OpenAI and Thrive Holdings just detailed Tax AI, an agent built with Codex that turns accountant corrections into structured evaluations to fix its own bugs. Piloted across Crete’s 30+ accounting firms on 7,000 tax returns, the system cut prep time by a third while its 75% accuracy rate jumped from 25% to 86% in six weeks. It shows how coding agents can autonomously maintain enterprise workflows in production.
Tech giants are hitting the brakes on autonomous coding tools after costs spiraled out of control. Uber blew through its entire 2026 AI coding budget in just four months, while Microsoft is cutting direct Claude Code licenses. Why it matters: Agentic workflows burn tokens faster than expected, forcing companies to admit cheaper tokens don't automatically mean cheaper AI.
Datacurve just released DeepSWE, a new benchmark tailored for agentic coding models. While frontier models often post nearly identical scores on existing leaderboards, DeepSWE simulates day-to-day developer workflows to expose where their actual capabilities diverge.
Anthropic just announced two enterprise-focused upgrades for Claude Managed Agents at its Code w/ Claude event in London. Self-hosted sandboxes move tool execution to your own infrastructure or cloud providers like Vercel and Cloudflare, keeping sensitive code inside your perimeter. Meanwhile, new MCP tunnels let agents securely reach internal MCP servers without exposing public endpoints. It basically solves one of the biggest security blockers for deploying coding agents in production.
Google DeepMind just introduced AlphaProof Nexus, an AI agent that solved nine open Erdős math problems and 44 OEIS conjectures. The system pairs an LLM with the Lean proof assistant to automatically generate and verify proofs, running at a cost of just a few hundred dollars per problem. Why it matters: AI is moving past standard contest benchmarks into genuine mathematical discovery, though problems demanding brand-new mathematical constructions remain out of reach.
xAI just launched Grok Build in beta for SuperGrok and X Premium+ subscribers. The developer suite brings a Plan Mode for task mapping, an Imagine tool to generate images and videos, and a dedicated CLI for crafting automations and orchestrators. It gives Grok power users a direct route from idea to finished agent workflow.
Ideogram just launched an MCP integration that turns AI assistants like Claude and ChatGPT into design agents. Users can now generate images, build designs, and even train custom models without leaving the chat interface. It also works with Cursor and Hermes, embedding creative workflows right into your existing tools.
Mistral just launched Mistral Medium 3.5, an open-weight 128B model powering new asynchronous cloud agents in Vibe and Le Chat. Instead of locking up your local terminal, agents now run across parallel cloud sandboxes and automatically open GitHub PRs when finished. It hits 77.6% on SWE-Bench Verified while fitting on as few as four GPUs.
OpenAI just landed in the Leader quadrant of Gartner's Enterprise AI Coding Agents report. Its Codex platform now counts over 4 million weekly users, with enterprise customers like Cisco using it to slash software delivery times from quarters to weeks. Why it matters: Coding assistants are rapidly graduating from simple autocomplete into full-blown autonomous software developers.
Microsoft just rolled out MagenticLite, a new agentic setup designed to run complex workflows across web browsers and local file systems using small models. It pairs a 14B orchestrator called MagenticBrain with Fara1.5, a specialized computer-use model that nearly doubles web navigation performance over its predecessor. It proves you don't need massive frontier models to build capable agents on local hardware.
Arq.live just unveiled GLITCH, a voice-to-film system that lets directors cast characters, build worlds, and block scenes simply by speaking. Multiple AI agents work in sequence like a virtual crew, producing a complete short film from a single laptop in under eight minutes. It moves AI video away from prompt typing toward how filmmakers actually direct.
ByteDance just open-sourced Bernini, a unified video framework pairing an MLLM semantic planner with a DiT renderer. It handles generation and editing across text, images, and video in a single pipeline, putting AI video closer to real directing. Code and model weights are now live on GitHub and Hugging Face.
Alibaba just introduced Qwen3.7-Max, a flagship proprietary model engineered specifically for long-horizon agent execution and multi-step coding. It integrates vision capabilities and plugs cleanly into frameworks like Claude Code. Why it matters: Alibaba is pushing hard into autonomous workflows that sustain thousands of steps, though this variant remains closed behind Alibaba Cloud APIs for now.
Figma just added native AI agents right into its collaborative canvas, letting designers create, edit, and iterate on files using plain text prompts. The design-tuned assistant even lets you run multiple agents at the same time for different tasks. Why it matters: you can now offload tedious layout tweaks and mockups without ever leaving the artboard.
OpenAI just revealed that an internal general-purpose reasoning model autonomously disproved the planar unit distance conjecture, first posed by Paul Erdős in 1946. By applying sophisticated tools from algebraic number theory to discrete geometry, the model overturned a decades-old belief about grid constructions. Fields medalist Tim Gowers and other mathematicians verified the proof, marking a massive leap for AI in scientific discovery.
Granola just launched Briefs, an AI feature that automatically preps you for calls right as you join. Overnight, it searches your past notes, connected Gmail threads, and the web to create a cited, 2-3 bullet summary of who you're meeting and where you left off. No prompting needed—it's designed to kill the last-minute pre-call Google panic.
Google just supercharged Gemini at I/O 2026, officially turning it from a reactive chatbot into an active digital coworker. The biggest leap is Gemini Spark, an autonomous cloud agent that keeps handling your Gmail and Docs even with your laptop shut, paired with a new cinematic video generator called Gemini Omni.
Google just gave Search a massive AI overhaul, making Gemini 3.5 Flash the default engine for AI Mode and debuting a redesigned search box that takes files, videos, and Chrome tabs. It is also rolling out 24/7 web-monitoring agents and on-the-fly coding that generates custom dashboards and mini apps. Search is officially moving from answering queries to building software for you.
Google just unveiled the Gemini 3.5 model family, kicking things off with 3.5 Flash. It beats Gemini 3.1 Pro across key coding and agentic benchmarks while outputting tokens four times faster than rival frontier models. The model is already live in the Gemini app, Google Search AI Mode, and APIs, with 3.5 Pro slated for next month.
Google just launched Gemini for Science, a new suite of agentic research tools on Google Labs designed to supercharge the scientific method. The prototypes handle multi-agent hypothesis debates, parallel code generation for computational experiments, and literature synthesis via NotebookLM. Two accompanying research papers were published in Nature.
