GLM-5.2 vs GPT-5.5: The Open Model That Just Made Frontier Coding Six Times Cheaper

Z.ai shipped a 753B open-weights model that beats GPT-5.5 on coding benchmarks at a sixth of the price. Here is what that actually means for your stack.

On June 16, 2026, a Chinese lab handed developers something the closed labs have been quietly hoping nobody would build yet: a frontier-grade coding model you can download, run on your own machines, and use commercially with almost no strings attached. GLM-5.2 from Z.ai beats GPT-5.5 on several of the coding benchmarks that actually matter, and it does it while costing roughly one-sixth as much to run.

If you write code for a living, or you pay the bill for people who do, this is the release worth paying attention to this month. Not because of the leaderboard bragging rights, but because of what the price column does to your monthly invoice.

Let me walk through what shipped, where the numbers hold up, where the marketing gets a little ahead of reality, and whether any of this should change how you work.

What GLM-5.2 Actually Is

GLM-5.2 is the flagship model from Z.ai, the lab formerly known as Zhipu AI. It is a Mixture-of-Experts design with somewhere between 744 and 753 billion total parameters, but it only fires about 40 billion of those on any single query. That routing trick is why a model this large can serve responses at a reasonable speed and cost. Each request gets sent to the slice of the network best suited to handle it, rather than waking up the entire 753 billion every time you ask it to fix a null pointer.

The headline feature is the context window. GLM-5.2 ships with a usable 1 million token window, up from 200,000 in GLM-5.1. That is a five-fold jump, and the word “usable” is doing real work in that sentence. Plenty of models claim a big context number and then fall apart somewhere past the halfway mark, forgetting what you told them at the top of a long session. Z.ai spent months training specifically on long, messy coding-agent trajectories so the model holds its quality deep into a million-token task. You can drop an entire mid-sized codebase into a single reasoning pass and have it stay coherent.

It also introduces effort levels. You pick “High” for a balance of speed and quality, or “Max” when you want the model to throw more computation at a genuinely hard problem. That control matters more than it sounds, because you are paying for those tokens and most tasks do not need maximum effort.

And then there is the license. Full weights are live on Hugging Face under the handle zai-org/GLM-5.2, released under a plain MIT license. No usage governance clauses, no acceptable-use bureaucracy bolted on. You can run it locally, modify it, fine-tune it, and ship it inside a commercial product without asking anyone.

The Benchmark Numbers, Read Honestly

Here is where most of the coverage this week stopped short. Everyone quoted the wins. Fewer people quoted the losses, and the losses tell you just as much about where this model fits.

Start with the wins, because they are real.

On SWE-bench Pro, which throws real-world software engineering tasks at a model, GLM-5.2 scored 62.1. GPT-5.5 managed 58.6. Its own predecessor, GLM-5.1, sat at 58.4. So this is a clear generational jump and a clear win over OpenAI’s flagship on a benchmark people respect.

On Terminal-Bench 2.1, it hit 81.0, up from 62.0 for GLM-5.1. That made it the first open-weights model to cross 80 on that test, and it lands within a few points of Claude Opus 4.8 at 85.0. Cline IDE flagged it publicly as beating every other open model available.

On FrontierSWE, which measures whether an agent can grind through open-ended projects that run for hours, GLM-5.2 reached 74.4 percent. GPT-5.5 was at 72.6. Claude Opus 4.8 sat at 75.1. Read that spread carefully. GLM-5.2 beat GPT-5.5 and finished within a single point of Opus on a test built specifically for the long-horizon work Z.ai is marketing.

On MCP-Atlas, a tool-usage evaluation, it scored 77.0 against GPT-5.5’s 75.3, just behind Opus 4.8 at 77.8. On Humanity’s Last Exam with tools enabled, it reached 54.7, ahead of GPT-5.5’s 52.2.

Now the part nobody put in their headline.

