Skip to content
Source · Daily Brief

AI Daily Brief — 16 April 2025

OpenAI’s most coordinated capability push of 2025 — o3 and o4-mini ship with agentic tool use across every ChatGPT tool, Codex CLI lands open-source, and the Preparedness Framework gets a major rewrite with a ‘competitive dynamics’ carve-out. xAI counter-programmed with Grok memory and Grok Studio. NVIDIA spent the day digesting the H20 ban.

Top stories

  • OpenAI launches o3 and o4-mini reasoning models. First reasoning models that can agentically use every ChatGPT tool — web browsing, Python, file analysis, image understanding and image generation — inside the chain of thought. Available to Pro, Plus and Team subscribers; o4-mini-high variant added. Replaces o1, o3-mini and o3-mini-high in the model selector. via OpenAI
  • Benchmark headlines: AIME, SWE-Bench, Codeforces. o4-mini scores 92.7% on AIME 2025 (no tools), beating o3 88.9%, o3-mini 86.5% and o1 79.2%; with Python, o4-mini hits 99.5% on AIME. SWE-Bench Verified: o3 69.1%, o4-mini 68.1%. Codeforces: o3 ~2727 Elo (top ~200 humans worldwide).
  • o3 ‘thinks with images.’ First OpenAI models that can manipulate images (zoom, crop, rotate) inside their reasoning steps. Users upload whiteboards, sketches or blurry photos and the model uses image-editing tools as part of its thought process. via CNBC
  • OpenAI open-sources Codex CLI. Lightweight terminal coding agent written in Rust. Installs via npm (npm i -g @openai/codex), Homebrew or direct binary; auth via ChatGPT plan or API key. Supports multimodal input. $1M API-credits grant program ($25K blocks) to seed adoption. via TechCrunch
  • o3 API pricing set at $10 / $40 per million tokens. 200K context window. o4-mini priced significantly lower. Flex Mode and o3-pro tiers signaled. via TechCrunch
  • OpenAI rewrites Preparedness Framework (v2). Streamlines risk thresholds to ‘High capability’ and ‘Critical capability’; drops ‘persuasion / mass-manipulation’ as a pre-release evaluation category, moving it to terms-of-service enforcement. Adds new research categories for capability concealment, safeguard evasion and self-replication. Introduces a ‘competitive dynamics’ clause: OpenAI may ‘adjust’ safeguards if a rival lab releases a comparable high-risk model. via OpenAI
  • xAI ships Grok memory feature. Grok now remembers details from past conversations and carries context across sessions. ‘Memories are transparent’ — users can see and delete entries. Beta on Grok.com and iOS/Android; not in UK or EU at launch. Catches up to ChatGPT and Gemini. via TechCrunch
  • xAI launches Grok Studio — Canvas-style collaborative workspace. Split-screen workspace for docs, code and even browser games. Supports Python, C++, JavaScript and TypeScript execution, plus Google Drive integration (Docs/Sheets/Slides). Free and premium tiers. Directly competes with ChatGPT Canvas and Claude Artifacts. via Maginative

Who shipped

OpenAI ran the day end-to-end: o3, o4-mini, Codex CLI, Preparedness v2 — its most coordinated launch since the 12-Days-of-OpenAI event. xAI ran a respectable counter-launch with memory and Studio. NVIDIA said it follows export laws ‘to the letter’ after the H20 ban took effect. AMD formally disclosed its $800M MI308 charge.

Open-source pulse

Codex CLI shipped open-source under a permissive license — OpenAI’s most aggressive open-source move since the Whisper release.

Money, infra & hardware

NVIDIA’s $5.5B charge and AMD’s $800M parallel disclosure together represent the biggest single AI-export-control hit on record.

Quiet corners

Anthropic, Google, Meta and Chinese labs all silent on the model side. The chip stocks continued to absorb the H20/MI308 disclosure.

By the numbers

  • 92.7% / 99.5% — o4-mini AIME 2025 without / with tools
  • 69.1% / 68.1% — o3 / o4-mini SWE-Bench Verified
  • ~2727 — o3 Codeforces Elo (top ~200 humans)
  • $10 / $40 per million tokens — o3 input/output
  • 200K — o3 context window
  • $1M / $25K — Codex CLI grant program total / block size
  • Most-mentioned company: OpenAI

Compiled by AI Feed’s editor from verified web sources for 16 April 2025.