Google's Gemini API Agent Skill boosts coding task success from 28% to 97%

NVIDIA released gpt-oss-puzzle-88B on March 26, 2026, a 88-billion parameter mixture-of-experts model optimized for inference efficiency on H100 hardware. Built using the Puzzle post-training neural architecture search framework, the model achieves 1.63× throughput improvement in long-context (64K/64K) scenarios and up to 2.82× improvement on single H100 GPUs compared to its parent gpt-oss-120B, while matching or exceeding accuracy across reasoning effort levels.

Researchers at Meta, University of British Columbia, and partner institutions have developed hyperagents—AI systems that optimize both their task performance and the mechanisms controlling their self-improvement. Unlike previous self-improvement approaches locked to coding tasks, DGM-Hyperagents (DGM-H) demonstrate significant gains across four domains and can transfer improvement strategies to entirely new tasks.

OpenAI is shutting down Sora in two phases: the web app and mobile application close April 26, 2026, followed by the API on September 24, 2026. Users must export their videos and images before the cutoff dates, as user data will be permanently deleted afterward. The discontinuation reflects OpenAI's strategic shift toward coding tools and enterprise products.

Kwaipilot released KAT-Coder-Pro V2, the latest model in its KAT-Coder series, on March 27, 2026. The model features a 256,000-token context window and is priced at $0.30 per million input tokens and $1.20 per million output tokens. It targets enterprise-grade software engineering with focus on multi-system coordination and web aesthetics generation.

Cohere has released Transcribe, a 2 billion parameter open-source automatic speech recognition model that the company claims tops the Hugging Face Open ASR Leaderboard with a 5.42% word error rate. The model supports 14 languages and is available under Apache 2.0 license, outperforming OpenAI's Whisper Large v3 and competing models on both accuracy and throughput metrics.

Meta's FAIR lab released TRIBE v2, an AI model that predicts human brain activity from images, audio, and text. Trained on over 1,000 hours of fMRI data from 720 subjects, the model maps predictions to 70,000 voxels and often matches group-average brain responses more accurately than individual brain scans.

Replit has released Agent 4, a major update that fundamentally restructures how product teams design, collaborate, and build. The release replaces Design Mode with an infinite Design Canvas, moves from fork-and-merge to shared real-time collaboration, and enables concurrent planning and execution.

Google has launched a memory import feature for Gemini that lets users transfer their chat history, memories, and preferences from competing AI services like ChatGPT and Claude. The feature works by exporting conversation data and personal context from your existing AI service, then importing it into Gemini to enable faster personalization.

Suno has released version 5.5 of its AI music generator with three new features: Voices (record/upload your singing voice for AI songs), Custom Models (fine-tune the model on your music library), and My Taste (personalized genre/mood recommendations). The update is available to Pro and Premier subscribers.

Cybersecurity stocks fell sharply Friday after reports that Anthropic is testing Mythos, described as its most powerful model yet, with enhanced capabilities that present potential security risks. CrowdStrike and Palo Alto Networks dropped 7%, while Tenable fell nearly 11%. Anthropic plans a cautious rollout due to cybersecurity implications.

OpenAI is introducing plugins for Codex that bundle skills, integrations, and connectors into shareable workflow packages. The move directly addresses Claude Code's lead among developers and positions Codex beyond coding into broader agentic work platforms. Over 20 plugins are currently available, including integrations with Figma, Notion, Gmail, Slack, and Google Drive.

A data breach at Anthropic exposed internal documents detailing an unreleased AI model the company describes as its most powerful to date. Anthropic confirmed it is already testing the model with select customers, claiming significant advances in reasoning, coding, and cybersecurity. The breach resulted from a misconfiguration in Anthropic's content management system that automatically made ~3,000 uploaded files publicly accessible.