| **This Week in AI**
OpenAI dominated the week with aggressive moves to expand developer access and ship specialized models. After its GPT-5.5 party sold out, the company extended 10× Codex rate limits to every applicant through June 5, effectively turning a logistics failure into a month-long coding sprint for thousands of builders. It also pushed GPT-5.5 Instant to all ChatGPT users and released three purpose-built realtime audio models, signaling a clear bet on voice-first agents and lower-friction experimentation.
DeepSeek countered with the full V4 technical paper, detailing FP4 quantization-aware training that keeps quality high while slashing KV cache and FLOPs. The release underscores how open-weight labs are closing the efficiency gap on frontier models. Meanwhile, Anthropic advanced constitutional alignment techniques and Google delivered practical inference speedups, showing the field is simultaneously racing on capability, cost, and safety.
**Model Tracker**
- **GPT-5.5 Instant** (OpenAI): lighter default ChatGPT model with improved factuality in medicine, law, and finance; API endpoint `gpt-5.5-chat-latest`; enhanced memory across chats, files, and Gmail.
- **GPT-Realtime-2 / Translate / Whisper** (OpenAI): three new Realtime API models for live voice reasoning, 70+ language speech-to-speech translation, and streaming transcription.
- **DeepSeek V4 Pro** (DeepSeek): trillion-parameter MoE with FP4 QAT; ties GPT-5.2 on FoodTruck Bench at ~17× lower cost; 62.7 % win rate over Gemini 3.1 Pro on Chinese writing.
- **Gemini 3.1 Flash Lite** (Google): exits preview with unchanged pricing; stable low-cost option for lighter workloads.
**Top Stories**
OpenAI emailed more than 8,000 developers who applied for its sold-out GPT-5.5 party with immediate 10× Codex rate limits valid through June 5. The boost applies regardless of acceptance status and gives builders a high-volume window for agentic coding just as Anthropic runs competing events. [▶ Episode 40 · 2026-05-05](https://nerranetwork.com/blog/models_agents/ep040.html)
OpenAI began rolling GPT-5.5 Instant to every ChatGPT user with measurable gains in factuality, image analysis, and memory that pulls from saved chats, files, and Gmail. The lighter model trades some capability for speed and cost, a deliberate choice Simon Willison noted. [▶ Episode 41 · 2026-05-06](https://nerranetwork.com/blog/models_agents/ep041.html)
DeepSeek published the complete V4 paper covering FP4 quantization-aware training in late-stage MoE optimization, anticipatory routing for stability, and generative reward modeling. Human evals show strong Chinese writing performance and users rating it ready as a default coding model. [▶ Episode 43 · 2026-05-09](https://nerranetwork.com/blog/models_agents/ep043.html)
OpenAI released GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper through the Realtime API, letting developers mix live voice reasoning, multi-language translation, and streaming transcription in one session. The models sit alongside existing endpoints rather than replacing generalist GPTs. [▶ Episode 42 · 2026-05-08](https://nerranetwork.com/blog/models_agents/ep042.html)
Anthropic shared new Model Spec Midtraining research showing that pre-training on detailed constitutions improves safety generalization in agentic settings, cutting misalignment by more than 3× in some tests. The work complements OpenAI’s automated CoT-grading detection updates. [▶ Episode 41 · 2026-05-06](https://nerranetwork.com/blog/models_agents/ep041.html)
**Agent & Tool Updates**
Developers can now route voice agents through dedicated realtime endpoints for reasoning, translation, or transcription without forcing everything through a single model. The 10× Codex boost gives immediate headroom for high-volume agentic coding experiments. Caliby, a new embedded vector database, launched with strong disk performance aimed at long-term agent memory. Google’s MTP Drafters deliver up to 3× faster Gemma 4 inference with no quality loss, useful for production agent pipelines.
**Open Source Spotlight**
A ggml port of Microsoft’s VibeVoice ships as a single binary supporting CPU, CUDA, and Metal for TTS, long-form ASR, and diarization with zero Python at inference time. Qwen3.6-35B-A3B received an uncensored fine-tune that preserves all 19 native MTP tensors. Community benchmarks on Qwen 3.6 27B quantization tradeoffs continue to surface practical recipes for local deployment.
**Safety & Regulation**
Anthropic’s Model Spec Midtraining work and OpenAI’s automated detection of CoT grading during RL both target safer agent behavior at training time. These techniques aim to reduce misalignment without heavy post-hoc filtering.
**What to Watch Next Week**
Watch for broader rollout of GPT-5.5 Instant personalization to mobile, community reaction to the realtime audio models in agent workflows, and any follow-up DeepSeek efficiency benchmarks on long-context or multi-agent tasks. |
