A Chinese open-weight model just closed most of the distance to the best coding AI on the planet, and it did it without a license fee standing in the way. GLM-5.2, built by the Beijing lab Z.ai (also known as Zhipu AI), has spent the summer climbing independent benchmark tables, and by mid-September 2026 the pattern is hard to miss: it beats every current GPT model on real-world coding tasks and sits within a few points of Anthropic’s Claude Opus 4.8 on the hardest agentic tests. It does this as a fully open-weight, MIT-licensed release that developers can download and run themselves.
That combination, frontier-adjacent performance plus free weights, is why GLM-5.2 keeps showing up in benchmark roundups from Codersera, Apidog, Layer3 Labs, and half a dozen other AI research outlets this quarter. The numbers are specific enough to check, and several outlets have done exactly that, sometimes finding gaps between what Zhipu claims and what independent labs measure. Below is what’s actually confirmed, what the discrepancies mean, and why the open-weight coding race matters more than another leaderboard shuffle.
GLM-5.2 Beats GPT-5.5 on Real Coding Work
The single number driving most of the coverage is GLM-5.2’s score on SWE-bench Pro, a benchmark that hands a model a real GitHub-style repository and a real bug report and checks whether it produces a working fix across multiple files. GLM-5.2 scored 62.1 on that test. GPT-5.5 scored 58.6. Claude Opus 4.8, still the reference point for coding-focused models, scored 69.2, according to benchmark tables published by Codersera and Layer3 Labs.
That puts GLM-5.2 ahead of OpenAI’s current flagship by 3.5 points and behind Anthropic’s by roughly 7. It’s also a real jump from Zhipu’s own prior release: GLM-5.1 scored 58.4 on the same test, meaning the newest version gained almost four points in one generation. For a model family that ships new versions every few months, that’s a fast climb.
On FrontierSWE, a separate benchmark built around long, open-ended engineering projects rather than single bug fixes, the gap to Claude nearly disappears. GLM-5.2 scored 74.4 against Claude Opus 4.8’s 75.1, a difference of seven tenths of a point, according to Layer3 Labs’ benchmark review. GPT-5.5 trailed both at 72.6 on the same test.
Where the Vendor’s Numbers and Independent Testing Disagree
Zhipu’s own launch materials put GLM-5.2 at 81.0 on Terminal-Bench 2.1, the benchmark that measures how well a model can operate a real command-line environment across multi-step tasks without human help. Most secondary coverage repeats that figure, and it would put GLM-5.2 close behind Claude Opus 4.8’s reported 85.0 and roughly level with GPT-5.5’s 84.0, per a side-by-side table published by Codersera.
But when Artificial Analysis, an independent benchmarking firm that runs its own evaluation harness rather than relying on vendor self-reports, tested GLM-5.2 on Terminal-Bench 2.1, it measured 77.9, roughly 3.1 points below Zhipu’s published number, according to a review published by The Planet Tools on August 3, 2026. The same review flags GLM-5.2’s SWE-bench Pro score as vendor self-reported and not part of Artificial Analysis’s own independent test battery, meaning the widely cited 62.1 figure has not been independently reproduced through that specific lab’s methodology, even though it appears consistently across benchmark write-ups from Apidog, Beam AI, and Fello AI.
None of this makes GLM-5.2 a weak model. A 77.9 on independently verified Terminal-Bench testing is still a serious agentic-coding result for an open-weight release. But the gap between vendor marketing and third-party verification is worth flagging, especially for a model still building trust outside China’s developer ecosystem.
What GLM-5.2 Actually Is
GLM-5.2 is Zhipu AI’s flagship coding model, built specifically for long-horizon, agentic software work rather than general chat. It uses a Mixture-of-Experts architecture with a parameter pool reported between 744 billion (per Codersera) and 753 billion (per MorphLLM), figures that fall within normal variance for how different outlets count active versus total parameters in an MoE system. It supports a 1 million token context window, a spec Zhipu markets heavily for tasks that require holding an entire large codebase in memory at once.
