trust erodes fast. Whether or not the suspicion is technically accurate, the perception is driving real subscription churn, and this isn't the first time routing opacity has been a complaint in developer communities.
Anthropic benefits here by contrast: the thread's members repeatedly said they could tell which model they were getting. In a subscription market, feeling like you got what you paid for turns out to be a feature.
Perhaps the most credible part of the discussion is the community's own testing. Multiple members compared Claude and Codex performance on "屎山" (shǐshān, literally "mountain of shit") — the affectionate term for inherited legacy codebases. These are the tasks that actually matter to working developers: understanding tangled old code, tracing bugs through layers of undocumented logic, and making safe changes. The community's verdict, backed by real projects rather than leaderboard scores, tilted toward Claude.
Community testing is not peer review — sample sizes are small, confirmation bias is real, and the winners of a forum argument aren't always the best models. But there's a reason these threads are influential: they test what leaderboards don't. A model that navigates a 200,000-line legacy codebase without breaking things is demonstrating something no multiple-choice benchmark measures.
Here's how the discussion stacks up the two options on the factors developers actually care about:
| Factor | Claude Pro ($20/mo) | ChatGPT (Pro/Codex) |
|---|---|---|
| Frontier coding strength | New flagship (Opus 5.5, Sept 2026); members say strong on Pro tier | Strong, but members say routing is opaque |
| Quota generosity | Members report generous on Pro plan | Varies; top tiers are expensive |
| Legacy-code handling | Community-tested, well regarded | Mixed community verdict |
| Model transparency | Clear which model you get | Routing often unclear |
| Ban/refund risk | Members report bans with weak refunds | Familiar, established |
Two rows need extra caution. The "$500/month" figure came from the same forum's code-diving, which reportedly found an upcoming "Pro Max" tier with regional pricing reaching $500 in the US — an unverified discovery from forum members poking at client code, not an OpenAI announcement. And the ban risk row reflects self-reports: the thread's original poster was asking precisely because a colleague had lost a Pro account without a refund, a reminder that payment friction — and recourse friction — is real for Chinese developers using international subscriptions.
It's worth stating plainly what this kind of article is and isn't. A 50-reply forum thread is a snapshot of one community's mood at one moment, filtered through the people motivated enough to post. The developers who quietly stayed on ChatGPT, or who tried Claude Pro and found the quota stingy, mostly didn't write in. The thread's verdicts — generous quotas, better legacy-code handling, opaque routing — are real signals, but they're signals from a specific, price-sensitive, deeply technical audience whose priorities (value per dollar, transparency, real-project testing) may not match yours. Treat this the way the forum itself treats the pelican test: a useful, honest instrument with known blind spots, not a peer-reviewed result.
Chinese developers are among the world's most intense AI-coding users — and some of the most price-sensitive, given payment and access frictions. Their collective subscription choices are a leading indicator of where the coding-assistant market is heading. When a developer community known for squeezing maximum value out of every dollar starts moving en masse from one product to another, it's worth paying attention.
The deeper lesson is about what actually wins subscriptions. It's not leaderboard scores or launch-day hype — it's a combination of: a model that demonstrably handles real legacy code, a quota generous enough that you stop thinking about it, and the feeling that you're getting what you paid for. As the community showed when they hunted for free cloud compute with the same intensity, this is a culture that measures value ruthlessly.
Whether Claude keeps this momentum depends on execution — quotas, reliability, and continued coding strength. But the verdict from one of the world's most demanding developer communities is clear: right now, Claude Pro is winning the wallet vote. Just remember it is a forum verdict — passionate, detailed, and worth reading, but one thread among many.