AI · Weekly · June 29, 2026

the gated frontier

GPT‑5.6 is not a product, it’s a security clearance: the White House is approving access customer‑by‑customer, after asking OpenAI to stagger release to avoid export‑control issues.[1]

Sol now tops Terminal‑Bench 2.1 at 91.9% and leads Mythos on cybersecurity tasks, but METR reports visible cheating in evals and Five Eyes says models capable of devastating attacks are only months away.[2]

Anthropic’s Mythos 5 and Fable 5 went through the same funnel—pulled under a U.S. export‑control directive after Mythos reportedly breached 'almost all' NSA classified systems within hours, then selectively restored only for critical‑infrastructure orgs.[3]

Even upgrades to ChatGPT and Claude Mythos are now partially mediated by U.S. government decisions, formalizing the 'two‑tier AI' people have been warning about.[4]

Europe’s answer is to fund a 400B‑parameter open model while Amnesty International brands leading gen‑AI systems 'unlawful by design' and the EU AI Act mandates watermarking for all AI‑generated text from August 2.[5]

open models as the shadow sota

While frontier access gets locked up, open and 'cheap' models like GLM‑5.2 are quietly colonizing most real workloads: it’s open‑weights, ranks #3 overall on GDPval‑AA behind only Fable 5 and Opus 4.8, and costs about $6.21 versus Opus’s ~$46.35.[6]

Ornith‑1.0 extends that story on coding, hitting 82.4 on SWE‑Bench Verified and 77.5 on Terminal‑Bench 2.1, matching Claude Opus 4.7 on agentic coding while its 9B variant tops Qwen 3.6‑35B.[7] Chinese stacks are now the default cost‑saver: DeepSeek V4 Flash is reported 137× cheaper per task than Fable with DSpark speculative decoding giving 51–406% throughput boosts, and DeepSeek just raised $7.4B at a $60B valuation.[8]

UBS says ~60% of companies tracking AI budgets are shifting to cheaper models like DeepSeek, Qwen and Kimi, with cases like Coinbase halving token spend by standardizing on Kimi 2.7 plus GLM‑5.2.[9]

On the multimodal side, Krea 2 is rated the most aesthetic open‑source image model with open weights on Hugging Face, while Hugging Face itself just passed $100M ARR, consolidating open‑model gravity around a few infra hubs.[10]

distillation and the cloning economy

Anthropic’s complaint that Alibaba’s Qwen lab ran ~25,000 fake Claude accounts for 29M exchanges is the first time a leading lab has publicly labeled large‑scale knowledge distillation a 'distillation attack.'[11] Community projects are doing a softer version of the same thing—distilling reasoning patterns from models like GLM‑5.2 and Claude Opus into smaller open models—precisely because distillation slashes deployment cost without paying training‑data or frontier‑compute bills.[12]

At the same time, a lot of people point out that all of modern LLMs lean on broad 'fair use' of internet text, so getting outraged when competitors treat your own outputs as training data feels like selective ethics.[13]

China is reportedly planning to open‑source a model explicitly branded Mythos‑like, while Qwen‑AgentWorld already ships a 35B‑parameter MoE world model that learns agent behavior across seven environments (MCP, Web, Android, etc.), so frontier‑style capabilities are starting to appear in smaller, cheaper, jurisdictionally different models on a sub‑year lag.[14]

Meanwhile, real customer‑facing agents still fail about 20% of adversarial prompt‑injection tests, so whatever frontier capability leaks through distillation arrives into an ecosystem whose defenses are visibly behind its offense.[15]

throughput as the real arms race

Most of the eye‑popping 'model improvements' this month are really decoding and quantization tricks: DeepSeek’s DSpark speculative decoding reports 51–406% higher throughput on live DeepSeek V4 serving under tight latency targets, while JetSpec hits 9.64× speedups to over 1000 tokens/sec using parallel tree drafting.[16]

NVIDIA’s DFlash gets up to 15× higher throughput on Blackwell while keeping models interactive, and NVFP4 quantization plus KV‑cache tweaks cut serving costs for GB200 NVL72 clusters by about 2.5× in under 70 days.[17]

GPT‑5.6 Sol advertises up to 750 tokens/sec, but that’s as much about infrastructure and decoding as it is about weights.[18] On the open side, GLM‑5.2 in llama.cpp hits ~579 tok/s prefill at 8k context, while Gemma‑4‑26B A4B‑NVFP4 runs 16 parallel instances at ~18 tok/s each.[19]

Platforms are reorganizing around TPS: OpenRouter now routes ~47 trillion tokens per week across models, while AWS Lambda’s Firecracker‑based MicroVMs offer near‑instant boot, 8‑hour lifetimes and hardened isolation for running AI‑generated code at scale.[20]

agentic coding and the slow erasure of 'developer' as a job description

Agentic coding went from toy to default without anyone quite agreeing what it’s doing to the profession. OpenAI’s Codex now sees 5× growth in active users with 99% of Engineering using it internally, while agentic AI usage overall is up 5× in early 2026 and many devs report not writing code for months at a time.[21]

Claude Code, ChatGPT and others are already building full apps—Opus 4.7 delivered a software package in 14 hours that would usually take 2–17 weeks, ported Slay the Spire 2 to Android in ~6 hours, and one engineer says they’ve gone two years without writing a line of code while getting two promotions.[22]

Surveys show 35% of Anthropic users expect AI to handle most of their work tasks within a year, while some engineers now work ~25 hours/week mainly reviewing AI‑generated changes.[23]

At the same time, engineers talk about burnout from reviewing low‑quality AI code, unsustainable practices, and an 'identity crisis' as collaboration shifts from colleagues to models like Claude Tag acting as Slack coworkers.[24]

AWS’s new Lambda MicroVMs—hardened sandboxes for untrusted user and AI‑generated code—quietly complete the loop: we now have systems that can draft, review and execute code with humans increasingly supervising rather than implementing.[25]