Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
In an industry first, the US government has compelled Anthropic to shut down global access to both Fable 5 and Mythos 5, citing jailbreak risks capable of enabling misuse at scale. This is the first known instance of the US government ordering an AI lab to withdraw a deployed model — a regulatory precedent with sweeping implications for every future frontier release.
Anthropic launched Fable 5 on June 9th as the first publicly accessible version of its Mythos model family, which had been partner-restricted since April due to its ability to autonomously discover and chain zero-day exploits. The public build included hardcoded blocks in cybersecurity, biology, and chemistry, falling back to Claude Opus 4.8 in those domains — a selective rollout strategy the government apparently found insufficient.
OpenAI's GPT-5.4 Thinking model scored 83.0% on GDPVal, a benchmark measuring real economic utility rather than academic performance, placing it at or above human expert level on economically valuable tasks. Unlike leaderboard games, GDPVal is designed to reflect AI impact on actual knowledge work — an 83% score is meaningful for anyone building AI into professional workflows.
Moonshot AI shipped Kimi K2.7-Code under a modified MIT license, claiming 30% lower reasoning token usage versus K2.6. Its predecessor K2.6 — a 1T-parameter MoE model with a 262K token context — is already the second most-used LLM on OpenRouter, putting Moonshot in direct benchmark competition with GPT-5.5 and Claude while open-sourcing aggressively.
Sam Altman, Dario Amodei, Demis Hassabis, and Mustafa Suleyman joined 65 other leaders in a joint letter to Congress backing S.3741, the Biosecurity Modernization and Innovation Act, which would require mandatory biosecurity screening of all synthetic DNA providers in the US. The signatories argue AI has lowered the expertise threshold for misusing synthetic biology to genuinely dangerous levels — the most coordinated cross-lab policy push to date.
The CEOs of OpenAI, Google DeepMind, and Anthropic are all attending the G7 summit in Evian-les-Bains, France, June 15–17 — the first G7 with all three major foundation model labs represented. The summit is expected to produce joint governance language on compute thresholds, safety standards, and possibly AI export controls.
Google's Gemini 3.5 Flash is now generally available worldwide, but Gemini 3.5 Pro — featuring a 2M-token context window and a Deep Think reasoning mode — has slipped to late June. Sundar Pichai's "wait one more month" at Google I/O reportedly drew audible groans from attendees. Every week of delay while competitors are shipping at speed is a compounding competitive cost.
Grok V9-Medium has completed training at 1.5 trillion parameters — triple the current production model — with public release expected imminently. Meanwhile, Grok 5 is training on Colossus 2, SpaceX's gigawatt-scale Memphis cluster targeting 6T parameters in a MoE architecture; xAI also launched a Grok Build Plugin Marketplace with integrations from MongoDB, Vercel, Sentry, and Cloudflare.
Moonshot AI is in talks to raise $1–2B at a $30B valuation — up from $20B just one month ago after a $2B close in May. ARR crossed $200M in April, doubling from $100M the month prior. Three funding rounds in six months from a Chinese lab most Western users haven't heard of is a signal worth paying attention to.
Google Research presented TurboQuant at ICLR 2026, combining PolarQuant vector rotation and Quantized Johnson-Lindenstrauss compression to significantly reduce KV cache memory cost during inference. KV cache is one of the primary latency and cost bottlenecks for long-context deployments — if the technique holds outside the lab, it has direct implications for making 2M-token models affordable at scale.
