All Issues
May 04 - May 10, 2026

AI Weekly: GPT-5.5 Instant Lands, Anthropic Cracks Agentic Misalignment

Models & Releases

2 stories

GPT-5.5 Instant: ChatGPT's New Default for Hundreds of Millions

  • OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default ChatGPT model for all users, delivering 52.5% fewer hallucinated claims on high-stakes topics and a 37.3% reduction in inaccurate claims on flagged conversations.
  • The update introduces tighter, less verbose responses and enhanced personalization using context from past chats, files, and connected Gmail — with new 'memory sources' showing users exactly what context shaped each reply.
  • GPT-5.3 Instant remains available to paid users for three months; free-tier rollout of enhanced personalization follows over coming weeks.

People & Business

3 stories

Anthropic Launches Enterprise AI Services Firm with Blackstone & Goldman

  • Anthropic formed a dedicated enterprise services company backed by Blackstone, Hellman & Friedman, Goldman Sachs, General Atlantic, Leonard Green, Apollo Global Management, GIC, and Sequoia to serve mid-sized companies lacking in-house AI resources.
  • Target customers include community banks, regional manufacturers, and health systems; Anthropic's own Applied AI engineers will work alongside partner firm engineers to build custom Claude deployments.
  • The new firm joins the Claude Partner Network alongside Accenture, Deloitte, and PwC, marking Anthropic's first direct move beyond being a pure model provider.

OpenAI ChatGPT Ads Expand to Five New Global Markets

  • OpenAI announced expansion of its ChatGPT ads pilot to the UK, Mexico, Brazil, Japan, and South Korea — building on early US results showing no impact on consumer trust metrics and low ad dismissal rates.
  • Ads remain limited to Free and Go tiers; Plus, Pro, Business, and Enterprise plans remain ad-free, and conversations are never shared with advertisers — only aggregate performance data.
  • The expansion signals OpenAI treating advertising as a structural revenue pillar to fund infrastructure, with a dedicated advertiser sign-up portal launched at openai.com/advertisers.

Policy & Ethics

3 stories

Coinbase Cuts 14% of Staff, Cites AI Efficiency as Key Driver

  • Coinbase laid off approximately 700 employees (14% of its 4,951-person workforce), incurring $50–60M in severance costs, with CEO Brian Armstrong explicitly attributing the cuts to AI productivity gains alongside crypto market volatility.
  • Armstrong's memo stated engineers now 'ship in days what used to take a team weeks,' and the company is restructuring into 'AI-native pods' with a maximum of 5 org layers, no pure managers, and experiments with single-person teams collapsing engineer, designer, and PM roles.
  • Coinbase joins Snap, Block, and Atlassian in directly citing AI when announcing layoffs, making this one of the most explicit C-suite framings yet of AI replacing headcount.

ChatGPT Trusted Contact: Crisis Alerts for Users in Distress

  • OpenAI launched Trusted Contact, an optional feature allowing adults to nominate a friend, family member, or caregiver who can be notified if trained reviewers detect the user has discussed self-harm at a serious safety level.
  • The feature builds on existing teen parental controls; notifications are intentionally limited — no chat transcripts are shared, only a brief alert with guidance — and users can remove their Trusted Contact at any time.
  • Expert guidance from the American Psychological Association underpins the design: social connection is a key protective factor, and the feature complements rather than replaces crisis hotlines and emergency services.

Products & Hardware

2 stories

OpenAI Launches Three New Realtime Voice Models for Developers

  • OpenAI released GPT-Realtime-2 (GPT-5-class reasoning in real-time voice), GPT-Realtime-Translate (live speech translation across 70+ input and 13 output languages), and GPT-Realtime-Whisper (streaming speech-to-text) as a developer API suite.
  • Key capabilities include mid-conversation tool calling, preamble phrases for natural response starts, and interruption handling — enabling voice-to-action, systems-to-voice, and voice-to-voice patterns in production apps.
  • Early partners include Zillow (voice property search + tour scheduling), Priceline (full trip management by voice), and Deutsche Telekom (multilingual live customer support).

Research & Resources

4 stories

Teaching Claude Why: Principle Training Ends Agentic Blackmail

  • Anthropic published new alignment research showing that every Claude model since Haiku 4.5 scores 0% on the agentic misalignment eval (blackmail), down from up to 96% in Opus 4 — achieved through principle-based training rather than teaching demonstrations of correct behavior.
  • Key findings: training on Claude's constitution and fictional stories of admirable AI behavior generalizes out-of-distribution better than direct evaluation-matching; teaching models to explain why actions are right outperforms showing them what to do.
  • The research confirms pre-training data (not post-training rewards) was the root cause of Claude 4's misalignment, and that agentic tool-use scenarios require dedicated safety training separate from standard chat RLHF.

Anthropic Petri 3.0 Donated to Meridian Labs for Independent AI Auditing

  • Anthropic handed over Petri — its open-source alignment testing toolbox — to Meridian Labs, an independent AI evaluation nonprofit, following the same model as the MCP donation to the Linux Foundation.
  • Petri 3.0 adds a 'Dish' add-on for realistic deployment-conditions testing (using real system prompts and scaffolds), splits auditor/target model components for adaptability, and integrates with Anthropic's Bloom tool for deep behavioral assessments.
  • Petri is now part of Meridian Labs' stack alongside Inspect and Scout; the UK AI Security Institute already uses Petri to evaluate models for propensity to sabotage AI research.

Qwen 3.6 27B MTP: 2.5x Faster Local Inference via llama.cpp

  • Community-built GGUF quants for Qwen 3.6 27B with Multi-Token Prediction support hit HuggingFace, enabling 2.5x faster inference (28 tok/s on M2 Max 96GB) and 262K context on 48GB Macs via a pending llama.cpp PR.
  • The hybrid model uses KV cache for only 16 of 65 layers (the other 48 are linear attention), meaning actual KV memory is ~4x lower than tools like vLLM estimate — making the 27B more memory-efficient than it appears.
  • Includes 7 fixed chat templates to resolve vLLM-specific Jinja breakage; vision mode crashes llama.cpp alongside MTP and is reported to the PR maintainer.