OpenAI ships new voice API models with GPT-5-class reasoning

OpenAI added three new voice models to its Realtime API on May 14: GPT-Realtime-2, a conversational model built on GPT-5-class reasoning; GPT-Realtime-Translate, covering 70 input languages and 13 output languages; and GPT-Realtime-Whisper for live streaming transcription. The release moves OpenAI's voice platform from call-and-response demos toward production-grade agents that can listen, reason, translate, and take action mid-conversation, with clear targets in customer service, education, and creator platforms.

Anthropic launches Claude Platform on AWS with full API parity

Anthropic made its Claude Platform generally available on AWS, giving AWS customers access to the full Claude API feature set — including Claude Managed Agents, Agent Skills, code execution, the Files API, web search, and prompt caching — with AWS billing and IAM authentication. Every new feature and beta now ships to Claude Platform on AWS the same day as the first-party API, a meaningful commitment for enterprise teams already standardized on AWS infrastructure.

Google DeepMind brings Gemini-powered AI pointer to Chrome

DeepMind published its AI Pointer research on May 12 and announced it is already integrating the system into Chrome: users can point at any part of a webpage and ask Gemini questions without copy-pasting into a separate chat window. A Magic Pointer version for the Googlebook laptop is coming soon. The framing is explicit — Google wants AI to meet users inside their existing tools rather than forcing them into a dedicated AI interface.

Meta completes Muse Spark rollout to all apps and AI glasses

Meta finished rolling out Muse Spark, its first model from Meta Superintelligence Labs, to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses, completing the wave that started with the Meta AI app on April 8. The update also added faster voice responses and new shopping-mode features that surface Facebook Marketplace listings alongside web results. Muse Spark marks Meta's break from its open-source Llama approach, with paid API access for third-party developers planned for later this year.

SubQ Launches First Commercial Subquadratic LLM With 12M-Token Context

Subquadratic (SubQ) shipped SubQ 1M-Preview on May 5, using sparse subquadratic attention end-to-end rather than a standard transformer, which scales linearly in context cost instead of quadratically. The model ships with a native 12 million token context window and the company claims roughly one-fifth the cost of frontier models on long-context tasks, with up to 52x faster attention at scale. Prior "1M context" transformer models carried quiet caveats about quality degradation at length; this is the first commercial release claiming to remove that ceiling architecturally.

Claude Mythos Preview Sets New GPQA Diamond Record at 94.6 Percent

Claude Mythos Preview now leads the LLM Stats leaderboard on GPQA Diamond with a score of 94.6%, making it the highest recorded result on the benchmark currently tracking frontier reasoning. The model also cleared a 32-step end-to-end cyber-attack range in evaluation, a threshold that only one other frontier model (GPT-5.5) has since matched. Anthropic has described the model internally as representing "a step change" in performance relative to Claude Opus 4.6, with substantially higher scores on coding, academic reasoning, and cybersecurity tasks.

USC's Attractor Models cut Transformer perplexity 46% with fixed-point refinement

Researchers at USC proposed Attractor Models, an architecture where a backbone proposes output embeddings that a second module refines by solving for a fixed point via implicit differentiation, keeping training memory constant regardless of iteration depth. In language modeling, the approach improves perplexity by up to 46.6% and downstream accuracy by up to 19.7% over standard Transformers at matched size, with a 770M Attractor Model outperforming a 1.3B Transformer trained on twice as many tokens. On hard structured-reasoning tasks, a 27M-parameter version hits 91.4% on Sudoku-Extreme and 93.1% on Maze-Hard, where frontier models like GPT o3 fail completely.

NVIDIA Ising open models deliver 2.5x faster quantum error correction

NVIDIA released the Ising family of open AI models, the first purpose-built for quantum computing, targeting the two hardest bottlenecks: processor calibration and quantum error correction decoding. The models deliver up to 2.5x faster decoding performance and 3x higher accuracy compared to prior methods, and integrate with NVIDIA's CUDA-Q platform and the NVQLink QPU-GPU interconnect for real-time control. Ising ships alongside a cookbook of quantum workflows and NIM microservices, letting researchers fine-tune for specific qubit hardware with minimal setup.

EU AI Act Omnibus deal extends deadlines, bans nudification apps

On May 7, 2026, the European Parliament and Council reached a provisional political agreement amending the EU AI Act, extending compliance deadlines for high-risk AI systems and adding an outright ban on AI-generated non-consensual intimate imagery and CSAM, effective December 2, 2026. The deal also delays watermarking obligations for AI-generated content to December 2, 2026, narrows which industrial AI products fall under high-risk rules, and clarifies that the EU AI Office holds supervisory authority over general-purpose AI models except in law enforcement, border management, and financial institution contexts. The agreement still requires formal adoption by both institutions before becoming law.

TAKE IT DOWN Act platform deadline arrives; FTC signals active enforcement

May 19, 2026 marks the compliance deadline for covered online platforms under the TAKE IT DOWN Act, the first federal law targeting AI-generated deepfake intimate imagery, requiring a notice-and-takedown process that removes flagged content within 48 hours. In the run-up to the deadline, FTC Chairman Andrew Ferguson sent formal warning letters to more than a dozen major platforms, including Meta, Apple, Microsoft, TikTok, Reddit, Snapchat, and X, signaling the agency's intent to pursue civil penalties of up to $53,088 per violation. The law, signed in May 2025 with near-unanimous congressional support, is the only standalone federal AI statute enacted to date.