Here are today's picks from AI news. The big stories: Anthropic's late-June announcements of Claude Sonnet 5 and the redeployment of Fable 5 and Mythos 5, which had been taken offline amid export controls. The latter in particular matters, as it wades into an industry-wide challenge: how to share safety evaluations of AI models. We've also rounded up developments around Google Earth's AI satellite imagery and a couple of articles from Japan's tech community.
Claude Sonnet 5 Arrives — Agentic Performance Moving Closer to Opus 4.8
On June 30, Anthropic announced a new model, Claude Sonnet 5 (model ID `claude-sonnet-5`). Gains in agentic capability had been led by the Opus class, but Sonnet 5 is billed as “the most agentic Sonnet yet,” substantially raising its ability to make plans, use tools such as browsers and terminals, and push tasks forward autonomously.
On performance, it comfortably clears the previous-generation Sonnet 4.6 and is said to approach the more expensive Opus 4.8. Cost-effectiveness got attention too: as an introductory price, it is set at $2 per million input tokens and $10 per million output tokens through August 31, moving to $3/$15 thereafter. It becomes the default model on the Free and Pro plans and is also available on Max, Team, and Enterprise, as well as in Claude Code and on the Claude Platform.
In safety evaluations, undesirable behaviors are down overall compared with Sonnet 4.6, and the model is considered safer in agentic use. Its capability on cybersecurity tasks, however, is deliberately kept below that of the current Opus model, and safeguards that detect and block dangerous cyber uses are enabled by default.
Sonnet 5 is also already being referenced as a benchmark for comparing models. The recent buzz claiming DeepSeek's new model is “on par with Sonnet 5” is one example — a sign that it is becoming a reference point for frontier model comparisons.
Fable 5 and Mythos 5 Redeployed, and a Proposed Framework for Grading Jailbreaks
The other Anthropic story is the redeployment of Claude Fable 5 and Claude Mythos 5, which had been forced offline in June. Announced on June 30 as well, Fable 5 became available again worldwide on July 1.
Tracing how we got here: on June 12, the US government applied export controls to both models. The direct trigger was a report that Amazon researchers had found a way around Fable 5's safeguards. The model identified software vulnerabilities and, in some cases, produced code demonstrating how to exploit them — which was what raised alarms. Because the controls took effect immediately and Anthropic had no reliable way to verify users' nationalities in real time, the company suspended access for all users.
After consultations with the government, the export controls were lifted on June 30. Anthropic says it trained a new safety classifier that surgically blocks the reported bypass, catching it more than 99% of the time. Notably, it also points out that other models it tested (including Opus 4.8, GPT-5.5, and Kimi K2.7) could identify the same vulnerabilities, emphasizing that no dangerous capability unique to Fable 5 was exposed.
How Should Jailbreak Severity Be Measured?
The most notable part of this story: working with Amazon, Microsoft, Google, and others, drafting has begun on a shared framework for evaluating the severity of AI jailbreaks (bypasses of safety mechanisms). There is no industry-wide standard today, and the motivating concern is that every new bypass technique makes severity judgments murkier.
The proposal rates a given jailbreak on four dimensions: (1) capability gain (how far beyond existing tools it goes), (2) breadth of the gain (how many kinds of attacks the same technique covers), (3) ease of weaponization (how much work it takes to turn into an actual attack), and (4) discoverability (how widely known the technique is). Paired with stronger coordination with governments and industry and a newly launched HackerOne program, the stated goal is to build industry-wide standards.
Following earlier reports that “jailbreaking frontier AI is frighteningly easy,” this move shifts things toward measuring severity in a shared vocabulary and systematizing the response. With model capabilities still climbing, the focus will be on how much traction common rules for safety evaluation can realistically gain.
Conversing with History Through AI — The Theodore Roosevelt Presidential Library
Here's an example of AI applied for public benefit. Microsoft announced that, as technology partner to the Theodore Roosevelt Presidential Library in Medora, North Dakota, which opened to the public on July 4, it helped deliver an AI-powered “living library.”
At the core is a knowledge base called “Box 1.” AI was used to organize and enrich hundreds of thousands of historical documents, reconstructing fragmented materials into a searchable form. On top of that foundation, the library built the Campfire Reading Room, a digital research tool that returns Roosevelt's own documents when anyone in the world asks a question in everyday language, plus an experience that lets visitors converse with an AI avatar of Roosevelt.
Led by Microsoft's AI For Good Lab, the project plans to open-source the software and publish a paper explaining how it all works, so that other cultural institutions can apply the technology. This “living library” approach — the avatar automatically updates as more materials are added — suggests one direction for AI in museums and libraries.
From Japan's Tech Community — Generate-and-Verify Loops and an OSS Agent Primer
Two picks from Japanese-language technical articles.
The first, “Generation Can Be Sloppy — the Verifier Catches Everything,” examines a generate-then-verify loop that still works with an acceptance rate of 0.26%. Taking verigen, a Go CLI that generates puzzles such as cryptarithmetic, as its subject, it articulates a separation of duties: the generator doesn't have to guarantee correctness — the verifier bears full responsibility for judging what is “correct.” If generation is cheap, mass-produce roughly and narrow things down through verification: that perspective carries over to practical work with LLMs on code, SQL, and structured data.
The second is “Learning AI Agents with Hermes,” a write-up of study group materials from a Waseda AI research club. Using the open-source AI agent Hermes Agent as its example, it explains, for beginners, what separates agents from chat-based AI like ChatGPT and the mechanics of how they run.
In Brief — Google Earth's AI Satellite Imagery, Chinese Researchers on X, and a Vulnerability Found by AI
Google Earth's AI satellite imagery:** Google reportedly incorporated AI-generated satellite imagery into Google Earth, only to pause it as concerns mounted over “deepfakes of the sky.” The episode threw fresh light on the problem of generated imagery seeping into maps — a medium that is supposed to convey facts.
Chinese AI researchers speaking up on X:** While staff at OpenAI and Anthropic have gone quieter online, researchers at Chinese AI labs are reportedly posting their own work on X, courting talent and shaping the discussion. It hints that the center of the AI conversation is gradually shifting.
AI finds an Oracle VirtualBox vulnerability:** A vulnerability in Oracle VM VirtualBox, CVE-2026-60161, was reportedly discovered by AI. It is drawing attention as one example of AI being put to use in defensive security research.