The frontier isn’t defined by the smartest model anymore, it’s defined by export licenses, VRAM limits, and how cleverly you stitch together mid-size models and runtimes. The US effectively turned Fable/Mythos-class systems into controlled munitions just as the ecosystem quietly split into 200B monsters and aggressively optimized 30B-ish workhorses running on everything from H100s to bags of old Pixels.
Security and reliability are lagging badly, so the real game is shifting to orchestration: which stacks can make this unstable mix actually usable in the real world.
Key Events
/The US formally classified frontier models like Fable 5 and Mythos 5 as controlled exports, adding nationality-based access restrictions.
/Fable 5 was launched on June 9 and pulled three days later after jailbreak reports and direct US government intervention.
/Deepseek V4 Flash, a ~200B-parameter model with image reading, launched with pricing around $224 for 1B input and 300M output tokens per month.
/Kimi K2.7 Code was dynamically 2-bit quantized from 1TB to 325GB, still achieving over 40 tok/s locally and up to ~180 tok/s on median inputs.
/vLLM added a new streaming parser for Qwen3+ that fixes mid-turn stopping and failed streaming tool calls in agentic workflows.
Report
Frontier AI did not slow down this month; access to it did. The loudest signal is that export controls, VRAM budgets, and model-mixing layers now decide who actually touches near-frontier capability.
frontier power is now an export license
The US formally classified frontier models like Fable 5 and Mythos 5 as controlled exports, adding nationality-based access rules on top of normal API gating.
Fable 5 launched June 9 and was pulled three days later after a reported jailbreak and direct outreach from Amazon’s CEO to the US government about its danger.
Cybersecurity veterans publicly argued the model was not uniquely risky, even as the Intelligence Frontier chart literally moved backward when Fable went offline and enterprises reported a trust hit.
Developers outside the US now talk about being told to 'go make AGI for other countries,' and there is open speculation about a black market for American AI, turning access into a geo-strategic commodity rather than a benchmark.
compute is breaking in weird directions, not just more H100s
Hetzner hiked US server prices 2.5× on AI demand, while users complain that training large models on a single GPU is impractical and that rising prices for high-VRAM cards are delaying hardware upgrades.
At the same time, the University of California built a low-carbon datacenter from 2,000 retired Pixel phones, and modern smartphone CPUs reportedly beat some multicore servers on single-core performance.
Tensordyne’s 3nm Napier AI chip promises 13× higher token throughput than NVIDIA’s Blackwell, and separate Logarithmic AI chips are advertised as significantly more efficient than Blackwell for inference.
Developers are also pushing decentralized training schemes like Bitcoin mining and leaning on FP8 quantization to offload work, splitting the ecosystem between centralized GPU clusters and increasingly odd, heterogeneous compute fleets.
the model race bifurcated into 35B hustlers vs 200B giants
New releases now cluster either in the 25B–35B range or above 200B parameters, with almost nothing fresh in the 100B–120B band that used to define 'frontier' a year ago.
Deepseek V4 Flash, a ~200B model that recently gained image reading, is aggressively priced at about $224 for 1B input and 300M output tokens per month, pulling the high end toward cheaper tokens instead of just bigger scales.
On the efficiency side, Kimi K2.7 Code shrank from 1TB to 325GB via dynamic 2-bit quantization and still runs at over 40 tok/s on 330GB rigs and around 180 tok/s on median-length inputs.
Microsoft’s 4B code model cuts token costs 10–50%, while Qwen 3.6‑27B and Gemma 4‑31B show that well-tuned mid-sized models can rival larger systems for many coding and agentic tasks, especially when run locally on dual-3090-style setups.
local runtimes and meta-routers are quietly becoming the real platform
Power users are consolidating on llama.cpp for local inference, reporting roughly 28–30 tok/s on an RTX 5080 plus 2×V100 and about 57 tok/s on an RTX 5090, while criticizing Ollama’s slower, more constrained, and licensing-messy experience.
On the server side, vLLM is becoming the default multi-user stack, with a new Qwen3+ streaming parser fixing mid-turn stops and tool-call failures and tight FP8/H100 integration, even as a bespoke Rust+GPU engine has outperformed it on Nemotron‑30B throughput.
