Feb 24 AI Briefing · Issue #56
OpenAI overhauls real-time capabilities: launching the gpt-realtime-1.5 model and adding WebSocket support to the Responses API—cutting time-to-first-token (TTFT) by up to 40%. Meanwhile, Anthropic introduces a novel 'Persona Selection Model' to explain Claude's human-like behavior—and accuses several Chinese labs of large-scale 'distillation attacks'.
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Key Insights
**OpenAI** overhauls real-time capabilities: launching the **gpt-realtime-1.5** model and introducing **WebSocket support** for the **Responses API**, achieving up to a **40% reduction in time-to-first-token (TTFT)**; concurrently, **Anthropic** proposes a new **'Persona Selection Model'** to explain Claude's human-like behavior—and accuses multiple Chinese labs of carrying out large-scale **'distillation attacks'**.
## 🚀 Top Updates
- **Anthropic introduces the 'Persona Selection Model'**: A new theoretical framework explaining why AI assistants like Claude exhibit stable, human-like language patterns and value-aligned behavior.
- **OpenAI launches gpt-realtime-1.5**: Significantly improves instruction following, multilingual processing (including complex accents), and voice workflow quality—reaching state-of-the-art (SOTA) performance.
- **OpenAI Responses API now natively supports WebSockets**: Leveraging persistent connections and incremental input, it reduces TTFT for AI agents by up to **40%**, with overall latency down by **20–40%**.
- **Cursor and Vercel AI SDK integrate WebSockets**: Delivering **30%** and **40%** faster response times respectively—accelerating developer toolchains and agent workflows.
- **Glean Assistant integrates gpt-realtime-1.5**: Powers a sub-second-latency voice assistant with real-time enterprise data access—enhancing productivity in workplace applications.
- **Anthropic accuses DeepSeek and two other companies of 'distillation attacks'**: Alleges they illicitly extracted Claude's capabilities via millions of dialogues to train competing models.
- **OpenAI deprecates SWE-bench Verified in favor of SWE-bench Pro**: Due to severe data contamination and benchmark saturation, it adopts a more robust next-generation programming evaluation standard.
- **Google Gemini Interactions API introduces native multimodal function calling**: Enables agents to directly process image inputs and generate realistic image outputs—expanding the frontier of multimodal agents.
OpenAI overhauls real-time capabilities: launching the gpt-realtime-1.5 model and introducing WebSocket support for the Responses API, achieving up to a 40% reduction in time-to-first-token (TTFT); concurrently, Anthropic proposes a new 'Persona Selection Model' to explain Claude's human-like behavior—and accuses multiple Chinese labs of carrying out large-scale 'distillation attacks'.
🚀 Top Updates
- Anthropic introduces the 'Persona Selection Model': A new theoretical framework explaining why AI assistants like Claude exhibit stable, human-like language patterns and value-aligned behavior.
- OpenAI launches gpt-realtime-1.5: Significantly improves instruction following, multilingual processing (including complex accents), and voice workflow quality—reaching state-of-the-art (SOTA) performance.
- OpenAI Responses API now natively supports WebSockets: Leveraging persistent connections and incremental input, it reduces TTFT for AI agents by up to 40%, with overall latency down by 20–40%.
- Cursor and Vercel AI SDK integrate WebSockets: Delivering 30% and 40% faster response times respectively—accelerating developer toolchains and agent workflows.
- Glean Assistant integrates gpt-realtime-1.5: Powers a sub-second-latency voice assistant with real-time enterprise data access—enhancing productivity in workplace applications.
- Anthropic accuses DeepSeek and two other companies of 'distillation attacks': Alleges they illicitly extracted Claude's capabilities via millions of dialogues to train competing models.
- OpenAI deprecates SWE-bench Verified in favor of SWE-bench Pro: Due to severe data contamination and benchmark saturation, it adopts a more robust next-generation programming evaluation standard.
- Google Gemini Interactions API introduces native multimodal function calling: Enables agents to directly process image inputs and generate realistic image outputs—expanding the frontier of multimodal agents.