OpenAI's Chief Revenue Officer sent an internal memo to employees this week that was not supposed to go public. It did. The memo, reviewed by CNBC, is unusually blunt about where OpenAI stands competitively, and a lot of it is about Anthropic.

The language is direct. Claude has "become a religion" among developers. Anthropic's strategy is built on "fear, restriction, and the idea that a small group of elites should control AI." They made a "strategic misstep to not acquire enough compute," meaning they did not secure enough of the computing power needed to scale, and are now falling behind on infrastructure. OpenAI's ramp, the memo says, is "materially ahead and widening."

The Microsoft section is equally candid. The partnership that helped build OpenAI has also "limited our ability to meet enterprises where they are," specifically pointing to AWS Bedrock, Amazon's cloud platform where many large companies prefer to run their AI tools. OpenAI is now pivoting toward Amazon as its primary compute partner, backed by a $50 billion investment, and is targeting 30 gigawatts of compute capacity by 2030 compared to Anthropic's 8.

The memo also challenges Anthropic's revenue numbers directly. OpenAI claims Anthropic's reported $30 billion annual run rate is inflated by around $8 billion because of how they count revenue sharing deals with Amazon and Google. By OpenAI's math, Anthropic's real number is closer to $22 billion. Anthropic has not responded publicly.

A company does not write a memo like this about a competitor it is not worried about.

Claude experienced a significant outage this morning during peak hours. Users across claude.ai, the API, and Claude Code all hit server errors, meaning the servers were crashing rather than simply being busy. Anthropic's status page showed all systems operational the entire time. For people paying $200 a month, that is a hard thing to ignore.

This is not isolated. There were server errors the week before, authentication failures the week before that. A different kind of failure every week.

Separately, Claude Opus 4.6's performance has been raising questions. Bridgebench, a third party benchmark that tests how often AI models make things up, retested Opus 4.6 this week and the numbers dropped hard. It fell from second place at 83.3% accuracy to tenth at 68.3%. No explanation from Anthropic. To make it stranger, the older Opus 4.5 is now outperforming it on the same benchmark.

The most likely cause is either a quiet model update pushed without announcement, or a shift in computing resources toward the next model internally. Opus 4.7 has apparently been spotted on an internal API endpoint, suggesting it is being tested at scale. If compute is being redirected toward the newer model, the current one would naturally run leaner and perform worse.

An Anthropic engineer said the company does not degrade models to manage demand. The benchmark data tells a different story. Either way, users deserve an explanation and have not gotten one.

OpenRouter quietly listed a new model this week called Elephant Alpha. A 100 billion parameter model, free to use, with a 256K context window and solid performance on coding, debugging, and document tasks. No lab name. No announcement. Just a stealth listing and a note that your prompts will be logged.

The AI community immediately started reverse engineering it. By analysing its internal structure, token patterns, and response style, users concluded it is almost certainly an unreleased model from Alibaba's Qwen team. OpenRouter has not confirmed this.

Labs are increasingly using OpenRouter to quietly test models at scale before any official announcement, collecting real world usage data in the process. Free access is the offer. Your prompts are the price.

Elephant Alpha is live on OpenRouter now if you want to try it.