Routers like OpenRouter’s discontinued Fusion Panel, and its local clone OpenFusion, compose multiple models via YAML graphs and claim near-Fable-5 quality at half price, putting orchestration and routing logic on the same footing as any individual model.
safety and security are failing in exactly the boring ways infosec predicted
Researchers showed that user-generated text as short as 13 words can steer systems like ChatGPT and Google’s AI search, demonstrating how brittle current guardrails are against prompt-level manipulation.
An attack turned Microsoft 365 Copilot into a one-click data-theft tool, making it trivial to exfiltrate corporate data from within the official productivity stack.
Fable 5 was reportedly jailbroken almost immediately after launch, even as Anthropic disputed the claims, and the US government was alerted to the exploit as part of the export-control saga.
In parallel, the OpenMythos cybersecurity LLM arrived trained on real CVEs, ArXiv papers, and vulnerable code, while Rust’s stricter treatment of unsafe functions is contrasted with C/C++ segfaults that rarely even get recorded as CVEs.
The LessWrong-influenced safety discourse and rising memory-safety costs, now exceeding half of smartphone hardware bills, show security and reliability being dragged into the core of AI tooling rather than left as afterthoughts.
What This Means
The interesting competition has shifted from 'who has the single strongest model' to who can stitch together export-gated frontier systems, aggressively optimized mid-sized models, and shaky security into something that actually runs on real hardware in a specific jurisdiction.
On Watch
/Europe is exploring training its own frontier AI model using EU computing resources, despite skepticism about whether it can marshal enough capital and coordination to matter.
/Atom Exhaustion reportedly accounts for about one third of all CVEs, an oddly specific pathology that could drive niche tooling and model specializations around regex- or parsing-heavy software.
/The 'Wall Street of AI Agents' project, where four SLM-based traders compete to grow $10,000 portfolios without going bankrupt, is quietly becoming a benchmark for small-model decision making.
Interesting
/A researcher claims to have bypassed safety layers in Claude Fable 5, exposing its 120,000-character system prompt.
/The Mirage paper's approach to using diffusion latent tokens could revolutionize video world models by enhancing efficiency.
/The trend towards a renters model for computing hardware is likely to benefit large corporations, potentially leading to a monopoly on computing power by hyperscalers.
/The debate over open-sourcing AI model weights is complicated by the size of frontier models, which may not be usable by average developers, potentially stalling open-source contributions.
/There is a growing concern about the ethical implications of using outputs from frontier models in open-source projects, especially regarding access restrictions by proprietary models.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/The US formally classified frontier models like Fable 5 and Mythos 5 as controlled exports, adding nationality-based access restrictions.
/Fable 5 was launched on June 9 and pulled three days later after jailbreak reports and direct US government intervention.
/Deepseek V4 Flash, a ~200B-parameter model with image reading, launched with pricing around $224 for 1B input and 300M output tokens per month.
/Kimi K2.7 Code was dynamically 2-bit quantized from 1TB to 325GB, still achieving over 40 tok/s locally and up to ~180 tok/s on median inputs.
/vLLM added a new streaming parser for Qwen3+ that fixes mid-turn stopping and failed streaming tool calls in agentic workflows.
On Watch
/Europe is exploring training its own frontier AI model using EU computing resources, despite skepticism about whether it can marshal enough capital and coordination to matter.
/Atom Exhaustion reportedly accounts for about one third of all CVEs, an oddly specific pathology that could drive niche tooling and model specializations around regex- or parsing-heavy software.
/The 'Wall Street of AI Agents' project, where four SLM-based traders compete to grow $10,000 portfolios without going bankrupt, is quietly becoming a benchmark for small-model decision making.
Interesting
/A researcher claims to have bypassed safety layers in Claude Fable 5, exposing its 120,000-character system prompt.
/The Mirage paper's approach to using diffusion latent tokens could revolutionize video world models by enhancing efficiency.
/The trend towards a renters model for computing hardware is likely to benefit large corporations, potentially leading to a monopoly on computing power by hyperscalers.
/The debate over open-sourcing AI model weights is complicated by the size of frontier models, which may not be usable by average developers, potentially stalling open-source contributions.
/There is a growing concern about the ethical implications of using outputs from frontier models in open-source projects, especially regarding access restrictions by proprietary models.