AI Engineering Signal #63
GLM-5.2 tops GPT-5 on Artificial Analysis's new agentic knowledge-work eval (AA-Briefcase) and leads coding benchmarks
Signals
GLM-5.2 tops GPT-5 on Artificial Analysis's new agentic knowledge-work eval (AA-Briefcase) and leads coding benchmarks
open-weight model routing decisions made against proprietary APIs need to be revisited now.
Web
DeepSeek-V4 targets million-token context at high efficiency
long-context retrieval cost assumptions built around current models need repricing.
ArXiv
Anthropic Mythos access persists for ~200 companies after US shutdown order
any pipeline relying on Mythos needs an audited fallback before enforcement tightens.
Web
OSS models overtook proprietary in OpenRouter market share over last three months
cost and routing models built around proprietary API dominance are now structurally stale.
Web
Zero-Touch OAuth spec lands for MCP
enterprise agent deployments can now delegate auth without manual credential injection, unblocking production rollouts.
Web
Human genome's 3D physical folding resists AI modeling
protein-structure playbook does not transfer directly; genomic AI pipelines need architecture rethinking.
Web
The Take
Open-weight models are no longer a cost compromise — they are the benchmark leaders, and the policy layer (export controls, access shutdowns) is now the primary risk variable, not model capability. Build your routing and compliance gates accordingly.
Subscribe
Related Signals