Skip to content
Source · Daily Brief

AI Daily Brief — 22 December 2025

Monday packed three structural shifts. Z.ai (Zhipu) open-sourced GLM-4.7 — coding-focused ~400B-param model with 200K context + 128K max output. SWE-bench Verified 73.8%; matches or beats Claude Sonnet 4.5 on LiveCodeBench v6 + Terminal Bench 2.0. Cements “China’s OpenAI” positioning. Anthropic released Bloom — open-source agentic framework for automated behavioral evaluations at github.com/safety-research/bloom. Takes researcher-specified behavior (sycophancy + sabotage + bias) and auto-generates many scenarios to measure frequency/severity in days rather than weeks. Complements earlier Petri tool; benchmark results released for 16 models across 4 alignment-relevant behaviors. OpenAI launched “Your Year with ChatGPT” Spotify-Wrapped-style recap to Free/Plus/Pro in US + Canada + UK + Australia + New Zealand. Generates custom awards + poem + pixel-art image from chat history. Requires “reference saved memories” + “reference chat history” toggled on + minimum-activity threshold. Pulitzer winner John Carreyrou + 5 authors filed suit against 6 AI giants in N.D. Cal. (case 25-cv-10897) accusing Anthropic + OpenAI + Google + Meta + xAI + Perplexity of “willful theft” for downloading their books from shadow libraries LibGen + Z-Library + OceanofPDF to train LLMs. First copyright suit against xAI + Perplexity by authors. Seeks $150K statutory damages per work per defendant ($900K per work total). Plaintiffs opted out of proposed $1.5B Anthropic settlement. xAI selected by US Department of War to deliver Frontier AI for government workloads. Same day, xAI made Grok Voice available to all developers + launched Collections API for document indexing + retrieval across large internal datasets (legal/financial archives). IBM Research CUGA on HF Spaces (Apache 2.0 enterprise multi-step framework) continues uptake. RedMonk: “10 Things Developers Want from Agentic IDEs in 2025” — long-horizon planning + granular permissions + transparent file diffs + persistent project memory + model-agnostic tool calling.

Top stories

  • Z.ai opens GLM-4.7 (~400B, 200K context). SWE-bench Verified 73.8%; matches/beats Sonnet 4.5 on LiveCodeBench v6 + Terminal Bench 2.0. via PR Newswire
  • Anthropic releases Bloom open-source agentic behavioral eval framework. Auto-generates scenarios; 16 models × 4 alignment behaviors benchmarked. github.com/safety-research/bloom. via Anthropic
  • OpenAI launches “Your Year with ChatGPT” recap. US/CA/UK/AU/NZ Free/Plus/Pro. Custom awards + poem + pixel-art image. Requires memories + chat-history toggles + min activity. via TechCrunch
  • Carreyrou + 5 authors sue 6 AI giants in N.D. Cal. First copyright suit against xAI + Perplexity by authors. $150K × 6 = $900K per work. Opted out of $1.5B Anthropic settlement. via Bloomberg Law
  • xAI selected by US Department of War for Frontier AI. Same day: Grok Voice opens to all devs + Collections API for document indexing/retrieval over legal/financial archives. via xAI
  • IBM Research CUGA on HF Spaces. Apache 2.0 enterprise multi-step workflows web + APIs. OpenAPI + MCP + LangChain. via InfoQ
  • RedMonk: “10 Things Devs Want from Agentic IDEs in 2025”. Long-horizon planning + granular permissions + transparent diffs + persistent project memory + model-agnostic tool calling. via RedMonk
  • Radical Data Science December 2025 AI News Briefs Bulletin Board. Monthly aggregation of December research + product + funding + benchmarks. via Radical DS

Who shipped

Z.ai shipped GLM-4.7. Anthropic shipped Bloom. OpenAI shipped Year-with-ChatGPT recap. xAI shipped Grok Voice + Collections API + DoW deal. IBM‘s CUGA continues uptake. Google, Meta, DeepSeek, Alibaba quiet.

Open-source pulse

GLM-4.7 + Bloom + CUGA = three open-source signals in one day. Z.ai narrowing gap to Sonnet 4.5 on coding benchmarks restructures open-source coding ceiling.

Money, infra & hardware

xAI’s Department of War selection signals frontier-lab military AI distribution maturation; Grok Voice + Collections API targets enterprise document workflows.

Quiet corners

Carreyrou suit notable as first by individually-named Pulitzer winner against full frontier-lab cohort + first targeting xAI + Perplexity in author copyright wave.

By the numbers

  • ~400B / 200K / 73.8% — GLM-4.7 params / context / SWE-bench Verified
  • 16 / 4 — Bloom models benchmarked / alignment behaviors
  • 6 / $150K / $900K — defendants in Carreyrou suit / per work per defendant / total per work
  • 5 / Free/Plus/Pro — Year-with-ChatGPT launch countries / tiers
  • Most-mentioned company: Z.ai
  • Quietest segment: Google + Meta launches

Compiled by AI Feed’s editor from verified web sources for 22 December 2025.