Anthropic in early May 2026 unveiled a new “dreaming” capability for its Claude Managed Agents at its Code with Claude developer event. Dreaming is a scheduled background process that reviews past sessions, memories and completed tasks to surface recurring mistakes, preferred approaches and cross-agent patterns, then locks selected learnings into agent memory. The feature is launching as a research preview on the Claude Platform; developers can allow automatic updates or review changes before they are applied. Alongside dreaming, Anthropic expanded Outcomes — rubric-style evaluations that grade outputs against developer-defined standards — and Multiagent Orchestration, which lets multiple agents delegate and coordinate sub‑tasks. Two Managed Agents tools also moved from research preview to public beta. Anthropic positions the updates as part of a broader push to make agents more autonomous and productive for long-running work such as software engineering, and to extend adoption into sectors including finance and law. Access to dreaming is currently gated by an application process for developers.

Meta Platforms this month carried out a global round of cuts affecting about 8,000 employees as the company shifts resources toward artificial intelligence. Julie Bone, a content designer for Facebook based in Los Angeles, said in a LinkedIn post that she had asked to be included in the layoffs after six years at the company. Bone said her personal ambitions had diverged from Meta’s AI-first direction, that the timing suited her life, and that volunteering might spare a colleague’s role. She described having upskilled with AI tools for prototyping and automating tasks but warned that AI training alone would not guarantee job security and called for broader worker protections. The account attracted mixed social media reactions. Other former employees who spoke out amid the cuts have raised different concerns about workplace culture and team decisions, illustrating varied responses to Meta’s restructuring drive and its global workforce impact.

Apple is preparing a major redesign of the AirPods settings interface in iOS 27, iPadOS 27 and macOS 27, Bloomberg’s Mark Gurman reports, with similar accounts from other outlets. The update will reorganize the existing AirPods panel in Settings to be more functional, better organized and easier to navigate, rather than introducing a standalone AirPods app. The move responds to years of incremental feature additions — including head gesture recognition, adaptive audio, hearing-aid functionality and sleep detection — that have left the current control surface cluttered and fragmented across Bluetooth, accessibility and device-specific menus. The revamped layout is expected to surface key functionality and improve information hierarchy for connected AirPods and AirPods Pro. Apple is set to unveil the new operating systems at its Worldwide Developers Conference keynote on June 8, where iOS 27’s broader changes, including a Siri redesign, will also be shown. While the reported changes aim to streamline device management, they stop short of promising a dedicated companion app for AirPods at this stage.

Chinese AI startup DeepSeek said on May 23-24, 2026 it has made a 75% cut to prices for its flagship V4‑Pro model permanent, lowering API costs to 0.025–6 yuan per million tokens (about $0.0035–$0.83) depending on usage. The move follows the model’s launch last month, when Pro access was priced far higher because of limited high-end compute capacity. Industry observers point to growing use of Huawei’s Ascend 950 chips inside China — aided by U.S. export controls that bar Nvidia’s top chips from the market — as a likely factor, though Huawei still faces equipment-related manufacturing constraints. Third‑party benchmarks and analysts now place V4‑Pro among the world’s best for cost-efficiency, and the pricing undercuts many Western frontier models. The cut intensifies an ongoing industry price war and raises questions about model quality, data provenance and commercial risks: Anthropic has publicly accused DeepSeek of using “distillation” on its outputs, an allegation DeepSeek has not fully addressed. For enterprises, the lower token costs materially reduce running expenses for long‑context tasks such as document analysis and codebase processing, but geopolitical, compliance and IP concerns could limit adoption in regulated sectors.

OpenAI this month posted a high‑paying role on its Preparedness safety team seeking researchers to prepare for the prospect of AI systems that can design and train improved versions of themselves. The listing offers $295,000 to $445,000 in total compensation and asks for “strong technical executors to support preparations for recursive self‑improvement.” Responsibilities highlighted include defending models from data‑poisoning, building interpretability tools, running safety experiments on increasingly agentic systems and tracking automation of technical staff. The posting comes as frontier models from OpenAI, Anthropic and others have advanced rapidly — researchers at METR estimate the length of tasks that top models can complete doubles roughly every seven months — and amid public comments from industry leaders including Demis Hassabis and OpenAI CEO Sam Altman. Altman has previously laid out goals to operate large‑scale automated research interns and ultimately a fully automated AI researcher by 2028. The vacancy underscores an escalating competition for alignment talent and reflects an effort by OpenAI to anticipate long‑term, systemic risks from self‑improving AI as it pursues commercialization and an anticipated public listing.

Anthropic said Project Glasswing, a restricted cybersecurity initiative powered by its unreleased Claude Mythos Preview model, has surfaced more than 10,000 high- or critical-severity vulnerability candidates in roughly a month of scanning. Of those, 1,726 have been validated as true positives and 1,094 confirmed as high- or critical-severity. Mythos flagged 6,202 high- or critical-severity candidates across more than 1,000 open-source projects; only 97 fixes have been merged upstream and 88 advisories published so far. Notable findings include a CVE-2026-5194 flaw in WolfSSL (CVSS 9.1) that could allow certificate forgery, and a separate use case in which a Glasswing partner bank says the model helped stop a $1.5 million fraudulent wire transfer. Access to the model is limited to about 50 vetted partners including major cloud and security firms; Anthropic has also launched a Cyber Verification Program for vetted defenders. The company and partners are urging developers to shorten patch cycles as vendors such as Microsoft and Oracle adjust release cadences. Anthropic says it is keeping Mythos-class models closed to the public until stronger safeguards are in place.