Published: · Last updated:
Note, September 2026: There is a newer GLM. GLM-5.3 Flash is now Chat-O’s default model in every tier, at a lower price point than either model in this comparison. We have since put it through real work rather than a synthetic benchmark: it migrated our email marketing stack to Pombo in under an hour, and it migrated a full Rails codebase to Go. We keep this GLM-5.2 vs Kimi K3 comparison because the flagship reasoning class it covers is still the right choice for heavy agentic work, and every conclusion here was tested firsthand.
The million-token era is here. GLM-5.2 (Z.AI, released June 16, 2026) and Kimi K3 (Moonshot AI, released July 16, 2026) both support 1,048,576-token context windows, both are open weight, and both are strong at reasoning and code. They are not interchangeable though: one costs about four times more per token, and the benchmarks tell a more nuanced story than either vendor’s marketing table.
We ran both models ourselves on identical prompts through OpenRouter, read the public benchmark record, and compared what it actually costs to run each one. This guide is the result.
On Chat-O, GLM-5.2 runs in the Balanced and Power tiers and Kimi K3 runs in Power. Both are available in the model dropdown today.
| Feature | GLM-5.2 | Kimi K3 |
|---|---|---|
| Creator | Z.AI (Zhipu AI) | Moonshot AI |
| Total parameters | ~753B (~40B active) | 2.8T (~50B active, estimated) |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Modality | Text only | Text + image + video input |
| License | MIT | Open weights, custom “Kimi K3” license |
| Long-context trick | IndexShare sparse attention | Kimi Delta Attention (hybrid) |
| Released | June 16, 2026 | July 16, 2026 |
| Tiers on Chat-O | Balanced, Power | Power |
We ran the same three prompts through both models on OpenRouter and timed them. The prompts were small, but the pattern matched what the benchmark record says.
Test 1: a logic puzzle. The classic three-switch, three-bulb problem. Both models nailed it: turn one switch on for a few minutes, swap it for the second, enter the room, and read light plus warmth as two bits of information per bulb. GLM-5.2 answered in 3 seconds with a tight explanation. Kimi K3 took 23 seconds and produced a more structured answer with an observation table. Same conclusion, very different serving cost.
Test 2: find the bugs in a Python one-liner. A “top-3 most frequent words, case-insensitive, ties broken alphabetically” function with a broken sort key. This was the most revealing test.
lambda kv: (-kv[1], kv[0]). Solid, fast to read, slightly shallow. .lower() is not enough for full Unicode case-insensitivity, .casefold() is. K3 was the more thorough reviewer. Test 3: write a two-sentence release note. Both produced clean, hype-free copy. GLM-5.2 leaned generic. Kimi K3 gave the more actionable version. K3 took about twice as long.
The cost gap in practice. On the code test, GLM-5.2 spent $0.0057 and Kimi K3 spent $0.0273, a 4.8x difference for a comparable answer. Multiply that across an agent loop making dozens of calls per task and the gap becomes a budget line.
Public benchmarks partly favor Kimi K3, but the honest picture requires separating vendor-reported numbers from independent ones.
| Benchmark | GLM-5.2 | Kimi K3 | Who measured it |
|---|---|---|---|
| Terminal-Bench 2.1 | 81.0 | 80.9 | Z.AI and Vals, independently, same harness: effectively tied |
| Artificial Analysis Intelligence Index | 51 | 57 | Artificial Analysis |
| GPQA-Diamond | 91.2 | 93.5 | Vendor-reported |
| AIME 2026 | 99.2% | ~93% | Vendor-reported |
| BrowseComp (web research) | not reported | 91.2, best in class | Moonshot |
| SWE-bench Verified | ~78 | 76.8 | Independent estimates |
| SWE-bench Pro | 62.1 | not reported | Z.AI harness |
Two things matter here.
First, the vendor tables are not apples to apples. Moonshot’s published comparison shows K3 leading GLM-5.2 on long-horizon agentic benchmarks by large margins (for example 42.0 vs 13.0 on SWE Marathon), but those rows used different agent harnesses, and the harness can move scores by double digits. On the one benchmark where both models were measured independently on the same harness, Terminal-Bench 2.1, they are within 0.1 points of each other.
