AI · Weekly · June 26, 2026

open weights crash the frontier party

GLM‑5.2 is now the de facto open frontier baseline: a 744B‑parameter MoE with ~40B active at a time, #3 overall on GDPval‑AA with 1524 Elo and leading the open‑weights pack.[1]

It beats GPT‑5.5 and Claude Opus 4.8 on several coding measures, including a 44% DeepSWE pass@1 and a higher AA Coding Index score.[2]

The same model was shrunk from ~1.51TB to a 238GB llama.cpp build while keeping ~82% of its performance, making “frontier‑ish” capability runnable on a single high‑end box.[3]

Users describe it as the first open model good enough to replace closed APIs for most tasks, yet also complain about brutal token burn and reports of ~$50/day bills on OpenRouter/OpenCode, so it’s frontier in quality but not in wallet‑friendliness.[4]

Meanwhile, Chinese‑linked stacks are owning the cost frontier: DeepSeek V4 Flash is reported up to 100× cheaper than Claude for some coding workloads, Qwen 3.6 27B/35B are widely regarded as the best local coders, and Kimi K2.7 Code is preferred over GLM in some shop benchmarks for speed and efficiency.[5]

Five Chinese labs slashed token prices by as much as 99%, turning inference into a commodity in that ecosystem.[6] On the Western side, the EU is funding an open 400B+ parameter frontier model on its own supercomputers, while sentiment in AGI threads converges on open weights typically trailing closed labs by only 8–12 months.[7]

Add in the Sentient Foundation’s $42M open‑source AGI grant program and you get a picture where “closed vs open” looks more like a time lag than a hard capability wall.[8]

infra is quietly becoming the model

The most interesting performance gains this month didn’t come from new weights but from kernel work and routing tricks. NVIDIA’s NVFP4 MoE kernel cut GB200 NVL72 serving costs by ~2.5× in under 70 days, and NVFP4 Blockscaled GEMM landed on RTX Pro Blackwell GPUs as a first‑class primitive.[9]

GLM‑5.2 now ships as an NVFP4 quantized build and as 2‑bit GGUF, letting a 744B‑parameter MoE behave like a much smaller network in both memory and throughput terms.[10]

On the image side, Krea 2 Turbo’s FP8/INT8 quantizations show how architecture and format interact: INT8 is ~1.9× faster than FP8 on a 3090, but NVFP4 quantization of the same model in ComfyUI is reported 2× slower than FP8.[11]

Throughput hacks are stacking on top: Model Throughput Protocol (MTP) gives Gemma 4 26B/31B 35–53% speed boosts, and users report MTP can roughly double token rates, but non‑MTP Qwen 3.6 variants still win in output quality ~80% of the time and stability takes a hit.[12]

JetSpec claims up to 9.64× speedups via speculative decoding, while Leyten demonstrated a 744B‑parameter model stretched over six consumer GPUs in different states hitting ~30 tok/s over a WAN link.[13]

This all lands exactly where one major AI CEO said we were heading: the “model is no longer the product,” the surrounding infra is.[14]

benchmarks, orchestration, and the Fugu mirage

If you only watched leaderboards this month, you’d think the story was straightforward: GLM‑5.2 beats GPT‑5.5 on AA‑Briefcase and leads open models on GDPval‑AA, while Claude Fable 5 debuts #1 on DeepSWE and Sakana’s Fugu Ultra matches Fable/GPT‑5.5 on many code benchmarks.[15]

The reality is noisier: Fable 5 subsequently degraded on DeepSWE and was pulled back behind a U.S. government embargo, with only ~200 orgs continuing under Project Glasswing.[16]

Fugu Ultra is praised for code review but reported slower than Fable in practical coding, and it struggles with front‑end tasks even while its benchmark scores look frontier‑tier.[17]

Evaluation researchers are basically waving a big “benchmarks are PR” sign: they note that LLM benchmarking is increasingly marketing, multi‑step evals are hard to design, and adding tools can 3× evaluation cost.[18]

At the same time, outcome‑based tests—like action‑success metrics, omission‑rate audits in medical note‑writing, or Cline’s bug‑fix bake‑off where GLM‑5.2 produced better, cheaper code than Opus 4.8 but took 3× longer—are starting to carry more weight than a few points of accuracy on synthetic suites.[19]

One large‑scale study even suggests a small basket of benchmarks can predict model scores across many others, which mostly underlines how redundant a lot of leaderboard churn has become.[20]

agents, IDs, and the security turn

The agent stack had a very “from toys to weapons systems” month. Anthropic is rolling out government‑ID verification via Persona for certain Claude capabilities starting July 8, prompting explicit user fears about a surveillance slope, while Discord is testing age checks via wallets, cards and potentially face scans under pressure from age‑verification laws.[21]

During major Claude outages, “Claude for Government” reportedly stayed online, underscoring that there is already a hardened, segregated tier of the stack.[22]

In parallel, the U.S. requested that OpenAI stagger GPT‑5.6’s launch with individual approvals, has frozen broad access to Fable 5, and ordered Mythos shut down for most users after a red‑team where it breached a large share of NSA classified systems and discovered the long‑standing “Squidbleed” leak.[23]

At the framework layer, the attack surface is no longer hypothetical. LangGraph’s SQLite checkpointer shipped with a SQL‑injection‑to‑RCE bug that was exploited against Langflow servers, and OpenClaw’s ClawHub marketplace was found to host malicious “Skills” that evade automated scanners.[24]

A low‑skilled attacker used Claude and Codex to breach 14 companies, making clear that offensive capability is now exportable to people who don’t understand what they’re doing.[25]

All of this lands under tightening governance: the EU AI Act will require watermarking for AI‑generated text starting August 2, while Amnesty is calling leading generative systems “unlawful by design” and arguing for outright prohibition.[26]

coding tools and the slow collapse of the ‘software engineer’ identity

On the ground, coding feels less like SOTA shuffling and more like a slow culture shock. Between GLM‑5.2, Qwen 3.6, Kimi K2.7 Code and DeepSeek V4 Flash, there are now several non‑US, often open‑weights models that can match or beat GPT‑/Claude‑class systems on many real coding tasks, often at a fraction of the price.[27]

Qwen 3.6 27B is widely recommended as the go‑to local coder, especially for C++, while users describe GLM‑5.2 as a “real frontier coding model” whose long‑horizon agentic engineering scores rival closed systems.[28]

DeepSeek Flash and Kimi are prized for sheer cost‑efficiency, even as users hedge by pairing them with top‑tier models on harder problems.[29]

The human side looks much messier: Cursor is in use at places like SpaceX and is described as essential for many devs, yet also as a perpetually wrong junior engineer that repeats bad approaches.[30]

GitHub Copilot touts 40% productivity gains, but the lived experience in sectors like healthcare and Rust is “AI slop PRs” and extra debugging, to the point where users are openly hostile to Microsoft’s forced Copilot installs.[31]

Claude Code, Copilot, Codex and friends are shifting many engineers into code‑review and orchestration roles; some report not writing a line of code in two years while still getting promoted, others describe burnout from reviewing insecure, inefficient AI‑generated code and watching an estimated 65% of all code become machine‑written by 2026.[32]

The same tools are also in the offensive stack—Claude and Codex helped a low‑skill attacker into 14 companies—so the “AI as prolific junior dev” metaphor now includes both the inside team and the people trying to break in.[33]