Researchers just used GPT-6 Astra to prove a Goldbach-type conjecture concerning the Liouville function, showing every even integer greater than 2 is the sum of two integers with an odd number of prime factors. The mathematical proof has been completely formalized and verified in Lean 4 without unverified axioms, marking another leap for AI in pure mathematics.
Google just open-sourced AX (Agent Executor), a declarative runtime built to schedule and isolate millions of autonomous AI agents on Kubernetes clusters. It treats agents as stateful workloads with K8s-style manifests, offering commands to pause tasks or shell directly into sandboxes. It is currently in active development under the Apache 2.0 license.
Chinese AI lab StepFun just launched Step 5 Preview, its flagship multimodal model designed specifically for agentic workflows. It packs a 600B total parameter MoE architecture that only activates 27B per token, alongside a 1M token context window and strong DeepSWE coding benchmarks. Developers can already test it via the StepFun Open Platform.
Google just started testing what looks like Gemini 4 Pro under a "gemini-3.8-flash" alias on the LMSYS Arena. The model generated an intricate mechanical butterfly SVG animation in just ten minutes and reportedly swept state-of-the-art scores in agent coding and OS benchmarks. Google might be back in the lead.
Anthropic just added support for OpenAI's AGENTS.md standard in Claude Code version 2.1.277. If a project lacks a CLAUDE.md file, Claude Code will now automatically fall back to AGENTS.md, with behavior configurable via /config. Why it matters: Developers juggling multiple coding assistants no longer need messy symlinks or duplicate instruction files across repos.
Researchers just released JEPA-Anything, an open-source framework designed to model predictive dynamics across biology, physics, and medicine. Instead of monolithic guesses, it uses Orthogonal Predictive Factorization to split latent states into independent components before recombining them. The repository provides core code and task-design tools, though pre-trained weights are not included.
Anthropic just overhauled Projects in Claude Code into an agentic coordinator for complex workflows. You describe an objective, and Claude scopes the work, spins up parallel cloud threads across different branches or repositories, runs tests, and opens PRs while maintaining shared project memory. It is currently rolling out in beta to select Claude Pro and Max subscribers.
xAI just rolled out persistent memory for its agentic coding tool Grok Build, ending the pain of re-explaining project rules in every new chat. After each turn, Grok quietly reviews the interaction in the background and writes key conventions and architectural decisions into Markdown files. When you start fresh sessions, it reads those topic files before touching related code—and you can inspect or organize everything with simple /memory and /dream commands.
TypeSafe AI just introduced Jev, a "System One" model built for automated software decisions rather than chat. Instead of generating tokens sequentially, Jev outputs typed, probabilistic answers in parallel, hitting 70–500ms response times with free output tokens. It gives developers a cheap, hallucination-free way to run fuzzy logic in code, though it only works when possible options are defined upfront.
CopilotKit just open-sourced OpenMuse, a personal AI agent you can self-host with a persistent Chromium browser, an isolated Linux terminal, and a local file workspace. It works through multi-step plans continuously and lets you take manual control whenever you want. Why it matters: If you want an autonomous assistant for emails and browsing without handing your personal data to big tech, this gives you full control.
Cline just launched Cline Desktop for macOS and Windows, a dedicated environment for running open-weight AI models. Users can tap free models via ClinePass or plug in their own API keys. It also packs scheduled agent runs, an MCP server marketplace, and cross-tool task imports, streamlining developer workflows.
OpenAI just launched the public beta of its Agents API, giving developers access to the exact same harness and infrastructure powering Codex. You can spin up production-ready agents via a single API call with automatic context compaction, parallel subagents, and managed sandboxes at no added fee beyond standard token costs.
DeepSeek just launched DeepSeek-V4.1-Flash, a 552B-parameter MoE model with native visual understanding. Built on a Causal Encoder-Decoder architecture, it activates just 8B parameters on input and 16B on output, outpacing the flagship V4-Pro while slashing KV cache to 1/4 the HBM and 1/8 the SSD storage.
Google just open-sourced ARTEMIS, an automation framework that lets coding assistants like Claude Code, Cursor, and Codex drive real Android devices or emulators via MCP. It runs cross-app workflows, grabs logcat crashes and screenshots, and scores a 99%+ success rate on AndroidWorld.
OpenAI just revealed that its coding agents generate 3.1 agent-workdays for every human workday inside its lab, with the median researcher running through over $600 in daily inference. The company says it hit its target of creating an "automated research intern" and aims for a full AI researcher by March 2028. Why it matters: AI is genuinely speeding up its own development, though recent safety halts prove human control is still catching up.
Anthropic just open-sourced the first complete, machine-checked proof of Fermat's Last Theorem in Lean 4, generated entirely by AI agents. Spanning over 13 million lines of code and 29,511 theorems, the massive artifact successfully passed checks by two independent verification kernels without extra axioms. It marks a historic milestone for AI-driven mathematics.
Google just released Gemini 3.8 Flash and 3.8 Flash Cyber, boosting agentic coding and security defenses while keeping pricing at $0.75 per million input tokens. Flash Cyber hits over 70% vulnerability discovery across 20 languages and produces 2.6x more correct patches for Chrome than larger models. Why it matters: Google is proving that specialized, agentic workhorses can beat brute-force frontier models on both cost and precision.
Meta just launched Muse Spark 1.3 across Muse Code and the Meta Model API. The update tackles long-horizon agent workflows by asking clarifying questions on vague prompts, while slashing tool calls by roughly 20% and token usage by 25% during coding tasks. It also bolsters prompt-injection defenses, with open weights promised down the road.
Alibaba just rolled out Qwen3.8-Max-0902, an upgraded snapshot targeting complex engineering code and collaborative agent workflows. It keeps the 1M context window and thinking mode while sharpening tool orchestration and multimodal vision, priced at $2 per million input tokens and $6 for output.
OpenAI's upcoming Astra model reportedly relies on a looped transformer technique called recurrent depth, which passes tokens through the same layers repeatedly to boost coding and math performance. The problem is that it obscures the model's chain-of-thought reasoning, making internal steps much harder to monitor. OpenAI Chief Scientist Jakub Pachocki has already warned against racing toward opaque systems.
