OpenAI just launched GPT-5.4 mini and nano, lightweight models built for high-volume workloads and tool-heavy coding. The mini variant runs over 2x faster than GPT-5 mini while nearly matching the flagship GPT-5.4 on SWE-Bench Pro, at just $0.75 per million input tokens. Nano drops the entry cost to $0.20, perfect for background subagents.
Mistral just unveiled Mistral Small 4, consolidating reasoning, multimodal vision, and agentic coding into a unified 119B MoE model with only 6B active parameters. It offers a 256k context window, a toggleable reasoning dial, and cuts completion latency by 40% compared to version 3. Best of all, it's released under Apache 2.0 and beats competing models with far fewer output tokens.
Emil Kowalski just open-sourced a collection of design and motion skills for AI agents. From fixing awkward easing curves to Apple-grade fluid transitions and mobile polish, the rules inject real design engineering taste into AI code generation.
Moonshot AI just introduced Attention Residuals (AttnRes), a drop-in replacement that swaps standard fixed residual connections for dynamic attention across depth. This lets layers selectively pull from earlier representations without suffering from representation dilution. Its practical block variant matches a baseline trained with 1.25x more compute, delivering sharp jumps in complex reasoning and coding benchmarks.
Meta has reportedly delayed its next-generation AI model, codenamed "Avocado," pushing its mid-March release to May at the earliest. Internal benchmarks showed the model lagging behind rivals from Google, OpenAI, and Anthropic in reasoning and coding, leading leadership to even weigh licensing Gemini.
Elon Musk just revealed that xAI "was not built right" and is undergoing a ground-up rebuild after 9 of its 11 original co-founders departed, including Grok Code lead Guodong Zhang. Musk admitted Grok is lagging in coding, prompting raids on Cursor talent as the startup reshuffles ahead of a massive planned IPO.
Anthropic just made the 1M-token context window generally available for Claude Opus 4.6 and Sonnet 4.6, ditching long-context price premiums entirely. Single-request media limits also jumped sixfold to 600 images or PDF pages. Claude Code users on paid tiers now get the full window by default, putting an end to messy mid-task context compaction.
Manus just rolled out a desktop app for macOS and Windows, introducing a feature called My Computer. Instead of being trapped in a cloud sandbox, the AI agent can now execute terminal commands to organize local files, build desktop apps, and tap into idle GPUs. Every command requires explicit user approval before running.
Researchers just introduced Terminator, a lightweight early-exit probe that detects when a reasoning model hits the final answer and stops generation on the spot. Across benchmarks like MATH-500 and GPQA, it cuts chain-of-thought length by up to 55% and doubles inference speed with virtually no accuracy loss. Code is expected soon.
xAI just hired Andrew Milich and Jason Ginsberg, two senior product engineering leads from hot coding startup Cursor, reporting directly to Elon Musk. Musk admitted that Grok is currently lagging behind rivals in coding and said the startup is undergoing an all-hands rebuild to leapfrog the competition. Why it matters: Cursor built the gold standard for developer workflows, and xAI wants that applied product magic inside its models.
alphaXiv just released a SKILL.md configuration for Claude Code, letting the coding agent fetch structured paper overviews directly via API. Instead of struggling with messy raw PDF parsing, agents can now pull clean summaries from arXiv and alphaXiv URLs in seconds. It marks one of the first killer third-party skills in the Claude Code ecosystem.
WordPress just launched my.WordPress.net, running a full, persistent site entirely inside your browser with zero sign-up or hosting needed. Built on WordPress Playground, it stores data locally and features one-click apps including an AI workspace that can safely modify plugins and manage private knowledge. Basically, it turns WordPress into a private sandbox where AI can safely tinker with code.
Baidu Cloud just launched DuClaw, a turnkey OpenClaw service that runs straight in the browser. Users can run full autonomous agent workflows without provisioning servers or API keys, complete with built-in Baidu search skills and instant switching between models like DeepSeek and Kimi-K2.5.
OpenAI just upgraded its Responses API with a native shell tool, hosted container workspaces, and built-in context compaction for GPT-5.2 and later models. Agents can now run Unix commands like curl or grep concurrently, stage files, and query SQLite databases in sandboxes. It basically saves developers from having to build their own execution harness for long-running workflows.
Alibaba Cloud's Wuying team just launched JVS Claw, an agent platform built directly on top of OpenClaw. Instead of forcing users to deploy scripts on local machines, it runs web browsing and code execution inside an isolated cloud sandbox called ClawSpace with cross-device sync. It turns a complex open-source agent setup into a one-click hosted service.
Claude Code just rolled out a new /btw command that lets you hold parallel side chats while background tasks keep running. Instead of waiting for a long automated process to finish or interrupting it, developers can now ask quick questions, clarify specs, or bounce ideas off the model on the side without losing momentum.
Google just launched Gemini Embedding 2 in public preview, its first natively multimodal embedding model. It maps text, images, audio, up to 120 seconds of video, and PDFs into a single vector space, even handling mixed inputs in one request. Matryoshka Representation Learning lets developers scale dimensions down from 3072 to cut storage costs.
Cloudflare just launched an open beta for its /crawl endpoint in Browser Rendering, letting developers scrape an entire website into Markdown or structured JSON with a single API call. The tool automatically obeys robots.txt and is built specifically to make gathering clean data for RAG pipelines and model training headache-free.
Andrej Karpathy just drafted AgentHub, a stripped-down GitHub alternative built specifically for swarms of AI agents. It strips away human-centric branches and PRs in favor of a sprawling commit DAG and an asynchronous message board. Why it matters: Karpathy took the repo private within days, leaving it behind as a conceptual sketch for autonomous agent research communities.
Andrej Karpathy just open-sourced autoresearch, a 630-line script that puts AI agents in charge of the scientific method while you sleep. The agent writes hypotheses, edits training code, runs quick GPU tests, and keeps winning changes—already boosting GPT-2 training efficiency by 11% in two days.