When you look at the very hardest, longest benchmarks, Claude Opus 4.8 is still clearly ahead, and sometimes by a lot. On SWE-Marathon, an ultra-long-horizon test covering things like building compilers and optimizing kernels, GLM-5.2 scored 13.0 against Opus 4.8’s 26.0. Opus literally doubled it. On Tool-Decathlon, Opus pulled ahead by nearly 12 points. On PostTrainBench, where each agent gets an H100 and is judged on how well it improves a smaller model through post-training, GLM-5.2 beat both GPT-5.5 and the older Opus 4.7 but still finished second to Opus 4.8.

So here is the honest summary, the one the marketing copy dances around. GLM-5.2 is the second-best coding model in the world on long-horizon tasks. “Second” means second to Opus 4.8, which still owns the messiest, longest, most punishing benchmarks. On the mid-length work that makes up most real development, GLM-5.2 trades blows with the frontier and beats GPT-5.5. On the absolute hardest marathon tasks, the gap is real and you will feel it.

That nuance is the difference between a useful decision and a hype-driven one.

The Price Is the Real Story

Benchmarks get the headlines. The pricing table is what changes behavior.

GLM-5.2 runs about $1.40 per million input tokens and $4.40 per million output tokens. Call it $5.80 combined per million. GPT-5.5 charges $5.00 for input and $30.00 for output, which lands around $35 per million. That is roughly six times more expensive for performance that, on a lot of coding work, is comparable or worse.

Against Claude Opus 4.8, the gap is similar in spirit. Opus runs $5 input and $25 output. GLM-5.2 undercuts it dramatically while landing within a point on several benchmarks.

Put that in concrete terms. Say an agent chews through 50 million input tokens and generates 10 million output tokens over a month of heavy work on a large repo. On GLM-5.2 that runs about $70 input plus $44 output, so roughly $114. The same volume on GPT-5.5 comes to $250 input plus $300 output, around $550. Same coding work, and the open model costs you a fifth as much. Scale that across a team of engineers each running agents all day and the annual gap turns into real headcount-sized money.

For a small studio or a solo builder, the math is more personal. That is the difference between AI-assisted development being a line item you barely notice and one you have to ration token by token.

And because the weights are open under MIT, there is a second pricing path that the API numbers do not even capture. If you have the hardware, you can self-host and pay only for compute and electricity. No per-token meter at all. That route is not realistic for most people, since a 753B model needs serious GPU clusters to run, but for an enterprise already sitting on that infrastructure, the marginal cost of inference drops toward the floor.

This is the thing closed labs have been bracing for. Not a model that wins every benchmark, but a model that gets close enough on open weights that the price difference becomes the whole conversation.

Under the Hood: Two Tricks Worth Knowing

Two architectural pieces explain how Z.ai got a million-token window to stay fast and affordable.

The first is IndexShare. At very long context, sparse attention models still spend a lot of compute figuring out which earlier tokens matter. IndexShare reuses one lightweight indexer across every group of four sparse-attention layers instead of computing a fresh one each time. At the full 1 million token length, Z.ai says this cuts per-token compute by 2.9 times. That single optimization is a big part of why the long context is cheap enough to actually use.

The second is an improved Multi-Token Prediction layer for speculative decoding. In plain terms, the model guesses several tokens ahead and verifies them in a batch, and the upgrade raises how many of those guesses get accepted by up to 20 percent. More accepted guesses means faster generation at the same quality.

You do not need to care about either of these to use the model. But they explain why the pricing is not just Z.ai eating a loss to buy market share. The efficiency is built into the architecture.

There is also a piece of backstory worth knowing for context. The GLM-5 line has been trained largely on non-Nvidia hardware, with the earlier GLM-5.1 reportedly trained entirely on Huawei Ascend chips. That detail matters for anyone thinking about supply-chain resilience and where frontier compute is allowed to come from.

Why This Release Landed When It Did

Timing is part of the story. GLM-5.2 arrived in the middle of a messy stretch for proprietary American models. A recent export control directive restricted foreign nationals from using one of Anthropic’s models, and that kind of geographic fencing makes enterprise buyers nervous about building critical workflows on top of a model they might lose access to.

