One of Opus 4.8’s most notable improvements is its honesty. …Early testers report that Opus 4.8 is more likely to warn of uncertainties about its behavior and less likely to make unsubstantiated claims. … Opus 4.8 is approximately four times less likely to have defects in the code you write ignored than previous versions. – Humanity
Anthropic releases Claude Opus 4.8Frontier AI Upgrade your model to Opus 4.7 for more powerful coding, agents, and professional work performance. In benchmarks, Opus 4.8 achieves a state-of-the-art 1890 in Knowledge Work’s GDPval-AA and 69.2% in SWE-Bench Pro. Anthropic also touts its increased integrity and honesty, making illusions of success and unverified claims less likely.
Improvements to Opus 4.7 are definite but gradual, and standard pricing remains unchanged from previous versions. We also launched effort control and fast mode, which is faster and cheaper. Can be used for high-throughput workloads.
Antropic is also available Claude Code’s dynamic workflow For large-scale tasks such as codebase-wide migrations. When a user requests a complex task, Claude breaks the target into subtasks and assigns subagents to do the work. Claude dynamically runs tens to hundreds of parallel subagents in a single session and checks their behavior through internal agent critique before final output.
OpenAI launches Rosalind BiodefenseThis will give trusted developers sponsored access to GPT-Rosalind for defensive biology work such as epidemiological modeling, early detection, screening, preparedness, diagnostics, and medical countermeasure development. OpenAI is also expanding trusted access to select U.S. government and public health and biodefense partners.
Mistral introduces search toolkit in public preview. The open source framework unifies the ingestion, retrieval, and evaluation of production search pipelines used in AI applications. Mistral’s pitch is that teams should spend less time connecting search infrastructure and more time improving search quality. The toolkit can run in the cloud, on-premises, or at the edge.
Mistral launches Vibe as Mistral’s live agent product And the main AI interface, available through Mistral chat Interface and mobile app. Currently, Vibe replaces LeChat and absorbs the previous Le Chat’s history, plans, and settings within the chat mode. Vibe has Work Mode AI agent for complex multi-step tasks and Code Mode as a new coding surface for Vibe web apps. This launch positions Mistral’s consumer and developer agents at the center of day-to-day operations and knowledge work.
We believe physics deserves its own cutting-edge AI models. – Mistral
Mistral announces “Physical AI” for industrial engineering. The company said it is bringing Emmi AI to Mistral to build AI models that learn from the output of physics solvers to predict physical fields from geometry, boundary conditions, or measured data. Intended use cases include accelerating design space exploration, tools, and process optimization. They aim to apply these physical AI models as real-time digital twins for industrial partners such as ASML, Airbus, Safran, and Siemens Energy.
Microsoft announces new MAI Image 2.5 image generation modelan upgraded text-to-image generator that replaces MAI Image 2.0, follows prompting more closely and renders text strings more reliably. climb to text-to-image 3rd place on Arena.ai leaderboardMAI Image 2.5 displays powerful visual inferences about the scene and lighting, combined with crisper and more accurate text rendering, making it suitable for branding and product concepts.
Microsoft rolls out a completely overhauled design for Microsoft 365 Copilot The company calls it a “consistent agent-like experience” across its office productivity suite. The new Copilot now has a consistent entry point across apps, allowing you to draw live data directly from other integrated Microsoft apps like email, calendar, and files to generate context-aware charts and graphs.
Microsoft is looking to keep Copilot competitive as rapidly evolving AI applications take over agent functions. for that, Microsoft is reportedly developing an integrated “super app” that will integrate GitHub Copilot, Copilot Chat, and Copilot Cowork to a single destination. This new platform features agent workflow capabilities, internally called Autopilot, and is expected to be released by the end of the summer.
Perplexity announced that Perplexity Computer functionality is now available directly within Microsoft 365 applications.Word, Excel, PowerPoint, etc. Tight integration allows users to request complex multi-step analytic actions beyond standard chat responses. For example, the tool can analyze legal documents based on templates, track changes, and generate issue lists with fallback clauses.
Eleven Labs releases upgraded Music V2 generated audio modelfocuses on producing higher fidelity music tracks. Eleven Lab claims:
Music v2 delivers better vocals, instrumentation, and arrangements across all genres with improved multilingual support and a set of new features.
The underlying model is trained on fully licensed data, ensuring commercial use rights are cleared for content creators. Testing showed that the model incorporates world knowledge and can correctly reference specific landmarks and pop culture elements when given local prompts.
Eleven Labo also releases Dubbing V2an automatic video localization tool that translates audio content while preserving its original attributes. The software takes an uploaded video file and converts its audio into one of over 90 target languages, preserving the speaker’s original tone of voice, emotion, and facial expressions. This makes the output more faithful to the original delivery.
