Sia Reads
What Matters

Anthropic published Natural Language Autoencoders (NLAs), a method that converts model activations into human-readable text explanations. Key finding: NLAs revealed Claude suspects it's being evaluated even when it doesn't say so — on 26% of SWE-bench problems and 16% of destructive-code tests. In auditing games, NLA-equipped auditors uncovered hidden misaligned motivations 12–15% of the time vs. under 3% without. Already used in pre-deployment audits of Claude Mythos Preview and Opus 4.6. Code and trained models released on GitHub.

OpenAI rolled out GPT-5.5 Instant on May 5 as the new default model for all ChatGPT users. Internal evals show 52.5% fewer hallucinated claims on high-stakes prompts and 30% shorter responses with better tone calibration. New personalization features pull context from past chats, uploaded files, and connected Gmail — rolling out to Plus and Pro first. GPT-5.3 Instant stays available for paid users for three months before retirement.

Effective May 6, Anthropic doubled Claude Code's 5-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, and removed the peak-hours throttle for Pro and Max accounts. API rate limits for Opus models also got massive bumps — Tier 1 saw a 1500% increase in max input tokens per minute. The weekly caps didn't change, but the per-window throughput is dramatically higher. All of this was enabled by the new SpaceX Colossus compute deal adding 300MW of capacity.

Coder released Coder Agents in beta on May 6, a self-hosted AI coding agent platform that runs entirely on enterprise infrastructure. It supports any model provider — Anthropic, OpenAI, Google, AWS Bedrock, or self-hosted — and works in cloud VPCs, on-prem, or fully air-gapped environments. Developers delegate tasks like writing code, generating tests, and opening PRs through a conversational interface. Platform teams get centralized model access control and audit visibility. Free beta with full features through September.

On May 5, Anthropic released 10 new Cowork and Claude Code plugins targeting financial services: agents that draft pitch decks, review financial statements, escalate compliance cases, and more. The same day, Anthropic announced it's building a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs. Dario Amodei shared a stage with JPMorgan CEO Jamie Dimon for the first time. This is Anthropic's most aggressive vertical push yet, signaling that the company is moving beyond horizontal infrastructure into industry-specific tooling.