← All posts
·3 min readModel PerformanceAPI Costs

GLM-5.2: The Open-Weight Coding Gap Just Got Uncomfortably Small

Computer screen displaying lines of program code
Photo by Rahul Mishra on Unsplash

On June 13, Zhipu AI (operating internationally as Z.ai) released GLM-5.2: a Mixture-of-Experts model with roughly 750B total parameters (~40B active per token), a 1M-token context window, up to 131K output tokens, and — the part that matters most — MIT-licensed open weights on Hugging Face and ModelScope.

Three weeks in, the picture is clear enough to write down: this is the strongest open-weight coding model available today, and it lands close enough to the closed frontier that the price difference is now the headline.

The numbers, with appropriate skepticism

From Z.ai's own announcement (so treat these as vendor-reported until independently reproduced):

Benchmark GLM-5.2 GPT-5.5 Claude Opus 4.8
SWE-bench Pro 62.1 58.6 69.2
Terminal-Bench 2.1 81.0 84.0 85.0
FrontierSWE 74.4% 72.6% 75.1%
GPQA-Diamond 91.2 — 93.6

The honest read: GLM-5.2 beats GPT-5.5 on some long-horizon coding benchmarks and trails Claude Opus 4.8 on essentially all of them. VentureBeat's framing — beats GPT-5.5 on multiple long-horizon coding benchmarks at about one-sixth the cost — is the fair one. "Beats the frontier" would not be.

But look at the trajectory. GLM-5.1 scored 58.4 on SWE-bench Pro; 5.2 hits 62.1 six months later. On FrontierSWE, Z.ai reports 74.4% against Opus 4.8's 75.1% — within a point. The open-vs-closed gap on agentic coding used to be measured in tiers. It's now measured in single digits, sometimes decimals.

The price is the story

Zhipu's first-party API lists GLM-5.2 at $1.40 per 1M input tokens and $4.40 per 1M output tokens, with cached input around $0.26/1M at some hosts. Because the weights are MIT-licensed, third-party hosts compete on price immediately: OpenRouter currently shows GLM-5.2 at $0.728/$2.288 per 1M under a promotional discount, and providers like SiliconFlow and DeepInfra serve it too.

This is what open weights actually buy you as an API consumer — not the romance of self-hosting, but hosting competition. Closed models have exactly one seller. GLM-5.2 has many, and the price floor drops accordingly.

About self-hosting: yes, it's permitted (MIT, no regional restrictions, commercial use fine), and vLLM/SGLang support landed at release. But this is a ~750B-parameter MoE. One local-AI guide bluntly called it "effectively cloud-only at home." Self-hosting makes sense for large teams with GPU fleets and data-residency requirements; for everyone else, the leverage is the multi-provider market, not your own rack.

Architecturally, the 1M context is more usable than the spec sheet suggests: Z.ai's IndexShare scheme reportedly cuts per-token compute ~2.9× at full context length, which is why long-context pricing hasn't exploded the way it did for earlier 1M-context models.

What this means for developers

  1. Recalibrate your "good enough" tier. If your workload is agentic coding — issue-to-patch, terminal automation, long-refactor loops — a model at ~$1.40/$4.40 now delivers what cost 5–6× more a year ago. The top closed models are still better; they're just no longer the only defensible choice.
  2. Vendor benchmarks need a discount factor. Z.ai's numbers were run on their harness. Independent leaderboards (Scale's SWE-bench Pro board, Terminal-Bench's official one) are where these claims get stress-tested — check those before migrating anything.
  3. Expect the closed labs to respond on price. Every time an open-weight model lands within a few points of the frontier, retail pricing on the tier below the flagship gets harder to justify. That pressure is the real gift of open weights, even if you never send Zhipu a single token.

This churn — new model, new price floor, every few weeks — is exactly why we built tier-based routing at TierUp: you pick the performance level, and repricing events like this one become our problem instead of yours.

Sources

Featured on Fazier