DeepSeek just launched and open-sourced DeepSeek-V4 Preview, making a 1M token context standard across its services. It delivers two MoE models—the flagship V4-Pro and the lightweight V4-Flash—built around hybrid compressed attention that slashes KV cache memory to just 2% of standard architectures. It basically tackles the persistent context blowout in long-running coding agents, and the new models are already live in chat and via API.
OpenAI just launched GPT-5.5 and GPT-5.5 Pro, heavily optimized for autonomous coding, knowledge work, and complex computer use. The model matches GPT-5.4's per-token latency while requiring fewer tokens and setting a new high score of 82.7% on Terminal-Bench 2.0. Why it matters: AI assistance is shifting from step-by-step chatting to letting agents independently plan, test, and finish messy engineering projects.
Anthropic just published a postmortem tracing recent Claude Code performance drops to three distinct issues: a downgraded default reasoning effort, a caching bug that wiped prior thinking, and a verbosity-limiting prompt that hurt coding quality. All three issues have been fixed as of April 20, and Anthropic is resetting usage limits for all subscribers.
Anthropic just added everyday lifestyle connectors to Claude, including Spotify, Uber, Instacart, and Booking.com. The directory now tops 200 tools, and Claude can dynamically suggest apps mid-conversation or chain them together, always confirming before booking or buying. It brings Claude one step closer to a true universal agent.
Microsoft just made agentic capabilities in Word, Excel, and PowerPoint generally available, turning Copilot from a passive suggester into an active co-worker. Instead of just listing steps, it can now directly build visuals, restructure documents, and execute multi-step native edits right on your canvas. The update is now rolling out to enterprise and personal Microsoft 365 subscribers.
xAI just rolled out Grok Voice 1.0 via API, designed for complex multi-step enterprise workflows like customer support and phone sales. It reasons in the background with zero added latency, handles noisy calls and strong accents across 25+ languages, and is already handling Starlink's support lines.
Microsoft just launched hosted agents in Foundry Agent Service in public preview. Instead of sharing containers, every agent session gets a dedicated hypervisor-isolated sandbox with a persistent filesystem that survives scale-to-zero idle events. It supports any orchestration framework like LangGraph or Claude SDK and connects directly to M365 Copilot.
Alibaba just open-sourced Qwen3.6-27B under the Apache 2.0 license, packing flagship-grade agentic coding into a 27B dense model. According to the team, it outperforms their much larger Qwen3.5-397B-A17B across major coding benchmarks while supporting both thinking and standard inference modes. It proves small dense models are catching up fast to monster MoE architectures in specialized developer workflows.
Google just unveiled its eighth-generation TPUs at Cloud Next, splitting the lineup into TPU 8t for training and TPU 8i for inference. Co-designed with Google DeepMind, the 8t delivers 121 ExaFlops per 9,600-chip superpod, while the 8i cuts reasoning latency with 80% better performance-per-dollar. It is Google's biggest hardware push yet to custom-build data centers around multi-agent workflows.
OpenAI just launched workspace agents in ChatGPT, leveling up traditional GPTs into collaborative, cloud-running automations for enterprise teams. Powered by Codex, these agents can handle multi-step tasks across apps like Slack, follow custom approval workflows, and keep working in the background. It is available in research preview and free until May 6, 2026.
Anthropic just published architectural patterns for connecting production agents to enterprise systems via Model Context Protocol (MCP), revealing monthly SDK downloads crossed 300 million. It highlights patterns like grouping tools by intent and dynamic tool search, which slashes tool-definition tokens by over 85%. Basically, MCP is solidifying its role as the standard plug for cloud-hosted agents.
OpenAI just rolled out WebSocket mode for its Responses API, tackling the cumulative network and validation overhead that drags down fast agent loops. By keeping a persistent connection and caching previous response states in memory, end-to-end latency drops by up to 40% with a 45% faster time to first token. Coding tools like Cursor and Vercel are already clocking major speedups.
Hugging Face just released ml-intern, an open-source agent that reads papers, pulls datasets, and fine-tunes LLMs without human hand-holding. In under 10 hours on a single GPU, it pushed a compact Qwen model's GPQA score from 10% baseline to over 30%. It basically automates the tedious post-training loop using native Hub tools.
Meta just started installing tracking software on U.S. employees' computers to record mouse movements, keystrokes, and occasional screenshots to train autonomous work agents. The company claims the data will not be used for performance reviews, but legal experts warn the practice brings gig-worker-level surveillance to white-collar desks.
Google just open-sourced the draft specification for DESIGN.md, the format behind its Stitch AI UI design tool. It lets users export and share design systems across any platform, giving AI agents explicit context on color roles and WCAG accessibility standards. Basically, AI can now build on-brand interfaces without reinventing the rules every time.
Google just upgraded its autonomous research agents with Gemini 3.1 Pro, introducing Deep Research for low-latency tasks and Deep Research Max for heavy-duty background diligence. Both now connect to proprietary datasets like FactSet via MCP and generate native visual charts. It basically turns Gemini into an enterprise-ready research analyst.
Kimi just launched Kimi K2.6, an open-source coding model built for long-horizon engineering tasks. It supports over 4,000 sequential tool calls, runs continuously for 12 hours, and orchestrates swarms of up to 300 parallel sub-agents to refactor 100+ files at once. Model weights are live on Hugging Face alongside web access on kimi.com.
Researchers just dropped OpenGame, an open-source agentic framework that can generate, test, and debug fully playable web games from simple text prompts. Instead of collapsing on complex game loops and multi-file logic, it evolves reusable project skeletons and verified debugging routines. Code and playable demos are out.
TwelveLabs just unveiled Pegasus 1.5 and Rodeo at NAB Show 2026, alongside an integration with Autodesk Flow Capture. Pegasus 1.5 extracts time-based structured metadata in a single API call, outperforming Gemini 2.5 Pro by 30% on video segmentation benchmarks. Meanwhile, Rodeo acts as an AI co-pilot that lets editors find, cut, and assemble footage using plain language.
Alibaba just released an early preview of Qwen3.6-Max-Preview, offering a first look at its next flagship model. Compared to Qwen3.6-Plus, it brings major leaps in agentic coding, instruction following, and real-world agent reliability. It is Alibaba's latest push to make autonomous agents dependable for complex tasks.