An MIT-licensed open-weights model sidesteps all of it. Download the weights, run them inside your own walls, and no policy change in another country can switch off your tooling. For a multinational engineering org, that is not a nice-to-have. It is a procurement requirement.

Z.ai also did not release this into a vacuum. The model launched with same-day support across more than 20 coding environments, including Claude Code, OpenClaw, Cline, Kilo Code, and Codex. You can run it through Ollama with a single command. The company also shipped a GLM Coding Plan with subscription tiers starting around $12.60 per month for Lite, $50.40 for Pro, and $112 for Max, billed annually, for people who want hosted access without managing infrastructure.

That distribution muscle is why this release is being felt immediately rather than six weeks from now once integrations catch up.

It also helps that Z.ai has been shipping at a pace that keeps the model in front of developers. GLM-5.2 follows GLM-5 and GLM-5.1 in a tight release cadence through the first half of 2026, and each step closed more of the gap to the closed frontier. GLM-5.1 was already the first open-weights model to top SWE-bench Pro. GLM-5.2 took that foundation, multiplied the context window by five, and pushed the coding scores up across the board. The trajectory matters as much as any single number, because it tells you the gap between open and closed is shrinking on a timeline measured in weeks, not years.

Should You Actually Switch?

Here is the practical decision, stripped of cheerleading.

Switch to GLM-5.2 if your coding workload is high-volume and cost-sensitive, if most of your tasks are short-to-medium in length rather than hours-long marathons, if open weights or self-hosting matter for compliance or supply-chain reasons, or if you are currently paying GPT-5.5 prices for work that does not need the absolute frontier. For a huge share of real development, you will not notice a quality drop, and you will notice the bill.

Stay on Opus 4.8 if your work lives at the extreme end of long-horizon difficulty, the compiler-building, kernel-optimizing, multi-hour autonomous engineering tasks where the benchmark gap is widest. If you are paying for the top one or two percent of capability and that margin pays for itself, GLM-5.2 is not yet a clean replacement there.

Run both if you are doing this seriously. The smartest setup right now is routing. Send the bulk of your volume to GLM-5.2 for cost, and escalate only the hardest tasks to Opus. The price difference is large enough that even a crude split saves real money while keeping a frontier model on call for the work that needs it.

One caution worth stating plainly. Benchmark scores depend heavily on the harness and prompting strategy used to produce them. The numbers above come from Z.ai’s own evaluations and early third-party testing. They are credible and broadly consistent across sources, but your results on your codebase with your tooling are the only benchmark that actually decides this for you. Run a real task through it before you migrate anything important.

Frequently Asked Questions

Is GLM-5.2 really free to use? The weights are free under an MIT license, so you can download and run the model at no licensing cost. You still pay for the hardware to run it, or for API access if you do not self-host. API pricing is about $1.40 per million input tokens and $4.40 per million output tokens.

Can I run GLM-5.2 on my own computer? Not on a normal machine. At 753 billion parameters it needs enterprise-grade GPU clusters to run the full model. Most individual developers will use it through the API or a hosted provider rather than self-hosting. The open license matters most for organizations that already own serious compute.

Does GLM-5.2 beat Claude Opus 4.8? On some benchmarks it comes within a point, like FrontierSWE (74.4 vs 75.1) and MCP-Atlas. On the hardest long-horizon tests like SWE-Marathon, Opus 4.8 is still clearly ahead, sometimes by double. GLM-5.2 is best understood as the strongest open model and a close second to Opus overall, at a fraction of the price.

How big is the context window, really? 1 million tokens, and Z.ai trained specifically to keep quality stable across that full length rather than just accepting more input. That is up from 200,000 tokens in GLM-5.1. It is enough to reason over an entire mid-sized codebase in one pass.

What coding tools support it? GLM-5.2 shipped with day-one support for over 20 environments, including Claude Code, Cline, Kilo Code, OpenClaw, Codex, and OpenCode. You can also run it through Ollama with a single command.

What is the difference between High and Max effort modes? High balances quality against speed and cost for everyday tasks. Max allocates extra computation for genuinely hard problems where you want peak performance and are willing to pay more tokens for it. Most work runs fine on High.