Figma has transformed its AI design assistant, Figma Make, into a live visual software editor that connects natively to your production codebase.. This update allows users to import existing Git repositories directly into the Figma desktop app, visually edit the underlying code, and push changes to engineering through GitHub pull requests. The platform leverages a multi-model AI system, switching between Anthropic’s Claude model and Google’s Gemini model to write code that adheres to established design system guidelines.
MiniMax releases technical report for M2 series and previews upcoming M3 model. The upcoming M3 series will feature ‘MiniMax Sparse Attendance’ (MSA), a secondary framework capable of 15.6x faster decoding speeds with a context length of 1 million tokens. of MiniMax-M2 Series Technical Report We focus on the sparse expert mixture architecture M2 and its training: an agent-driven data pipeline. Forge is a reinforcement learning system for agent native training. M2.7 takes a step towards self-evolution by autonomously debugging training executions.
Meta is developing an AI-powered pendant and plans to start testing it next year.. The device will be built on technology from AI startup Limitless, which Meta acquired in late 2025. Meta also plans to expand its lineup of AI glasses and launch a “Wearables for Work” business subscription.
OpenAI adds Codex computing capabilities to Windows. Apps can display screens and perform tasks on your device. Users can also manage and review Codex jobs through the ChatGPT app.
OpenAI removes Canvas functionality for GPT-5.5 models. Parallel editing functionality will no longer be available in GPT-5.5 Instant or GPT-5.5 Thinking. OpenAI also shortens GPT-5.5 instant responses and reduces the use of bullet points in text.
paper”Efficient extraction of large-scale language models preserving inferences through activation-aware initialization” argue that some efficient distillation methods damage multilevel reasoning through “inference collapse.” To fix this, the proposed RED method uses activation-aware initialization to better preserve hidden representation ranks. Experiments with Llama and Qwen models show that RED recovers inference while preserving the efficiency benefits of compressed LLM.
Anthropic raises $65 billion in Series H funding at a staggering $965 billion post-money valuationAnthropic says the proceeds will support safety and interpretability research, compute expansion, and product scaling. Anthropic’s run-rate revenue exceeded $47 billion in early May, leading OpenAI in revenue, and it has major compute deals with Amazon, Google, and SpaceX to increase its AI service capabilities. Anthropic also expanded its European footprint by opening an office in Milan.
OpenAI publishes Frontier Governance Framework This week, we discuss how OpenAI’s safety and security practices align with existing and emerging legal requirements, including in the US, California, and the EU. Frontier governance framework Learn how OpenAI addresses AI risk assessment and mitigation in areas such as cyberattacks, CBRN, harmful operations, and loss of control, and provides guidance on model reporting, security management, and incident response.
The Verge investigated the rapid normalization of AI in warfare In a special feature that argues that military AI is no longer a future scenario. In this article, we discussed migrating from Project Maven to modern AI-enabled surveillance, object detection, and targeting workflows. Tensions between government demands for broad “lawful use” and AI companies seeking to define ethical red lines around autonomous weapons and surveillance.
In the age of artificial intelligence, when human dignity is threatened by new forms of dehumanization, we have an urgent obligation to remain deeply human. – Pope Leo XIV
Pope Leo XIV issued an encyclical on AI. “Magnifica Humitas”stands for “great humanity” and focuses on “protecting humans” in the age of AI. This is a detailed, nuanced document covering the impact of AI and how to approach it. The Pope stressed that humans have a unique and inherent dignity that should not be ignored as AI capabilities grow.
Although the Pope does not completely reject AI or embrace accelerationist arguments, he does raise serious concerns and social consequences arising from the development of AI, such as the impact of AI companionship on human relationships. He criticizes how AI development is controlled by a small number of private entities, complicating the management of these technologies for the “common good.” The Pope advocates the “disarmament” of AI, which means we must move away from the idea of “arms competition” among AI species and instead foster open, human-friendly cooperation.
Pope Leo XIV and the new social issues of AI reviews Pope Leo XIV’s AI teachings in the context of Pope Leo XIII’s Revum Novarum, which faced the challenges of industrialization more than 100 years ago.
The Guardian scrutinized Anthropic’s relationship with Pope Leo XIV’s AI encyclicalshared criticism that if Anthropic were to improve its safety image without addressing AI concerns, its involvement with the Vatican could become a “Vatican wash.” Anthropic is also using AI “concerns” as a way to shut down AI development through “regulatory capture.”
The Pope advanced the discussion of AI ethics, addressing AI in religious, social, labor, and geopolitical contexts.
“I would like to use the expression disarmament, which is close to my heart. Disarmament of AI means freeing it from the spirit of armed competition… Today, it is not just limited to a military context, but is also an economic and cognitive phenomenon. This involves a competition for ever more powerful algorithms and larger data sets, driven by the desire to secure geopolitical or commercial advantage.” – Pope Leo XIV
