Frontier lab intelligence

A cross-card read of GPT-5.6, GLM-5.2 and Kimi-K2-Thinking: the closed leader publishes zero capability comparisons while the two open labs publish incompatible ones — so the leaderboard everyone wants cannot be built from primary sources. That void is the eval product.

Master benchmark inventory of the GPT-5.6 system card with web-verified cross-vendor comparison and the eval-shift Sankey — every claim tied to a page.

Zhipu’s MIT-licensed flagship is in the Opus/GPT-5.5/Gemini tier on agentic SWE while trailing on broad knowledge — and its own cross-vendor table shows why "the harness is the benchmark".

Moonshot’s open 1T-MoE thinking agent holds the top of agentic web research (BrowseComp SOTA, 200–300 tool calls without drift) while trailing closed frontier on raw coding — a distinct open frontier from GLM-5.2’s agentic engineering.

What 128 open eval-relevant roles reveal about frontier eval investment, for two audiences: people who want to get hired, and people who want to sell to the labs.

What 715 open infrastructure/systems roles reveal about the GPU buildout — physical first — for two audiences: people who want to get hired in infra, and people who want to sell infra to the labs and neoclouds.

What 127 open safety/alignment/red-team roles reveal about the safety org (OpenAI out-hires Anthropic in raw count), for two audiences: get hired into safety, and sell safety tooling to the labs.

What 86 open human-data / annotation / data-quality roles reveal about the fuel layer — the clearest "sell to the labs" buy signal, since labs structurally buy data rather than build it.

Anthropic is running three simultaneous buildouts: (1) physical compute and multi-cloud capacity, evidenced by a repeat cluster of data-center, TPU, and compute-financing roles ; (2) the commercial/enterprise layer of the Claude platform, evidenced by GTM, public-sector, regulate

Baidu's public footprint in this pack runs on two parallel tracks. First, a high-velocity open-weight release engine: the ERNIE/Qianfan model families plus PaddleOCR/PaddleX tooling are shipping continuously, spanning text, vision-language, image, video, and document-parsing mode

Amazon has collapsed its portfolio of high-end Nova models into a single flagship effort — Frontier Model Research (FMR) under Pieter Abbeel — while Nova Premier, Omni, Reel, and Canvas move to keep-the-lights-on status; only Nova 2 Lite/Sonic plus the Nova Forge customization an

Cohere's observable motion in this pack is a security-first enterprise platform play converging on three fronts: (1) retrieval and evaluation infrastructure — Embed 5, the RCP-nDCG LLM-judge metric, and Compass search ; (2) a deep multilingual data-and-models research line (Aya,

DeepSeek is running two parallel buildouts in this pack: a fast-moving open-source agent platform (DeepSeek Harness, Everything is a Plugin) that it is packaging into desktop apps and developer tooling , and a porting wave that brings its own training/inference kernels onto Huawe