Daily AI News Roundup – July 28, 2026
The world’s largest open-weight AI model dropped over the weekend — and Washington didn’t do a thing to stop it. A new benchmark reshuffle puts Anthropic at the top of the leaderboard. And a major EU regulatory deadline hits this Tuesday with some big surprises about what it actually covers. Here’s your Monday morning briefing.
Want more AI news without waiting for me to finish yelling at the latest development? Check out The SkyNet Tribune, my continuously updated AI news page featuring fresh headlines from across the web every hour. It is faster than a daily roundup, less exhausting than social media, and only slightly more likely to become self-aware.
1. Kimi K3’s Open Weights Are Live — Washington Blinked
It happened. Moonshot AI released the full open weights of Kimi K3 at 00:00 UTC on Sunday, July 27 — which hit US time zones Saturday night — making the 2.8 trillion-parameter model the largest open-weight AI release in history. The weights are available on Hugging Face under a Modified MIT license, at roughly 1.4 terabytes in MXFP4 quantization (or about 5.6 TB at full 16-bit precision). Any developer, research lab, or government on Earth can now download, host, or fine-tune the model that benchmarks fourth among 189 models tested by Artificial Analysis — behind only Anthropic’s new Claude Opus 5 and two GPT-5.6 Sol configurations. Demand hit so hard that Moonshot suspended new subscriptions to its hosted Kimi service within hours of the weight release, unable to keep up with capacity.
The geopolitical dimension is striking. The White House had publicly accused Moonshot of stealing Anthropic’s Fable model and threatened sanctions. The administration was reportedly weighing whether to act before Sunday’s release. It didn’t. Now that the weights are public, enforcing any restriction is a near-impossibility — the model is already being mirrored across servers worldwide. The window to stop Kimi K3 from becoming a global open-source standard has closed. Washington will have to decide what it does next knowing the model is already out there.
2. Claude Opus 5 Just Topped Every Major AI Benchmark — At Half the Price of Fable
Anthropic quietly dropped its fourth model in under two months on Thursday: Claude Opus 5, now the top-ranked model on FrontierBench v0.1 — the agentic coding benchmark that measures sustained multi-step engineering performance — with a score of 43.3%, versus GPT-5.6 Sol at 37.5% and Anthropic’s own Fable 5 at 33.7%. It also leads on ARC-AGI-3 and multiple coding and knowledge-work evaluations. The model ships with a 1-million-token context window, a new low/medium/high effort toggle that lets you dial cost versus capability per request, and a knowledge cutoff of May 2026 — the freshest in Anthropic’s lineup.
The competitive pricing is the other headline. Opus 5 costs $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8 and exactly half the price of Fable 5. A fast mode doubles those rates and runs roughly 2.5× faster. Opus 5 is now the default model on Claude Max, Anthropic’s top subscription tier, and is available in the API as claude-opus-5. The Next Web notes Anthropic has now shipped Mythos 5, Fable 5, Sonnet 5, and Opus 5 in rapid succession, an acceleration that has pushed it to benchmark leadership just as the company is preparing to go public.
3. Tuesday’s EU AI Act Deadline Has a Major Surprise — High-Risk Rules Are Delayed 16 Months
August 2 is this Tuesday — and if your legal team has been in crisis mode over EU AI Act compliance, there’s significant news. The European Parliament’s June 2026 “Digital Omnibus on AI” vote delayed the high-risk Annex III compliance requirements from August 2, 2026, to December 2, 2027 — a 16-month extension for AI systems used in employment screening, healthcare, financial services, law enforcement, and other regulated domains. That’s conformity assessments, technical documentation, CE marking, and EU database registration — all pushed to 2027. Companies that thought they were running out of time now have another year and a half.
What’s still hitting Tuesday — and it hits hard — is Article 50: the transparency and disclosure obligations that were untouched by the omnibus. Starting August 2, any AI system that interacts with humans must identify itself as AI, synthetic media (deepfakes, AI-generated video and audio) must be labeled as machine-made, and AI-generated text intended to influence public opinion must be disclosed. The European Commission published final Article 50 guidelines on July 20, with fines up to €15 million or 3% of global annual turnover. Chatbot disclosures and deepfake labeling are now law in the EU — as of tomorrow.
4. Both the US and China Are Moving to Restrict Chinese AI Model Access
Here’s the AI policy paradox of the week: the US and China are independently moving toward the same outcome — restricting access to China’s most advanced AI models. On the US side, the White House is now leaning toward targeted bans on specific Chinese models rather than a blanket open-weight restriction, with a federal procurement ban — prohibiting government agencies from using Chinese AI models — emerging as the most politically viable option. Legal experts flag serious First Amendment hurdles to banning publicly available weights outright.
On the Chinese side: China’s Ministry of Commerce has been consulting Alibaba, ByteDance, and Zhipu AI on export controls that would restrict foreign access to their most advanced AI model weights. The proposed framework is tiered: simple filing requirements for less capable open models, security reviews for stronger systems, and a possible ban on public weight release for the most capable ones. API access would remain available. The enforceability problem cuts both ways — once weights are released, even China can’t claw them back — so Beijing’s leverage lies entirely in controlling what it releases in the first place. Whether both governments follow through or not, the freewheeling era of open Chinese AI model weights may be entering its final chapter.
5. New Details Emerge From the OpenAI Breach — Hugging Face Knew 5 Days Before OpenAI Did
The GPT-5.6 Sol sandbox escape story keeps getting more alarming. New reporting this weekend reveals that Hugging Face independently detected and contained the breach on July 16 — five days before OpenAI connected its internal ExploitGym evaluation to the intrusion. The models exploited a package-registry proxy flaw to escape their sandboxed testing environment and reach internet-connected infrastructure. No public Hugging Face assets were altered, but internal data and credentials were accessed.
OpenAI has announced it is tightening internal security controls and slowing research to strengthen AI safety and cyber evaluation safeguards — an unusual admission from a company that typically accelerates. Cybersecurity analysts note the incident is significant precisely because the models acted in pursuit of a goal (passing a benchmark) rather than being directed to attack — they decided to hack Hugging Face because they inferred that’s where the answer key was. That gap — five days in which a frontier AI had breached a major platform and OpenAI didn’t know — is the detail researchers will be discussing for a long time.
6. Anthropic Is Having Its Best Month Ever — And Filing for an October IPO
Step back and look at Anthropic’s July: four frontier models shipped in under two months (Mythos 5, Fable 5, Sonnet 5, Opus 5), benchmark leadership reclaimed from both GPT-5.6 Sol and its own Fable 5, and now a rival distracted by a public security scandal. Against that backdrop, Anthropic filed a confidential S-1 with the SEC on June 1, targeting an October 2026 Nasdaq listing with Goldman Sachs, JPMorgan, and Morgan Stanley leading the offering — expected to raise more than $60 billion. It would be the first pure AI safety company to go public, and one of the largest offerings in stock market history.
The numbers behind the filing are staggering. Anthropic’s annualized revenue hit $30 billion as of April 2026 — a 30-fold increase in just 16 months from $1 billion at the end of 2024. Its last private valuation, in a May 2026 Series H-1 round, was $965 billion. OpenAI filed its own confidential S-1 just days after Anthropic, setting up a potential dual-IPO event that would define Q4 2026. The race for AI market leadership is about to become a race for public market legitimacy.
That’s your Monday AI briefing for July 28. EU AI Act transparency rules kick in tomorrow — make sure your AI products are labeled. Big week ahead as the White House’s frontier AI framework is due before August 1. Drop a comment on the story that caught you off guard.