Archive

Today's Key Insights

  • OpenAI Makes Codex-Powered Agents Public Beta — For developers, OpenAI now manages orchestration, long-running sessions and tool use through a Codex-powered service, while DeepSeek's V4.1-Flash cuts KV-cache memory to one-quarter of its predecessor.
  • Anthropic Alleges Escalating Distillation Attacks From Three China-Based AI Firms — The report puts Alibaba, Moonshot AI, and DeepSeek at the center of Anthropic's allegations as competition intensifies and alleged distillation attacks escalate.
  • Huang Forecasts 70% Nvidia Growth Next Year — Nvidia investors are getting a bullish 70% growth forecast from a CEO who says the company has its finger in every AI-market pie and that its deals are not circular.

Top Story

OpenAI Makes Codex-Powered Agents Public Beta #

The Decoder AI • TechCrunch AI • OpenAI Blog

OpenAI is releasing its Agents API as a public beta. The managed service lets developers build cloud agents that run autonomously for hours, execute code, use tools and hand off tasks to sub-agents. OpenAI says developers pay no additional platform fee beyond token usage.

OpenAI has also released GPT-Live-1 as a developer API for applications that listen and speak simultaneously. The full-duplex speech model scored 80.1% on interactivity tests, up from 45.4% for its predecessor, and costs $0.05 per minute.

DeepSeek's V4.1-Flash offers a different efficiency tradeoff: its 552 billion-parameter model uses one-quarter the KV-cache memory of its predecessor while activating 16 billion parameters per token. It narrowly beat Opus 5 and GPT-5.6 Sol on the DeepSWE coding benchmark. Separately, TechCrunch reports that AI agents are flooding public services with new requests; one researcher said the vast majority involved people claiming something they were entitled to receive.

Why it matters: For developers, OpenAI now manages orchestration, long-running sessions and tool use through a Codex-powered service, while DeepSeek's V4.1-Flash cuts KV-cache memory to one-quarter of its predecessor.

Key Takeaways

  • GPT-Live-1 scored 80.1% on interactivity tests, compared with 45.4% for its predecessor, at $0.05 per minute
  • DeepSeek V4.1-Flash contains 552 billion parameters but activates 16 billion per token and narrowly beat Opus 5 and GPT-5.6 Sol on DeepSWE
  • TechCrunch reports AI agents are flooding public services with requests, and a researcher said most cases involved people claiming benefits or services they were entitled to receive

Industry Updates

Anthropic Alleges Escalating Distillation Attacks From Three China-Based AI Firms #

TechCrunch AI

Anthropic's new report alleges persistent distillation attacks by Alibaba, Moonshot AI, and DeepSeek. The report says the attacks have escalated in recent months as competition among AI companies has intensified.

Why it matters: The report puts Alibaba, Moonshot AI, and DeepSeek at the center of Anthropic's allegations as competition intensifies and alleged distillation attacks escalate.

Huang Forecasts 70% Nvidia Growth Next Year #

TechCrunch AI

Nvidia CEO Jensen Huang forecasts 70% growth for the company next year. TechCrunch AI says Huang sees “another year of plenty” ahead for Nvidia, which it describes as having its finger in every pie.

Huang also rejects the claim that Nvidia’s business deals are circular.

Why it matters: Nvidia investors are getting a bullish 70% growth forecast from a CEO who says the company has its finger in every AI-market pie and that its deals are not circular.