AI · Daily · June 4, 2026

the new scaling law is 'good-enough model + smart routing'

A hybrid legal stack using GLM 5.1 as the main worker and Opus 4.7 as a specialist outperformed Opus alone on quality.[2] The same workflow hit an 18% all‑pass rate where pure Opus managed 14%, using the expensive model only intermittently instead of as the default.[3]

On the benchmark side, Microsoft’s MAI coding model scores 51% on SWE‑Bench Pro, while the MAI‑Thinking‑1 MoE variant reaches 53%.[4] The InstinctRazor‑Qwen3.5‑122B‑A10B‑GGUF MoE runs with roughly 8GB of active GPU VRAM by offloading experts to CPU and compresses to about 50GB on disk, making 'too big' models effectively local.[5]

At the small end, Qwen 3.5‑9B beating Gemma‑4‑12B‑it on 5 of 8 benchmarks shows that routing, training, and architecture now matter more than raw parameter count.[6]

gemma 4 and the 'good enough, everywhere' open stack

Gemma 4 12B is a unified, encoder‑free multimodal model that handles text, images, and audio with a 256K‑token context window.[7]

It runs on consumer laptops with 16GB of RAM, ships under Apache‑2.0, and explicitly allows commercial use, which is a rare combination at this capability level.[8]

Since launch it has reportedly passed 150 million downloads, and open‑source AI usage jumped about 500% in the same window, with many attributing the spike to Gemma.[9]

Community benchmarks and anecdotes still show Qwen models winning several coding and reasoning tests against Gemma‑4‑12B, and some users prefer Gemma mainly for retrieval and multilingual work rather than as a frontier replacement.[10]

DeepSeek’s open‑weight models being roughly 10–30× cheaper than US SaaS APIs, while also seeing 500% usage growth in two months, reinforces Gemma’s role as one of several strong open bases rather than a singular frontier.[11]

codegen crossed 50% of code; the bottleneck moved upstream

Some projects now report over 80% of their code being AI‑generated inside IDEs like VS Code, mostly via tools such as GitHub Copilot and similar assistants.[12]

Claude Code is rated S‑tier on the SemiAnalysis VibeMAX benchmark for coding, and Microsoft’s MAI coding model hits around 51% on SWE‑Bench Pro, moving 'good enough' coding into genuinely hard problem territory.[13]

OpenAI’s Codex has evolved from autocomplete into a desktop app that can spin up local servers, generate full apps, and even build and host live websites directly from prompts.[14]

Developers report the bottleneck moving from typing to reviewing, with complaints about mounting technical debt, tools that lose project context over time, and fears that heavy reliance on AI is hollowing out foundational skills in newer engineers.[15]

Costs are also biting, with Cursor users reporting up to $100 per day in usage, coding tools moving toward metered pricing, and proprietary models like Gemini 3.5 Flash coming in about three times more expensive than earlier versions.[16]

reasoning is sharper on benchmarks and still brittle in the wild

Claude 4.8 and Opus 4.6 show big gains on logic tasks, with Anthropic reporting 4× fewer unremarked flaws compared to earlier models.[17]

DeepSeek scored 103 out of 120 on a Putnam‑style math exam, indicating near‑top human undergraduate performance on that benchmark.[18] The LEAP framework separately reached a 70% one‑shot solve rate on complex formal math problems, far beyond earlier systems.[19]

Yet the same era of models—GPT‑4o, Claude 3.5, Gemini—still fail the classic Stroop attention test and routinely hallucinate, which is exactly the kind of brittle, pattern‑based 'reasoning' failure that keeps resurfacing.[20]

Outside benchmarks, peptide companies are reportedly gaming ChatGPT and Google AI Search by spamming biohacker subreddits to steer optimization advice, exploiting how models overweight recent and highly‑linked content.[21]

models are commoditizing; ecosystems and sentiment aren’t

ChatGPT is at roughly 1B monthly active users with 62% year‑over‑year growth, but Claude is smaller at 56M users and growing at about 640% annually.[22]

Claude is also being wired directly into curricula at places like U Chicago and U Penn, giving Anthropic a privileged path into the next generation of power users.[23]

On the quality front, users increasingly describe the big US models as 'commoditized', with many saying there is little difference in everyday outputs across Gemini, GPT‑class systems, and Claude.[24]

At the same time, open‑source AI projects have surged by about 500% in two months, DeepSeek’s open weights are reportedly 10–30× cheaper than proprietary APIs, and Chinese competitors are emerging at even lower price points.[25]

Routing layers like OpenRouter and the MCP AI Gateway, plus new low‑cost inference servers claiming 5–10× cheaper hosting, make it trivial to swap underlying models, pushing the real competition up into UX, support, and ecosystem lock‑in.[26]