The model shipped under an MIT license, meaning the weights can be downloaded and self-hosted without the usage restrictions that come with many “open” releases from larger labs. Zhipu released GLM-5.2 on June 13, 2026, with public benchmark results circulating widely from June 16 onward, according to MorphLLM and Techtimes. That release cadence, a new flagship roughly every few months, has become one of Zhipu’s competitive advantages against slower-moving Western labs.
| Benchmark | GLM-5.2 | GLM-5.1 (prior) | GPT-5.5 | Claude Opus 4.8 |
|---|---|---|---|---|
| SWE-bench Pro (vendor) | 62.1 | 58.4 | 58.6 | 69.2 |
| Terminal-Bench 2.1 (vendor) | 81.0 | 63.5 | 84.0 | 85.0 |
| Terminal-Bench 2.1 (Artificial Analysis, independent) | 77.9 | n/a | n/a | n/a |
| FrontierSWE | 74.4 | n/a | 72.6 | 75.1 |
| GPQA Diamond | 91.2 | n/a | n/a | n/a |
Figures compiled from benchmark tables published by Codersera, Layer3 Labs, Apidog, Beam AI, and The Planet Tools between June and September 2026. Vendor rows reflect Zhipu’s published launch benchmarks. The Artificial Analysis row reflects independent third-party testing reported separately.
Pricing and How Developers Are Accessing It
Because the weights are open, developers have two paths into GLM-5.2: self-host it, or pay Zhipu for hosted API access and the bundled coding tools. Zhipu’s GLM Coding Plan starts at $18 a month for the Lite tier, $72 a month for Pro, and $160 a month for Max, with a 20 percent discount for annual billing, according to pricing details published by Fello AI. Raw API access runs approximately $1.40 per million input tokens and $4.40 per million output tokens, per MorphLLM’s pricing comparison against rival open model Kimi K3.
For a team already running self-hosted inference infrastructure, that per-token API price is close to irrelevant, since the whole appeal of an MIT-licensed 744-billion-parameter model is the ability to run it on owned or rented hardware without paying a vendor per call at all.
curl https://open.bigmodel.cn/api/paas/v4/chat/completions \
-H "Authorization: Bearer $ZHIPU_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "glm-5.2",
"messages": [{"role": "user", "content": "Fix the failing test in auth.py"}]
}'
That’s a representative request shape for Zhipu’s hosted API, not a guaranteed-working snippet. Endpoint paths and authentication details can change between releases, so developers should check Zhipu’s current documentation before integrating.
Where GLM-5.2 Sits in the Artificial Analysis Rankings
Zhipu’s own launch notes claimed the top open-weight score on the Artificial Analysis Intelligence Index at release in June. That claim has aged in the way most launch-day rankings do. A snapshot from ranking tracker Synthszr, dated early September 2026, places GLM-5.2 at a score of 51 on Artificial Analysis’s Intelligence Index v4.1, ranked sixth among open-weight models rather than first. Fello AI’s coverage attributes part of that slide to Kimi K3, a rival open-weight model that has since taken the open-weight performance lead.
That’s a normal trajectory for fast-moving open-weight releases: a model launches at or near the top of its category, holds that position for weeks, then gets passed as the next wave of open releases ships. GLM-5.2’s real staying power looks less like a world-beating open model and more like a consistently top-five one, with specific coding benchmarks where it still beats every closed OpenAI model available.
The China Factor: Data Jurisdiction Risk
Zhipu AI is a Chinese company, and GLM-5.2’s hosted API runs through infrastructure under Chinese jurisdiction. Techtimes flagged this directly in its June 17 coverage of the model’s launch, noting that while the open weights themselves carry no jurisdictional strings once downloaded, sending proprietary code or sensitive data through Zhipu’s managed API means that data is subject to Chinese data-handling rules rather than the data-residency terms enterprise buyers typically negotiate with US or European vendors.
For self-hosted deployments, that concern mostly disappears. A company running GLM-5.2 on its own GPUs never sends code to Zhipu at all. For teams using the hosted API or the GLM Coding Plan for convenience, it’s a real procurement question, and one that has shaped how some US and European engineering teams evaluate the model: strong benchmark results, but a harder sell to a security or legal team than an equivalent US-hosted option.
From GLM-4 to GLM-5.2: A Fast Release Cadence
Zhipu’s GLM line has moved through major version numbers faster than most Western labs move through minor ones. The jump from GLM-5.1 to GLM-5.2 alone added roughly four points on SWE-bench Pro and, per Codersera’s benchmark table, nearly 18 points on Terminal-Bench 2.1 (63.5 to 81.0) in a single release cycle. That kind of quarter-over-quarter gain is a big part of why Chinese open-weight labs, Zhipu among them, have become a fixture in benchmark conversations that used to be dominated entirely by OpenAI, Anthropic, and Google.
It also means benchmark snapshots age fast. Numbers accurate in June were already being revised by independent testers in August, and a new GLM release could reset the comparison again before the end of 2026. Treat any single benchmark table, including the one in this article, as a point-in-time reading rather than a permanent ranking.
