Zhipu's 320B GLM-5.3-Flash runs 17x cheaper than full GLM-5.3 on coding tasks
Multiple independent benchmarks show a 320B open-weight model running on Chinese chips matching or beating closed frontier models on coding at a fraction of the cost, with detailed cost-quality tradeoff data now available across both the full model and its Flash variant. Bloomberg Intelligence has flagged Z.ai's unit economics as unsustainable given rising inference costs.
Together AI's 900-rollout DeepSWE test finds GLM-5.3-Flash costs 17x less than full GLM-5.3, with Flash giving up 5.6 pass@1 points that narrow to 2.6 at pass@4. A Blender architectural task via MCP showed comparable output from Flash for 16.7x less ($0.0526 versus $0.8807). Z.ai disclosed on August 26, 2026 that GLM-5.3-Flash is the model it had been previewing as Ox Alpha: a 320B-parameter MoE with 18 billion active parameters, MIT-licensed, natively multimodal, with a 1M-token context window, running on domestic Chinese accelerators. The full GLM-5.3, released August 14, 2026, scored 91.25% on KingBench 3, outperforming Opus 5 and Kimi K3, tied Fable 5 on pass@1 in DeepSWE testing at $3.99 per rollout versus $21.63, and trailed GPT-5.6 Sol by 3.7 pass@1 points while winning on pass@4 at roughly half the cost. Z.ai's own benchmarks show GLM-5.3 outperformed models from Anthropic and OpenAI on a cyber performance test. Bloomberg Intelligence called Z.ai's commercial footing unsustainable; GLM-5.3's release coincided with a 9% drop in Z.ai shares.
Three independent comparisons published August 28 put Z.ai's GLM-5.3-Flash at roughly 17x less cost than the full GLM-5.3, with the pass@1 gap narrowing from 5.6 to 2.6 points at pass@4 in a 900-rollout DeepSWE test and 16.7x less on a Blender MCP task.
A three-model price test placed Flash at $0.03 versus $0.29 for Claude Opus 4.6 on the same prompt. Rohan Paul found that wrapping full GLM-5.3 in the Atomic Agent harness roughly doubled token usage while adding only $0.77 to cost.
Blender MCP task: Flash produces comparable architectural output for 16.7x less than full GLM-5.3 with better prompt following
GLM-5.3-Flash became available on Databricks on August 27, adding a second major cloud platform alongside CoreWeave for Z.ai's 320B-parameter mixture-of-experts model. Databricks reports it delivers 10% higher quality than GLM-5.2 at one-tenth the cost on OfficeQA Pro v2 and adds multimodal support for reading figures and verifying web pages. A Hugging Face analysis the same day found Chinese labs are releasing open-weight models several times larger than any open models US labs have shipped publicly in 2025-2026.
GLM-5.3-Flash launches on Databricks, delivering 10% higher quality than GLM-5.2 at one-tenth the cost on OfficeQA Pro v2
Together AI published 904-rollout DeepSWE comparisons August 27 showing Z.ai's GLM-5.3, released August 14, tied Fable 5 on pass@1 and beat it on pass@4 at 5.4x lower cost; KingBench 3 placed it at 91.25%, ahead of Opus 5 and Kimi K3. Bloomberg Intelligence analyst Robert Lea said Z.ai 'remains on a completely unsustainable commercial footing,' with shares down 9% on the launch day, and founder Tang Jie argued post-training is now the highest-value scaling dimension.
CoreWeave announces GLM-5.3-Flash coming to its Serverless Inference platform
Together AI: GLM-5.3 trails GPT-5.6 Sol by 3.7 pass@1 points but wins pass@4 at half the cost; GLM-first cascade hits 85.9%
Tang Jie argues parameter count is incomplete as a capability metric and frames GLM-5.3 as a controlled post-training experiment
Hugging Face analysis finds Chinese labs releasing open-weight models several times larger than any US labs have shipped publicly in 2025-2026
I send a short email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free