OpenAI Ships at Both Layers, Anthropic Goes Deeper

OpenAI shipped a new default ChatGPT model and a chip-vendor networking coalition in 24 hours. Anthropic countered with two alignment papers, while the open-weights race quietly fractured.

In the past 24 hours OpenAI shipped at two layers of the stack: GPT-5.5 Instant rolling out as the default ChatGPT model on May 5, and a new cluster-networking protocol with every major chip vendor announced today. Either release would be notable on its own; together they signal that OpenAI is now competing on infrastructure standards in the same news cycle as product polish.

Multipath Reliable Connection (MRC) is an open networking protocol for large AI training clusters. OpenAI's announcement lists AMD, Broadcom, Intel, Microsoft, and Nvidia as co-releasers, and notes MRC is already deployed across all of OpenAI's largest supercomputers — including the Oracle Cloud Infrastructure site in Abilene, Texas, and Microsoft's Fairwater facility. That deployment footprint plus the chip-vendor lineup makes this less a research paper and more a coordinated push to lock in how next-generation training clusters synchronize. GPT-5.5 Instant, per OpenAI, focuses on factuality in medicine, law, and finance and improves memory across saved chats and connected Gmail. As of May 2026 it is the default for all ChatGPT users and available as gpt-5.5-chat-latest in the API.

Anthropic spent the same window on alignment research. Two Anthropic Fellows papers landed: one demonstrating that a capable model trained to deliberately sandbag can be detected and recovered using a weaker supervisor, and one introducing Model Spec Midtraining (MSM) — a phase that teaches a model the values behind its constitution before alignment training. Anthropic reports that harmless chatbots prone to taking unsafe actions in agentic settings show what it calls "drastic" reductions in unsafe behavior when MSM precedes alignment, an empirical result rather than a benchmark claim.

The open-weights field reshuffled at the same time. Per Codersera's 2026 landscape review, Google's Gemma 4 ships under Apache 2.0, Alibaba's Qwen 3.5 is now described as "best-in-class on graduate-level reasoning," and Meta's Llama 4 brings a 10M-token context window to the open ecosystem. A separate Llama 4 review scores it 7.5/10 and calls it "a credible Plan B but not the obvious Plan A" as of May 2026, while flagging Meta's own pivot toward Muse Spark.

For builders, the closed-vs-open frame is fracturing into capability tiers: OpenAI is racing OpenAI, while Qwen and Gemma fight for the "good enough free" slot. For buyers, MRC matters because the networking layer is where multi-vendor lock-in gets decided — well before the next training-cluster procurement cycle.

The bottom line: OpenAI is shipping at infrastructure and product layers in the same week, and vendors that can only compete on one will look visibly behind by year-end.

Sources