
Thinking Fast and Slow
There are two kinds of inference: someone is waiting, or nobody is watching. The alpha lives at the ends. The open stack serves the middle.
Kimi K3 is #1 on Design Arena: Fable-tier, served at 100 tok/s. GLM-5.2 is Opus-tier with 1M context. Both on kernels tuned for codegen, plus the open-weight lineup: Qwen, MiniMax, DeepSeek.

Specialized models: WarpGrep for search. Fast Apply for applying edits at 10,500 tok/s. Compact at 33k tok/sec.

OpenAI-compatible at api.morphllm.com. In Claude Code, Cursor, and Codex viaMCPplus the AI SDKand
Deploy Morph on your own infra - on-prem or cloud.
Flexible, high-capacity rate limits.
99.9% uptime SLA with top-tier support .
Ready-to-sign agreements for enterprise compliance.

There are two kinds of inference: someone is waiting, or nobody is watching. The alpha lives at the ends. The open stack serves the middle.

Z.ai's GLM-5.2 lands within a few points of Claude Opus 4.8 on agent benchmarks at a fraction of the cost. Open weights don't serve themselves — here's how we make it fast, and how to reach it on Morph and OpenRouter.

Claude Code cost climbs because the whole agent loop runs on one frontier model. Here is what you actually pay for, why spend caps do not help, and how per-call model routing keeps the bill flat while token usage keeps growing.

An agent loops, a user gives up, a jailbreak slips past, and the trace still says 200 OK. The failures that matter to agents are semantic, and your logs can't see them. Reflexes are small classifiers that read every turn and label the ones that broke: frustration, jailbreaks, loops, policy violations. Eight ship out of the box, one API call, under 90ms. Train your own in an hour.

Reflexes are a new way to run LLM-as-a-judge at scale - by reusing the compute used for other reflexes

How we serve open source models to be the best in class at coding agent workloads.
Critical takes on the latest in codegen.