For weeks now there has been a viral thread on X about developers migrating en masse from Claude Code to GLM. Screenshots of cancelled Anthropic subscriptions, glowing quotes, and the headline claim on top: as good as Opus, the bolder ones even reach for Fable, and all of it five, maybe seven times cheaper.
My entire working stack runs on Claude, and the API bills come to me. So the question was not idle. But the thread with tens of thousands of likes was missing the main thing: what exactly people measured when they got their cherished five. I had to measure it myself.
I took two tasks, and picked them not for show but from what I actually get paid to do.
The first one is business: triaging inbound support email. Classify requests, extract data, call tools, draft a reply. Almost every client brings me this task in some form, whatever industry they are in. Boring, high-volume, repeated thousands of times a month, which is exactly where token savings turn into real money.
The second is engineering: a bug from a real project. After payment, an order sometimes hangs in an intermediate status even though the money was charged. Sometimes. Not every time. A race between the payment webhook and the state update, plus a component reading the status from a stale cache. To find the cause you have to connect the behaviour of three files, each of which looks innocent on its own.
I wanted the comparison to be fair, and fair means only the model changes. Here a curious thing surfaced that both sides keep quiet about: the GLM Coding Plan plugs directly into Claude Code through environment variables. Same harness, same tools, only the brain changes. Anthropic does not advertise it, because their harness starts feeding someone else's model. Z.ai does not advertise it, because then their own app turns out to be an option rather than a necessity. For a test, though, you could not ask for better conditions.
GLM handled the email. Not almost handled it, but properly: routing without misses, confident tool calls, decent drafts, German and Polish no worse than English. The furious all-caps letter was correctly escalated to a human. The difference with Opus is cosmetic: Opus holds brand tone better without reminders, and that is it. Nobody pays five times more for that.
The economics, though, did not match the thread. On the price list GLM is 5.7x cheaper, but the logs show it thinks out loud, at length. Where Opus answers in three lines, GLM unrolls a wall of reasoning. Cheap tokens that come in one and a half to two times the quantity end up giving you savings closer to four, not seven. Still a lot. If my production consisted of emails, the article could end here.
But then came the bug.
GLM found the cause in a minute. Or rather, what it believed was the cause. There really is a suspicious-looking line in the webhook handler, and GLM wrapped it in checks, wrote a confident commit message and reported victory. The code compiles, the tests are green, the diagnosis sounds convincing. One problem: it is a symptom, not the cause. The race lived on, now covered by a neat patch, ready to surface later in a place where it will take twice as long to find.
On the second attempt, with heavy hints, GLM connected two files out of three. The cache stayed outside its picture of the world. Opus built the full trace on the first run: here is the race, here is why it is rare, here is the stale cache, here is the fix and the regression test. You read it and nod, because you once reached the same conclusion yourself after losing an evening to it.
And here the thread's arithmetic flips. The GLM run cost less, but it did not solve the task. A cheap unsolved attempt is the most expensive kind there is: you pay for tokens, then with your own time reviewing plausible garbage, then once more when the covered-up bug fires in production. A model is cheap exactly until the moment it stops coping.
A few more things came up along the way that no comparison table mentions. GLM is blind, literally: it cannot look at a screenshot, and for frontend work that is a wall (Z.ai has a sighted crutch through a second model, but a crutch is what it is). The tariffs are counted slyly: the promised 80 prompts per 5 hours actually divide by three, because the flagship model burns quota with a triple multiplier. And prompts do not transfer for free: mine have spent a couple of years moulding themselves to Claude's behaviour, and for the first few hours GLM looked worse than it is. Migrating a process to another model costs days, and that line is not in any price list.
The most interesting part turned up when I went to the live leaderboards to check my impressions. On the benches that measure the cost of a task carried through to the end, GLM plays in the Sonnet league: it solves about half at a modest price per attempt. Respectable, but no Opus killer. One line above, though, sits Kimi K3 from Moonshot: 72.8% of tasks solved at $48 per attempt, against 71.5% at $113 for Claude Opus 5. An open-weight model above Anthropic's flagship, at less than half the price per attempt. So the thread is broadly right: cheaper really is no worse. It is just that all of X is pointing at the wrong model (Z.ai's marketing budget is bigger, apparently).
The same leaderboard holds the counterexample, for symmetry: a model at $0.50 per million tokens, ten times cheaper than Opus on paper, that solves 19% of tasks and burns over a hundred dollars per attempt. Great fuel economy. The car just does not arrive.
In the end nothing got replaced in my stack; it got rearranged. High-volume routine, where the cost of a mistake is an inconvenience, moved to the cheap model: across thousands of calls a day a fourfold saving is the client's money, not percentages in a table. Multi-file engineering and long agent chains stayed on the expensive one: there, plausible lying costs more than a model honestly admitting it does not know. The price of the expensive model in those tasks is an insurance premium, not an expense.
The numbers here will go stale within a quarter: GLM 5.3 is already out, Kimi keeps updating, Anthropic knows how to answer. But the conclusion should live longer. Price per million tokens measures fuel consumption without asking whether the car arrived. The only number that means anything is the cost of a solved task. Your task, on your data.