AI News This Week: Enterprise AI News | AI Weekly Pulse #2

Anthropic quietly made Claude Sonnet 5 the new default for Free and Pro users on June 30, 2026 - and the numbers behind that decision are hard to argue with. It hits 63.2% on SWE-bench Pro, edges out Claude Opus 4.8 on Terminal-Bench 2.1 (80.4% vs 74.6%), and comes in at $2 per million input tokens and $10 per million output tokens through August 31, 2026. It also ships with a 1M token context window, stronger tool use, and noticeably better performance on agentic and coding tasks.

After the US Department of Commerce lifted export controls, Anthropic brought Claude Fable 5 back to global users on July 1, 2026. The model had been off the table since June 12 following cybersecurity concerns - a three-week window that left international teams scrambling. It's back now, but not unchanged: the redeployment comes with tighter safety classifiers and stricter behavioral constraints baked in. Anthropic also used the moment to propose something broader - an industry-wide jailbreak severity scoring framework, developed alongside Amazon, Microsoft, and Google. Claude Mythos 5 is a different story; it stays locked to approved US organizations for now.

Etched just hit a $5 billion valuation, and the math behind it is straightforward: $1 billion in sales orders for its Sohu Transformer chip. The company's whole bet is that building silicon exclusively for transformer model inference - rather than repurposing general-purpose GPUs - will win on cost-per-token and energy efficiency when it counts most: in production.

Anthropic is in early talks with Samsung Electronics to build a custom AI accelerator chip on Samsung's 2-nanometer SF2P process. It's still exploratory, but the direction is clear. The initiative follows a string of hires from OpenAI's custom silicon program, and the goal is straightforward: reduce dependence on existing GPU suppliers by bringing more of the compute stack in-house. The existing partnerships with Google, Amazon, and Nvidia aren't going anywhere - this would sit alongside them, not replace them.

Meta is building something it's calling "Meta Compute" - an internal initiative to turn its sprawling AI infrastructure into a commercial product. The plan is to offer enterprises two things: API access to hosted AI models, much like Amazon Bedrock, and raw GPU compute capacity along the lines of what CoreWeave sells today. The underlying motivation isn't hard to read. Meta has poured $600 billion into data centers, and those clusters sit idle between training runs. Selling that capacity is a way to claw some of that back - and it puts Meta in direct competition with AWS, Google Cloud, and Azure.

New York didn't ease into this one. The Synthetic Performer Law took effect on June 29, 2026, and went into active enforcement the same day. Any advertiser running an ad that features a synthetic human performer, a deepfaked voice, or an AI-generated likeness of a real person now has to disclose it - clearly and conspicuously. No grace period, no phased rollout. Agencies and brands that skip the disclosure face financial penalties and civil litigation exposure. And because New York is New York, the industry expectation is that this effectively becomes a national standard whether other states follow or not.