Monday: A mystery model called Elephant Alpha appeared on OpenRouter with no lab name, no announcement, and a 256K context window. Free to use. Your prompts are the price.

Tuesday: Claude Code got a full desktop redesign with parallel session support, integrated terminal, and a side chat that lets you ask questions without interrupting the agent mid-task. OpenAI launched GPT-5.4-Cyber, a less restricted version of its model for verified security researchers. Microsoft updated Copilot in Word with tracked changes, inline editing, and organizational context through Work IQ.

Wednesday: Google shipped subagents for Gemini CLI, letting developers run parallel specialist agents across complex tasks simultaneously. Google also launched Gemini 3.1 Flash TTS, a text-to-speech model with over 200 audio tags for director-level control over delivery, scoring competitively against ElevenLabs. And Gemini landed on Mac with a global shortcut and window-sharing for contextual help.

Thursday: Anthropic launched Claude Opus 4.7, scoring 64.3% on standard coding benchmarks, ahead of GPT-5.4's 57.7%. OpenAI updated Codex to run in the background on your Mac. OpenAI also launched GPT-Rosalind, a model for biology, chemistry, and protein research.

Friday: Anthropic launched Claude Design, a conversational design tool that exports to Canva, PDF, PowerPoint, or directly into Claude Code. Google launched Flow Music, letting you generate fully arranged tracks from a text prompt.

The leaked OpenAI memo is the story of the week. The Chief Revenue Officer sent an internal note that was not meant to go public, and it reads like a company taking stock of a fight it is no longer sure it is winning. Claude has "become a religion" among developers. Anthropic made a "strategic misstep" on compute. The Microsoft partnership has "limited" OpenAI's ability to reach enterprise customers where they actually are. They also argue Anthropic's reported $30 billion annual run rate is inflated by around $8 billion due to how revenue-sharing deals with Amazon and Google get counted. Anthropic has not responded.

The other rough patch was Anthropic's own. Claude went down during peak hours while the status page showed all systems operational, the third different category of failure in as many weeks. Opus 4.6's Bridgebench accuracy scores also dropped from 83.3% to 68.3% with no explanation, while the older Opus 4.5 outperformed it on the same test. Anthropic denied quietly degrading the model, but neither of the plausible explanations is reassuring if you are paying $200 a month.

The three OpenAI departures deserve more than a bullet point. Bill Peebles built Sora, Kevin Weil ran the science division, and Srinivas Narayanan led B2B. All three left on the same day, and the products they built are either shelved or folded into Codex. OpenAI is narrowing hard ahead of Spud. Whether that kind of organizational tightening helps or hurts the landing is a real question.

The memo is the most honest thing OpenAI has published in years, and they did not mean to publish it. Anthropic is the competitor they are losing sleep over. Everything else this week, the departures, the org narrowing, the Codex push, reads as a company trying to close the gap before Spud arrives.

Anthropic gave them reasons to worry and reasons to feel better in the same week. The product momentum is real. So are the reliability issues. A lab that talks about safety more than anyone else having unexplained benchmark drops and outages without public acknowledgment is a trust problem, not just a technical one.

Google had the quietest good week in the industry. Three major AI tools, zero drama. While OpenAI and Anthropic spent the week generating headlines about memos and model degradation, Google just shipped.

Buckle up for next week.