Meta Superintelligence Labs released Muse Spark on April 8, a natively multimodal reasoning model that accepts text, image, and voice inputs and offers both instant and thinking modes. Backed by a $14.3 billion investment commitment, Muse Spark is already integrated across Facebook, Instagram, WhatsApp, and Meta's Ray-Ban smart glasses. Third-party trackers rate it below GPT-5.4 and Claude Opus 4.6 on overall intelligence but note it is strong on medical benchmarks and multimodal reasoning.
Muse Spark's multi-agent coordination architecture is the standout feature: the model can spawn specialized sub-agents for different modalities within a single interaction session, allowing a user to, for example, take a photo, ask a question about it, and have the system generate a voice response — all handled by different agents coordinated under a single orchestration layer. The model runs free through meta.ai and the Meta AI app.
Anthropic signed a multi-year cloud infrastructure agreement with CoreWeave on April 10, securing additional data center capacity to meet surging demand for Claude's AI services. The partnership diversifies Anthropic's infrastructure providers beyond its existing deals with Google (TPU chips) and Amazon (up to 5GW of compute). The announcement coincided with revelations that Anthropic added approximately $6 billion in annual recurring revenue over the first three months of 2026 — a pace that would put the company on track for $47B ARR by year-end.
The CoreWeave deal signals Anthropic's urgency to scale: the company's API has seen demand outstrip supply in recent weeks, with some enterprise customers reporting rate limits on Claude Opus 4.5 and 4.6. By adding a third infrastructure partner, Anthropic reduces reliance on any single cloud vendor and provides more stable API availability as its user base expands beyond software developers into enterprise procurement.
Anthropic confirmed it had blocked the public release of its Mythos model after internal red-teaming revealed the model could convert browser crash logs into working computer exploits with a 72% success rate. The finding prompted Anthropic to launch Project Glasswing, a restricted-access program partnering with Amazon, Microsoft, Apple, Google, and Nvidia to test Mythos exclusively for cybersecurity defense. Only approximately 50 vetted partners currently have access.
Mythos represents a new category of 'dual-use' AI — models whose defensive capabilities (identifying thousands of OS vulnerabilities) are matched by their offensive potential (generating working exploits from minimal input). Anthropic's decision to gate rather than block the model entirely sets a precedent that other frontier labs are watching closely: how to deploy a model that is too capable to release broadly but too useful to keep entirely private.