Open weights used to mean accepting a real capability drop to escape metered API pricing. GLM-5.2 makes that trade-off much less obvious. It scores 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 2.1, which puts it above GPT-5.5 on the first and four points behind Claude Opus 4.8 on the second, at roughly a sixth of frontier API cost.
It doesn't come in first. That's not the point.
How does GLM-5.2 compare on coding benchmarks?
Two benchmarks matter for agentic work. SWE-bench Pro measures whether the model resolves a real repository issue end to end. Terminal-Bench 2.1 measures whether it holds up through a long shell session without losing the thread.
| Model | SWE-bench Pro | Terminal-Bench 2.1 | Weights |
|---|---|---|---|
| Claude Opus 4.8 | 69.2 | 85.0 | Closed |
| GLM-5.2 | 62.1 | 81.0 | Open, MIT |
| GPT-5.5 | 58.6 | n/a | Closed |
Read the gaps, not the ranking. GLM-5.2 sits 7.1 points below Opus 4.8 on SWE-bench Pro and 4 points below on Terminal-Bench. And it sits 3.5 points above GPT-5.5 on SWE-bench Pro. An open-weight model beating one frontier lab and losing to another by single digits is a different situation than the one that existed a year ago.
How much does each one cost on the official API?
| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Claude Opus 4.8 | US$ 5 | US$ 25 |
| GLM-5.2 | US$ 1.40 | US$ 4.40 |
Source: official launch pricing, Anthropic (anthropic.com) and Z.ai (z.ai), checked on 2026-09-15. On output, where a reasoning model spends most of its budget on a coding task, GLM-5.2 comes out 5.7 times cheaper. On input, the gap drops to 3.6 times. "About a sixth of the cost," as this article's opening puts it, is the output-side approximation, the one that matters most in a long agentic session.
What's inside the model?
| Total parameters | 753B |
| Active per token | about 40B |
| Context window | 1M tokens |
| Max output | 128K tokens |
| License | MIT |
| Released | June 13, 2026 |
The MoE ratio is the line worth noting. A 753B model that activates about 40B per token costs roughly what it costs to serve a 40B dense model, while carrying the knowledge of something nineteen times bigger. That ratio is what puts the price-per-point where it is.
The 1M context is the second thing. In a long agentic run, context is what keeps the agent from re-reading the same files, re-deriving the same conclusions, and burning a second pass on work it already did.
Why does the MIT license matter more than the score?
MIT means you run, modify, and sell it commercially without asking anyone. No usage clause that changes next quarter, no per-seat audit, no paragraph about competing products.
Most open-weight releases ship with a custom license and some hidden exception. A genuinely permissive license on a model this close to the frontier undercuts the argument that open weights are only good for hobby projects.
The benchmark score tells you what the model does today. The license tells you whether you're allowed to build a business on top of it for the next five years. Only one of those two numbers is stable.
Where does GLM-5.2 still lose?
Honestly, at the top of the difficulty curve. Those 7.1 points on SWE-bench Pro aren't spread evenly across easy and hard tasks. A gap like that concentrates in the hardest cases: the ambiguous issue, the refactor that touches a dozen files, the bug whose reproduction is already a puzzle.
If that's most of your work, Opus 4.8 is still the better tool and the price is justified. If your work is the normal ticket distribution any codebase generates, you're paying six times more for a difference you won't feel most days.
How do you actually decide?
- Look at your hardest ticket from last month, not the average one. The benchmark gap only shows up there. If the frontier model wouldn't have saved you either, the difference is theoretical.
- Price per point, not price per token. Six times the cost for eleven percent more capability is a bad trade in volume and a great trade for one critical run.
- Stop treating it as a single choice. The real answer is usually routing: the open model for the bulk of the work, the frontier one for the few tasks that earn it.
The catch no benchmark table shows
Every number above assumes the model finishes. Under metered billing, it often doesn't. A 1M context window is only an advantage if you can afford to fill it, and long agentic runs are exactly the kind of task that gets cut short when someone's watching the token bill.
Those 7.1 points of difference will only show up on your hardest ticket this month. For everything else, what decides isn't the score, it's how many attempts you can afford. On Verboo Code, GLM-5.2 runs on dedicated GPU with unlimited tokens, so the fifth attempt costs the same as the first, and the 1M window stops being a spec-sheet number.
