#1

Anthropic leaked its own model. OpenAI killed Sora. MCP hit 97 million installs.

It's been a month. Let's get into it.

The big stories

Anthropic accidentally leaked its most powerful model. A draft blog post about "Claude Mythos" was found in an unsecured, publicly searchable data store. Anthropic confirmed the model exists, called it a "step change" in capabilities, and acknowledged it was human error in their CMS configuration. The leaked details are striking: Mythos dramatically outscores Opus 4.6 on coding, reasoning, and cybersecurity benchmarks. Anthropic is privately warning government officials that it makes large-scale cyberattacks significantly more likely. The model is being tested by early access customers but is extremely compute-intensive -- no public release date yet.

OpenAI killed Sora after 6 months. The AI video generator was burning $15 million per day in inference costs against $2.1 million in total lifetime revenue. Users peaked at a million and dropped to under 500,000. Disney had committed $1 billion to the partnership and found out about the shutdown less than an hour before the public announcement. Sam Altman told employees the compute is being redirected to "Spud," an upcoming model he says will "meaningfully accelerate the overall economy." No details on what Spud actually is.

MCP crossed 97 million installs. OpenAI announced full MCP support across their products. The spec jumped to v2 with Streamable HTTP transport and OAuth 2.1 baked in. Six months ago, MCP was an Anthropic experiment. Now every major AI company is building on it.

CLI tool updates

Claude Code got push-to-talk voice mode, computer use (Claude can open files and click around your screen), Opus 4.6 as default with a 1M token context window, and /loop for recurring tasks. That last one is quietly the most useful -- it's basically a session-level cron job for your AI agent.

Codex went deep on multi-agent workflows. Sub-agents use readable path-based addresses, plugins are now first-class, and the app-server can send shell commands and watch filesystem changes over WebSocket. GPT-5.4 mini and Codex-Spark (research preview) landed in the CLI.

Gemini CLI shipped native sandboxing (bubblewrap + seccomp on Linux, Seatbelt on macOS) and turned on subagents by default with parallel tool scheduling. Gemini Pro models are now paid-only in the CLI -- free tier gets Flash only.

Junie CLI entered beta. JetBrains' coding agent is now a standalone CLI tool, not just an IDE plugin. LLM-agnostic -- works with OpenAI, Anthropic, Google, and Grok models.

Starlette hit 1.0 after 8 years on zero-ver. Not an AI tool, but it underpins FastAPI and the Python MCP SDK. 325 million downloads a month. If you're building MCP servers in Python, your foundation just went stable.

The take

The CLI is winning. Not the IDE, not the browser, not the chat window. The command line.

Claude Code, Codex, Gemini CLI, Junie -- they all converged on the same form factor: an agent that lives in your terminal, reads your repo, edits files, runs tests, and talks to external services through MCP. The terminal went from "where you run git commands" to "where your AI agent lives."

The interesting question isn't which CLI tool is best right now. It's what happens when these agents start talking to each other through MCP, running as background processes, and operating on schedules instead of waiting for you to type a prompt.

We're closer to that than most people think. Claude Code's /loop and Codex's multi-agent v2 are early versions of exactly that pattern.

One thing

If you're still using AI through a chat window and copy-pasting code back into your editor, try one of the CLI tools for a week. Any of them. The workflow difference isn't incremental -- it's a different way of working. A good terminal helps too -- Yaw was built for exactly this kind of workflow.

Token Limit News

Token Limit News