Market Impact: Pressure on OpenAI’s Pricing Power
The practical effect of a free, downloadable model beating GPT-5.5 on a widely cited coding benchmark isn’t that GPT-5.5 users switch overnight. It’s pricing pressure. Every enterprise buyer now has a credible open-weight option to cite in a renewal negotiation, whether or not they actually plan to migrate. That leverage matters more in coding tools specifically, where SWE-bench Pro and Terminal-Bench scores get quoted directly in vendor sales decks and procurement comparisons.
It also changes the calculus for startups building AI coding products on top of a foundation model API. A team that would have defaulted to GPT-5.5 or Claude a year ago now has a self-hostable option that lands within striking distance of both on the specific tasks, multi-file bug fixes and terminal automation, that coding-agent products actually sell. That doesn’t threaten Claude Opus 4.8’s lead at the top end, but it does compress the middle of the market, where cost per token matters as much as raw benchmark score.
The Open-Weight Race: Kimi K3 and What Comes Next
GLM-5.2 doesn’t have the open-weight category to itself. Kimi K3, covered in MorphLLM’s direct comparison against GLM-5.2, has already overtaken it on Artificial Analysis’s broader index, according to Fello AI’s ranking commentary. That competition between Chinese open-weight labs, not just between China and the US, is arguably the more interesting story: multiple well-funded teams are now shipping frontier-adjacent coding models on a cadence measured in weeks, each one leapfrogging the last on a different benchmark.
For developers, that turnover is mostly good news. It means real competitive pressure on API pricing, faster iteration on agentic coding capabilities, and multiple viable self-hosting options instead of one. For anyone trying to pin down a fixed ranking of the best open model, that ranking has a shelf life of roughly one product cycle.
Why This Isn’t a Simple Beats-ChatGPT Story
Some coverage of GLM-5.2 has flattened its results into a single beats-every-ChatGPT-model headline. The actual picture is narrower and, honestly, more useful. GLM-5.2 beats GPT-5.5 specifically on SWE-bench Pro and runs close to even with it on Terminal-Bench, depending on whose harness you trust. It does not have a confirmed, independently verified head-to-head result against Google’s Gemini models on either benchmark. None of the sources reviewed for this piece, including Layer3 Labs, Codersera, and The Planet Tools, published a direct Gemini comparison on SWE-bench Pro or Terminal-Bench 2.1. Any claim that GLM-5.2 beats every model from OpenAI and Google on those specific tests overstates what’s actually been measured and published.
What is well supported, across nearly a dozen independent write-ups, is narrower and still notable: a Chinese open-weight coding model now beats OpenAI’s current flagship on real bug-fixing tasks, runs within a point of Claude Opus 4.8 on long-horizon engineering work, and does both under an MIT license anyone can download.
Historical Context: How Fast Open Models Closed the Gap
Two years ago, the gap between the best open-weight model and the best closed model on coding benchmarks was wide enough that few enterprise teams seriously considered self-hosting as a competitive alternative. That gap has been shrinking steadily, and GLM-5.2’s SWE-bench Pro result, ahead of a current-generation GPT model, is one of the clearest data points yet that the gap on specific, well-defined coding tasks has effectively closed for at least one major open release. The remaining gap to Claude Opus 4.8, roughly seven points on SWE-bench Pro and under a point on FrontierSWE, is now smaller than the gap between GLM-5.1 and GLM-5.2 was just one release ago.
| Spec | GLM-5.2 |
|---|---|
| Architecture | Mixture-of-Experts, ~744-753B parameters |
| Context window | 1,000,000 tokens |
| License | MIT (open-weight) |
| Release date | June 13, 2026 |
| Public benchmarking began | June 16, 2026 |
| Coding Plan pricing | $18/mo Lite, $72/mo Pro, $160/mo Max (20% off annual) |
| API pricing | ~$1.40 per million input tokens, ~$4.40 per million output tokens |
| Artificial Analysis Intelligence Index v4.1 | Score 51, rank #6 among open-weight models (early September 2026 snapshot) |
Specs and pricing compiled from Zhipu’s launch documentation as reported by MorphLLM, Fello AI, and Synthszr.
What GLM-5.2 Still Gets Wrong
Benchmark scores describe averages, not guarantees. Codersera’s guide to the model, one of the more detailed independent write-ups available, includes a dedicated limitations section alongside its benchmark tables, a structure most vendor marketing skips entirely. The gap between GLM-5.2’s vendor-reported 81.0 on Terminal-Bench and Artificial Analysis’s independently measured 77.9 is itself evidence that self-reported numbers deserve a discount, even from a lab that’s otherwise been transparent about publishing its methodology.
