What Happened

Three significant AI developments emerged during the week of April 6–12 , 2026, per a roundup published by Juejin AI . Anthropic quietly released its most capable model to date under restricted access, a violent incident targeting OpenAI's CEO escal ated AI-related social tensions into the physical world, and Alibaba's video generation model topped two benchmark leaderboards while operating anonym ously before the company claimed ownership.

Claude Mythos: Anthrop ic's Most Capable Model, Restricted to 12 Partners

On April 7, Anthropic released Claude Mythos Preview — internally codenamed "Capybara" — positioning it above the existing Opus tier. Unlike all prior Anthropic model launches, Mythos is not publicly available . Access is limited to 12 core institutional partners and approximately 40 critical infrastructure organizations via invitation only, according to the Juejin report.

The rationale is explicit. Anthropic's own risk assessment found that Mythos can autonomously discover and exploit zero-day vulnerabilities in mainstream operating systems and browsers, producing complete proof-of-concept exploit programs. On a cybersecurity vulnerability reproduction benchmark, Mythos scored 83.1%, compared to 66.6% for Opus 4.6 — a 16.5 percentage point gap. Anthropic determined that public release could accelerate large-scale cyberattacks at scale in 2026.

Pricing reflects the capability tier. The Mythos API is priced at $25 per million input tokens and $125 per million output tokens — approximately 67% higher than Opus 4.6, according to the source article. The pricing signals a directional shift: at sufficient capability levels, per -token costs may increase rather than continue their historical downward trend.

Why It Matters

The restricted release model represents a structural departure from Anthropic's prior public API approach. By concentrating access to its most capable model among a small set of enterprise and infrastructure partners, Anthropic is making an explicit governance decision: who gets to use frontier AI, and under what conditions. The April 4 shutdown of OpenClaw subscription channels, followed three days later by the Mythos announcement, suggests a deliberate tightening of the open-access model in parallel with capability increases.

  • Cybersecurity benchmark score: 83.1% (Mythos) vs. 66.6% ( Opus 4.6)
  • API pricing: $25/M input tokens, $125/M output tokens
  • Access : 12 core partners, ~40 critical infrastructure organizations
  • No public API availability announced

Altman Residence Targeted in Fi rebomb Attack

In the early hours of April 10, San Francisco police responded to a report of a firebomb thrown at OpenAI CEO Sam Altman's residence in the North Beach neighborhood, igniting the exterior gate. The suspect, identified as 20-year-old Daniel Moreno- Gama, fled the scene and approximately one hour later appeared outside OpenAI's headquarters and threatened to set the building on fire. Police arrested Moreno-Gama; no injuries were reported, per the Juejin summary.

According to the report, Moreno-Gama is believed to have been affiliated with the "Pause AI" movement and had previously posted online content related to AI-driven human extinction. The attack occurred days after The New Yorker published a lengthy critical profile of Altman by Ronan Farrow. Altman subsequently published a blog post stating he had "underestimated the power of words and narratives" and called on the industry to reduce "the intensity of rhetoric and tactics."

Alibaba's Happy Horse Tops Video Benchmarks Anonymously

On the night of April 7, AI evaluation platform Artificial Analysis announced on X the addition of a "pseudonymous" video model — designated HappyHorse-1.0 — to its Video Arena leaderboard. No organization claimed the model at launch. Within 48 hours, HappyHorse-1.0 reached an Elo score of 1,389 in the text -to-video category and 1,403 in the image-to-video category, topping both leaderboards and displacing ByteDance's Seedance 2 .0 and Kuaishou's Kling 3.0, according to the Juejin report.

On April 10, Alibaba officially claimed the model. HappyHorse is developed by the innovation division under Alibaba's ATH business group and is currently in closed beta. Alibaba stated the API will open on April 30. Lin Junyang, former head of Alibaba's Tongyi Qianwen team, publicly endorsed the release on social media, writing : "happy horse is insanely happy."

Technical Detail

HappyHorse-1.0 specifications per the official announcement:

  • Parameter count: 15 billion
  • Supported modes : text-to-video, image-to-video, video-to-video
  • Audio: native synchronized generation (no separate pipeline )
  • Output resolution: 1080p
  • Watermark: none
  • Commercial use: permitted
  • API availability: April 30, 2026

The anonymous benchmark strategy allowed HappyHorse to accumulate Elo scores without the pre-release expectations that typically accompany major platform announcements. The timing also coincided with Sora's shutdown, leaving a competitive vacancy in the text-to-video segment .

What To Watch

  • HappyHorse API launch (April 30): Pricing, rate limits, and enterprise terms will determine whether the benchmark performance translates to developer adoption. Watch for compar isons against Kling 3.0 and Seedance 2.0 on production workloads.
  • Claude Mythos partner disclosures: Anthropic has not named the 12 core partners. Any partner disclosures in the next 30 days will clarify which enterprise segments gain early access to the 83.1%-benchmark cybersecurity capability tier.
  • Altman security and OpenAI public posture: The arrest of Moreno-Gama does not resolve underlying social pressure dynamics. Watch for any changes to OpenAI's executive security disclosures or public communications strategy in response to the incident.
  • Restricted-model preced ents: If Mythos performance data leaks through partner integ rations, competitive pressure on Google DeepMind and Meta to match both the benchmark scores and the selective-access model will intensify.