GLM-5.2 costs a fifth of GPT-5.5 and says it codes better. One of those claims you can check.
Z.ai's open-weight model undercuts OpenAI's flagship by roughly 5x per token, and its own benchmark sheet shows it edging GPT-5.5 on coding. The price is on the rate card, in writing. The benchmark win is on a sheet Z.ai graded itself. Worth keeping those two facts in different mental buckets.
Z.ai turned on the per-token API and open weights for GLM-5.2 around June 16, a few days after the model first showed up behind its coding subscription. Within a day, the framing was already set by a VentureBeat headline: an open model beating GPT-5.5 on coding for a fraction of the cost. That is a great story. It is also two claims welded together, and they are not equally solid.
We have written up GLM-5.2 on its own and the wider price war it sits inside. This piece is narrower: should the cost gap with GPT-5.5 pull your coding agent off OpenAI? The honest answer depends on how much you trust a benchmark nobody outside Z.ai has reproduced, so let's start with the part that needs no trust at all.
The price gap is the boring, verified part
Both prices come straight off published rate cards, dollars per million tokens. No interpretation required.
| Model | Input / 1M | Output / 1M | Cached input | Context |
|---|---|---|---|---|
| GLM-5.2 | $1.40 | $4.40 | ~$0.26 | 1M |
| GPT-5.5 | $5.00 | $30.00 | $0.50 | 1.05M |
GPT-5.5 charges 3.6x more to read a token and just under 7x more to write one. Output is where the spread is brutal: $30 against $4.40. If you have priced coding agents before you know why that matters, but if not, the next table makes it concrete. The cached column is worth a glance too: GLM-5.2's $0.26 cache-hit rate, listed on Z.ai's own pricing page, comes in under half of GPT-5.5's $0.50, so repeated prompts stay cheaper on GLM even after OpenAI's caching discount kicks in. OpenRouter, for what it's worth, lists GLM-5.2 a shade lower at $1.20/$4.10 through its routing.
Run the same month through both meters
Coding agents read far more than they write. They keep re-loading files, diffs, and tool output on every step, then emit a small patch. So the bill is mostly input volume, and even though GPT-5.5's output rate is the scarier number, the input rate does most of the damage at scale. I priced three real shapes of coding work, all at standard uncached rates, all under the 272K context line.
| Who's running it | Monthly tokens | GLM-5.2 | GPT-5.5 | GPT-5.5 premium |
|---|---|---|---|---|
| Indie dev on an agent | 25M in / 5M out | $57.00 | $275.00 | +$218.00 |
| Small team, shared agent | 50M in / 10M out | $114.00 | $550.00 | +$436.00 |
| Monorepo CI bot, all day | 250M in / 50M out | $570.00 | $2,750.00 | +$2,180.00 |
The multiplier holds steady at about 4.8x down the whole column, because the workload mix barely moves. An indie dev pockets a couple hundred a month. The team running a bot against a monorepo all day clears north of two grand. Whether that is decisive depends entirely on the next section. Plug your own split into the cost calculator if your ratio looks nothing like mine.
And the gap gets worse past 272K tokens
One detail that does not show up in the headline rate: GPT-5.5 bills any request over 272,000 input tokens at 2x input and 1.5x output, for the whole request, and that surcharge stomps on Batch and Flex discounts. We dug into that context cliff separately. It bites exactly when you would reach for a big context window, which is repo-scale agent work. GLM-5.2 has no such penalty and a full 1M window, so the moment your prompts get large, the 4.8x gap quietly stretches wider. For the kind of person who feeds an entire service into the model and lets the agent wander, that is the difference that actually shows up on the invoice.
Now the part Z.ai graded itself
The reason this comparison is interesting and not just "cheap model is cheap" is that GLM-5.2 claims to win on quality too. Here is what Z.ai published. Read the third column before the first two.
| Benchmark | GLM-5.2 | GPT-5.5 | Who measured it |
|---|---|---|---|
| SWE-Bench Pro | 62.1 | 58.6 | Z.ai self-report |
| FrontierSWE | 74.4 | 72.6 | Z.ai self-report |
| Terminal-Bench | 81.0 (v2.1) | 82.7 (v2.0) | Different versions |
| Intelligence Index | ~51 | higher | Artificial Analysis |
Three of those four rows came from Z.ai's own harness. Vendor scaffolds flatter vendor models, that is just how it goes, and nobody has reproduced the SWE-Bench Pro or FrontierSWE wins on a neutral setup. The Terminal-Bench row is worse than self-reported: the two scores are on different versions of the test, 2.1 against 2.0, so putting them side by side is closer to noise than to a result. Strip out the unverified rows and the only independent signal is Artificial Analysis, which puts GLM-5.2's Intelligence Index around 51. That is genuinely strong for an open-weight model. It is not the same thing as "beats GPT-5.5 at coding."
So where does that leave the headline? GLM-5.2 might well be in GPT-5.5's league on real coding tasks. It might also be a couple of points behind once someone runs it cleanly. We do not know, and pretending the self-reported sheet settles it would be the easiest way to give you bad advice.
What the extra money still buys you on GPT-5.5
Cheaper is not free of tradeoffs. GPT-5.5 has been in production since April 23 and sits at or near the top of the independent intelligence rankings, with a long track record of real-world agent runs behind it. OpenAI also claims GPT-5.5 burns meaningfully fewer output tokens per task than GPT-5.4, which would shave its effective cost below the sticker gap, though that is OpenAI's own number and unverified too. And the tooling around it, every SDK, every agent framework, every Codex integration, is mature in a way a model that shipped its rate card last week simply is not.
There is also the question nobody loves bringing up: GLM-5.2's hosted API runs through Z.ai's infrastructure in China, which is a non-starter for some compliance teams regardless of price. The escape hatch is the open weights. Pull GLM-5.2 under its MIT license, run it on your own boxes, and the data-residency worry and the per-token bill both disappear at once. GPT-5.5 gives you no such door.
The call
If your coding workload is large and somewhat forgiving, GLM-5.2 is hard to argue against on cost. You are looking at roughly a fifth of the GPT-5.5 bill, a window four times bigger, no long-context surcharge, and an MIT license that lets you walk away from the meter entirely. For batch refactors, test generation, and the kind of high-volume agent work where a wrong answer is cheap to catch, the math is on Z.ai's side by a wide margin.
Stay on GPT-5.5 when the work is unforgiving and you are paying for a known quantity: a proven model, a mature toolchain, and a compliance story your security team has already signed off. The benchmark sheet says you would be overpaying. The benchmark sheet was written by the company that wants your business. Until an independent lab reruns those coding numbers, the safe read is that you are buying down risk, and only you can price what that is worth.
Sources
- GLM-5.2 pricing and listing - OpenRouter ($1.40/$4.40 on Z.ai, $1.20/$4.10 routed; 1M context)
- GLM-5.2 independent evaluation - Artificial Analysis (Intelligence Index ~51, throughput, the one outside measurement)
- GLM-5.2 beats GPT-5.5 framing - VentureBeat (self-reported SWE-Bench Pro 62.1 vs 58.6, FrontierSWE, MCP-Atlas)
- GPT-5.5 pricing - aipricing.guru ($5/$30, cached $0.50, 272K long-context surcharge 2x/1.5x)
- GLM-5.2 open weights - Hugging Face (MIT license, downloadable for self-hosting)