Running a 744-billion-parameter Mixture-of-Experts model also isn’t casual infrastructure. Self-hosting GLM-5.2 at usable speed requires serious GPU capacity, which is why most individual developers and smaller teams will access it through Zhipu’s hosted API or the GLM Coding Plan rather than standing up their own inference cluster, even with the weights freely available.
Five Predictions for the Rest of 2026
- Zhipu ships a GLM-5.3 or similar point release before year-end, continuing the roughly quarterly cadence seen since GLM-5.1.
- Artificial Analysis or a comparable independent lab publishes a direct GLM-5.2 versus Gemini comparison, closing the current gap in public data.
- At least one more Chinese lab, building on Kimi K3’s current open-weight lead, ships a model that beats GLM-5.2 on Terminal-Bench 2.1 before Q1 2027.
- Enterprise procurement teams increasingly ask coding-tool vendors to disclose whether backend inference runs through a China-jurisdiction API, treating it as a standard security questionnaire item rather than an edge case.
- OpenAI and Anthropic face renewed pricing pressure on coding-specific API tiers as open-weight alternatives get cited more often in enterprise contract renewals.
These are informed projections based on the release cadence and competitive dynamics described above, not confirmed roadmap items from any of the companies involved.
Frequently Asked Questions
What is GLM-5.2?
GLM-5.2 is an open-weight, MIT-licensed large language model built by the Chinese AI company Z.ai (Zhipu AI), focused on coding and long-horizon agentic tasks. It released June 13, 2026, with a Mixture-of-Experts architecture reported at roughly 744 to 753 billion parameters and a 1 million token context window.
Does GLM-5.2 beat GPT-5.5?
On SWE-bench Pro, a real-world software bug-fixing benchmark, GLM-5.2 scored 62.1 against GPT-5.5’s 58.6, a confirmed lead across multiple independent benchmark write-ups. On Terminal-Bench 2.1, results are closer and vary by which harness is used, with some tables showing GLM-5.2 slightly ahead and others showing it roughly even.
Does GLM-5.2 beat Claude Opus 4.8?
No. Claude Opus 4.8 still leads GLM-5.2 on SWE-bench Pro (69.2 vs 62.1) and Terminal-Bench 2.1 (85.0 vs 81.0 vendor-reported). The gap narrows to under a point on FrontierSWE, a long-horizon engineering benchmark, where Claude Opus 4.8 scored 75.1 against GLM-5.2’s 74.4.
Is GLM-5.2 free to use?
The model weights are free to download and self-host under an MIT license. Zhipu also offers a hosted GLM Coding Plan starting at $18 a month for the Lite tier, with Pro at $72 a month and Max at $160 a month, plus pay-per-token API access at roughly $1.40 per million input tokens and $4.40 per million output tokens.
Is it safe to send company code to GLM-5.2’s API?
Zhipu’s hosted API runs through infrastructure under Chinese jurisdiction, which Techtimes has flagged as a data-residency consideration for enterprise users sending proprietary code through the managed service. Self-hosting the open weights avoids that concern entirely, since no data leaves the organization’s own infrastructure.
Is GLM-5.2 the best open-weight AI model available?
Not currently. A September 2026 ranking snapshot from Synthszr places GLM-5.2 sixth among open-weight models on the Artificial Analysis Intelligence Index v4.1, with a score of 51. Rival open-weight model Kimi K3 has taken the top spot on that index, according to coverage from Fello AI.
How does GLM-5.2 compare to its predecessor, GLM-5.1?
GLM-5.2 gained roughly 4 points on SWE-bench Pro (62.1 vs 58.4) and nearly 18 points on Terminal-Bench 2.1 (81.0 vs 63.5) compared with GLM-5.1, according to benchmark tables published by Codersera and Beam AI, a large jump for a single release cycle.
Has GLM-5.2 been independently verified, or are the scores only from Zhipu?
Partially. Independent lab Artificial Analysis tested GLM-5.2 on Terminal-Bench 2.1 and measured 77.9, about 3 points below Zhipu’s published 81.0, according to a review from The Planet Tools published August 3, 2026. The same review notes that GLM-5.2’s SWE-bench Pro score has not been run through Artificial Analysis’s own independent test battery and remains a vendor-reported figure, even though it’s widely repeated across secondary coverage.
