How the math works: GLM 5.2 on the public API costs about 1/5 to 1/6 the per-token price of GPT-5.5 or Claude Opus 4.8 GD. On pure output, GLM-5.2 at $4.40 per million tokens against GPT-5.5's $30 is about 1/6.8 the cost OF. Cached input drops further to $0.26 per million tokens GF.
This is a notably different hardware strategy from the NVIDIA H100/H200 deployments common in the industry, and Featherless frames it as a way to avoid NVIDIA supply constraints I.
| Specification | Details |
|---|---|
| Architecture | 744B total parameters, Mixture-of-Experts (MoE) with ~40B active per token GS |
| Context window | Up to 256K tokens on public cloud; up to 1 million tokens on private cloud deployments FI |
| Quantization | FP8 by default (~750 GB VRAM for weights) FOR |
| Open weights | Yes — Z.ai released GLM 5.2 as open-weights under MIT license IDA |
| Training focus | Coding-first; specialized for software engineering tasks GDS |
At FP8 quantization, the model's weights require roughly 750 GB of VRAM, which is comfortably accommodated by the 1 TB total VRAM across 4× MI325X GPUs (4 × 256 GB), with headroom for KV cache at extended context lengths OR.
SWE-bench Pro (real-world software engineering):
Terminal-Bench 2.1 (agentic coding tasks):
GLM 5.2 ranks #2 on the Arena WebDev coding leaderboard according to Featherless F.
The cost advantage is where GLM 5.2 really stands out. At Z.ai's official API rates:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GLM 5.2 | $1.40 | $4.40 |
| GPT-5.5 | ~$5.00 | ~$30.00 |
| Claude Opus 4.8 | ~$8.00 | ~$40.00 |
On a realistic 3:1 output-to-input workload blend, GLM-5.2 lands near $3.65 per million tokens against GPT-5.5's roughly $23.75 — a ratio of about 1/6.5 O. Independent trackers list a lower median across providers serving the open weights (closer to ~$0.55 input and ~$1.85 output) DD.
The Featherless private cloud for GLM 5.2 is most compelling for:
Featherless also offers lower-tier flat-rate plans starting at $25/month for serverless access to smaller models, but the $7,500/month dedicated node is explicitly for teams needing sustained, high-volume inference on the full GLM 5.2 model F.
Featherless's new service offers a $7,500/month flat-fee private cloud for GLM 5.2 on 4× AMD MI325X GPUs, claiming 94% cost savings over proprietary APIs for high-volume agentic usage. The 744B MoE model beats GPT-5.5 on SWE-bench Pro at roughly one-sixth the per-token cost, making it a compelling open-weights alternative for organizations with heavy coding and agentic inference workloads. While Claude Opus 4.8 still leads on some benchmarks, the cost advantage of the GLM 5.2 + Featherless combination is large enough to make it a serious option for budget-conscious teams that don't want to compromise on performance.