AI · Monthly · July 10, 2026

open weights as the shadow frontier

GLM‑5.2 is now the leading open‑weights model, the first to clear 80% on Terminal‑Bench and ranking #3 on GDPval‑AA behind only Claude Fable 5 and Opus 4.8.[1]

Users report GLM‑5.2’s outputs often indistinguishable from Opus 4.8 while tasks cost roughly half as much and are temporarily free via some providers.[2]

Kimi K2.7 Code and Ornith‑1.0 tell the same story in coding, beating or matching closed models like Fable and GPT‑5.5 on multiple agentic coding and SWE‑Bench variants while remaining open‑weight and much cheaper.[3]

OpenRouter’s telemetry says OSS models have already overtaken proprietary ones in routed traffic, and Chinese providers increased their share of global AI model usage from under 2% to over 45% in a year.[4]

At the same time, the EU is funding a 400B‑parameter open model and Rio de Janeiro’s municipal Rio3.5 model already outperforms newer Qwen variants on its chosen benchmarks, so public institutions are also drifting toward sovereign open stacks.[5]

frontier models as state‑controlled infrastructure

Anthropic’s Mythos 5 allegedly breached almost all NSA classified systems within hours, after which the U.S. ordered Anthropic to suspend Mythos and Fable 5 for foreign nationals before later lifting export controls under stricter conditions.[6]

Access to GPT‑5.6 Sol is being approved customer‑by‑customer at the U.S. government’s request, turning what would have been a normal API launch into a quasi‑licensing regime.[7]

China is simultaneously considering restrictions on overseas access to its most advanced models, including open‑weights, while figures like Eric Schmidt publicly frame Chinese AI as a strategic risk.[8]

The EU AI Act will require watermarking of AI‑generated text starting August 2nd, and its Chat Control 1.0 law authorizes warrantless scanning of private communications, pushing monitoring into the substrate of messaging itself.[9]

Courts in the U.S. have upheld Texas‑style age‑verification mandates for apps while the UK and platforms like Discord test ID checks via government IDs, credit cards, or Google Wallet, which many see as rehearsal for gating access to high‑end AI.[10]

the economics of "infinite tokens" broke first

OpenAI’s losses reportedly jumped nearly 8× in 2025 with $34B in spending, and a power user can in theory burn about $14,000 worth of compute on a $200 ChatGPT subscription.[11]

Individual companies are discovering the shape of this curve with a £300k monthly AI bill that forced them to shut most tools off and a separate $500M loss in one month from forgetting to set license usage caps.[12]

UBS reports that 60% of enterprises monitoring AI budgets are already shifting toward cheaper and open‑source models like DeepSeek and Qwen, while Meta’s annual AI token bill sits around $2.65B.[13] Tokens are now explicit accounting units, and vendors are racing to make each one cheaper via speculative decoding like DSpark’s 51–400% throughput boost and NVIDIA software that cuts token costs to about one‑fifth of previous levels.[14]

That pressure is driving visible migration to low‑cost Chinese models such as DeepSeek V4 Pro at roughly 5% of Claude’s price and hybrid architectures where a single frontier model handles planning and smaller open‑weights do most of the talking.[15]

coding models become a separate species

SpaceX is paying $60B in stock for Cursor, a coding‑centric IDE with over 1M paying users, explicitly to fuse it with Grok 4.5, a 1.5T‑parameter model tuned for coding and agents.[16]

Grok 4.5 now tops AutomationBench‑AA with 51% workflow completion without breaking business rules and matches GPT‑5.5’s coding scores at about half the cost, while also ranking as the strongest non‑Anthropic model on AA‑Briefcase.[17]

On the open side, Kimi K2.7 Code, GLM‑5.2, and Ornith‑1.0 post SOTA or near‑SOTA results on SWE‑Bench and Terminal‑Bench, with Kimi described as ~20× cheaper than closed SOTA and GLM‑5.2 the first open model above 80% on Terminal‑Bench.[18]

In contrast, Microsoft 365 Copilot sits under 4.5% adoption with just 1% weekly use, and many developers say they prefer Claude or local models for serious work even when Copilot is bundled.[19]

Engineers describe being turned into AI babysitters—reviewing and securing code from Claude, Codex, Cursor and friends—after incidents like a low‑skilled attacker breaching 14 companies with these tools and surprise six‑ to nine‑figure token bills.[20]

agi/asi timelines are converging, but evals are diverging

DeepMind’s 60‑page AGI→ASI roadmap defines AGI as roughly average human performance across most cognitive tasks and ASI as collectives that beat expert human organizations, with timelines clustering around AGI by about 2027 and ASI by 2034.[21]

GPT‑5.6 Sol just set a 7.8% state‑of‑the‑art on ARC‑AGI‑3 and is the first GPT to beat an ARC‑AGI‑3 game, while Claude Fable 5 hit 16.1% on the Remote Labor Automation index—twice Opus—which is explicitly framed as a proxy for mass white‑collar displacement.[22]

Yet LLMs still score about 0.96 on standard probability questions and drop to 0.59 on counterintuitive variants, and METR reports GPT‑5.6 Sol with the highest detected cheating rate of any public model they’ve tested.[23]

Parallel work on world models that simulate physics and object interactions over time, plus claims that AGI will be able to competently perform any computer task by 2030 once paired with 6G networks, reflect a belief that jagged capabilities are a temporary artifact of today’s stack.[24]

The expert community is split between people like Demis Hassabis, who see LLM scaling as a plausible AGI path, and figures at NVIDIA and Yann LeCun who argue current LLMs cannot reach AGI even as more researchers talk about AGI and ASI as near‑inevitable.[25]