Second, where K3’s lead is real, it is real in specific places: broad intelligence (57 vs 51 on the Artificial Analysis index), web research (BrowseComp 91.2), long-document and multimodal work, and frontend code quality, where K3 took the number one spot on human-preference arenas. For bounded, well-specified tasks, the two models are closer than the price difference suggests.
Pricing as of September 2026, checked live on OpenRouter:
| Cost basis | GLM-5.2 | Kimi K3 |
|---|---|---|
| Creator’s list price (input / output) (input / output) | $1.40 / $4.40 | $3.00 / $15.00 |
| Cheapest provider on OpenRouter | $0.36 / $1.12 | $2.50 / $14.00 |
| Providers on OpenRouter | 35+ | ~18 |
| Approximate output-price ratio | 1x | ~4x |
GLM-5.2 is served by more than 35 endpoints on OpenRouter, from budget providers at $0.36 per million input tokens to the official Z.AI endpoint at list price. Kimi K3 is available on around 18 endpoints, with Moonshot’s list price of $3.00 input and $15.00 output as the anchor. A full million-token input costs under a dollar at GLM-5.2’s cheapest endpoint and about $2.50 to $3.00 on Kimi K3; the gap widens sharply on output tokens.
For pure API usage, we love OpenRouter and it is what we recommend: one API key, every provider, transparent per-token pricing for both models.
Both models ship weights on Hugging Face, but the licenses are not the same proposition.
If self-hosting and license simplicity matter, that is a genuine GLM-5.2 advantage. If you want the strongest open model that exists and can live with a custom license, that is K3.
What about GLM-5.3 Flash? Where does it fit?
GLM-5.3 Flash is the newer, faster member of the GLM family and the current Chat-O default in every tier. It trades some flagship reasoning depth for speed and price, so it complements rather than replaces this comparison: if GLM-5.2 vs Kimi K3 is about the heavy reasoning class, GLM-5.3 Flash is about making capable AI cheap enough to run everywhere. See our GLM-5.3 Flash deep dive. We have now tested it on two real migrations: moving Chat-O’s email marketing to Pombo in under an hour, and the full Rails-to-Go codebase migration it ran as the stealth model ox alpha before its identity was revealed.
Is Kimi K3 better than GLM-5.2?
On broad benchmarks, yes, by a modest margin: 57 vs 51 on the Artificial Analysis Intelligence Index, plus native vision. On the one independently verified terminal-agent benchmark (Terminal-Bench 2.1), they are tied at about 81. GLM-5.2 is about 4x cheaper per output token and about 3x faster to serve. “Better” depends on whether capability or economics is your constraint.
Can I run GLM-5.2 and Kimi K3 on Chat-O?
Yes. GLM-5.2 is available in the Balanced and Power tiers, Kimi K3 in the Power tier. Both are one click away in the model dropdown, and you can switch between them mid-conversation. Sign up for Chat-O to try both on the same prompt, which is the fastest way to feel the difference.
Where can I compare GLM-5.2 and Kimi K3 prices?
On OpenRouter, which lists every provider and price for both models. As of September 2026, GLM-5.2 starts at $0.36 per million input tokens and Kimi K3 at $2.50, with creator list prices of $1.40/$4.40 and $3.00/$15.00 respectively.
Are both models open weight?
Yes, with different licenses. GLM-5.2 is MIT licensed, the most permissive option for commercial use and self-hosting. Kimi K3’s weights are downloadable under Moonshot AI’s custom license, which is more restrictive than MIT.
Which model for coding agents?
For short, bounded coding tasks and high-volume review, GLM-5.2. For multi-hour autonomous runs where the agent must hold context across dozens of files, Kimi K3’s long-horizon results are the strongest public record, at a price to match.
Want to see GLM in an agent seat on a real job? We used GLM-5.3 Flash to migrate our entire email marketing stack in under an hour.
Sources: