GLM-5.3 is builders' new value coding pick, at ~1/10 frontier cost
Z.ai's 744B coder tops open-weight benchmarks and undercuts US frontier models 10x; open weights are expected within days and it's flooding Hacker News.
Copy markdown
Why it's on every builder's feed
GLM-5.3 owns the Hacker News front page today: 'beat Anthropic/OpenAI for 1/5 the cost,' plus a builder who finished a $266, four-model project with it in a day. The pitch is frontier-class agentic coding without the frontier bill.
The benchmark jump Z.ai claims
Against GLM-5.2, Z.ai reports Terminal-Bench 3.0 climbing 4.6 to 28.3, DeepSWE v1.1 46.2 to 66.9, and CyberGym 77.2% to 84.5%, calling 5.3 the top open-weight coding model. These are vendor numbers, so retest on your own harness.
Open weights land within days
The 744B mixture-of-experts (~40B active) is due to ship in stages after safety review, targeted around Aug 28, with an MIT license expected per GLM-5.1/5.2. Until then it is API and Coding-Plan only.
What you'll need to self-host
This is not a laptop model: an fp8 build (~750GB) fits a single 8xH200 node, and a 2-bit quant (~245GB) targets 256GB-class machines, with a 256K-token context. Serious homelab or one rented node territory.
It shipped a 2,436-bug ledger
Z.ai says GLM-5.3 found 2,436 vulnerabilities across 269 open-source projects, undiscovered for 26.6 years on average, but only 53 are public with 2,383 under embargo. Your dependencies may be on that list.
The price gap, in numbers
API pricing runs at roughly a tenth of US frontier rates (GLM-5.2 was $1.40 in / $4.40 out per million), and Coding-Plan tiers are $18 Lite, $72 Pro, and $160 Max a month. That price lever is what's driving the switching chatter.