Skip to content
Source · Daily Brief

AI Daily Brief — 17 June 2025

A Google-and-OpenAI Tuesday with a Chinese open-source kicker. Google flipped the switch on Gemini 2.5 Pro and Flash GA and shipped Flash-Lite preview at the lowest paid-tier price it has ever charged for a frontier model. OpenAI booked a $200M Pentagon contract and consolidated its federal work under a new “OpenAI for Government” umbrella — awkwardly routing around Microsoft Azure. From Shanghai, MiniMax open-sourced M1, a 456B-parameter MoE reasoner with a 1M-token context window trained on $534,700 of RL compute.

Top stories

  • Google Gemini 2.5 Pro and Flash hit GA. Both stable endpoints launched on Vertex AI, Gemini API and AI Studio, with the older preview endpoints sunsetting July 15. via Google Cloud Blog
  • Gemini 2.5 Flash-Lite preview at $0.10/$0.40. The cheapest paid frontier model Google has ever shipped, thinking off by default, multimodal input and Search grounding included. Independent benchmarks later clocked it at 887 output tokens/sec — fastest proprietary model measured. via Google Developers Blog
  • OpenAI lands $200M Pentagon CDAO contract. One-year prototype ceiling award from the Chief Digital and AI Office covers warfighting and enterprise-admin frontier-AI work, signed Jun 16 and announced Jun 17. via TechCrunch
  • “OpenAI for Government” initiative launches. Folds existing US National Labs, AFRL, NASA, NIH and Treasury collaborations under one banner anchored by the new CDAO award. via OpenAI
  • MiniMax open-sources M1. Shanghai’s MiniMax released a 456B-parameter MoE (45.9B active) with 1M-token context, 80K reasoning output and Apache 2.0 weights — trained on $534,700 of RL compute. The technical report claims roughly 30% of the FLOPs of DeepSeek R1 at 80K generations. via MiniMax
  • MiniMax-M1 paper drops alongside weights. arxiv 2506.13585 details the Lightning-Attention hybrid design and SWE-Bench Verified scores of 55.6% / 56.0% for the 40K and 80K variants. via arxiv

Who shipped

Google shipped GA versions of Gemini 2.5 Pro and Flash plus the Flash-Lite preview. OpenAI launched the Government initiative and the $200M CDAO contract. MiniMax open-sourced its 456B-parameter M1 reasoner with paper attached.

Open-source pulse

MiniMax-M1 reshapes the open-weight reasoning frontier — a 1M-context hybrid-attention MoE under Apache 2.0 that closes the gap with DeepSeek R1 on SWE-Bench Verified while training on less than $535K of RL compute.

Money, infra & hardware

The Pentagon CDAO booking marks one of the largest single frontier-AI defense contracts to date, structured as a one-year, $200M-ceiling prototype award. The “OpenAI for Government” framing implicitly sidesteps Microsoft Azure as the Pentagon’s incumbent OpenAI reseller.

Quiet corners

Pope Leo XIV released his message to the Second Annual Rome Conference on AI, Ethics & Corporate Governance (Jun 19-20), calling AI “an exceptional product of human genius” but “above all else a tool” — previewing his first encyclical. via Vatican.va

By the numbers

  • $200M — Pentagon CDAO contract ceiling for OpenAI
  • $0.10 / $0.40 — Flash-Lite per million input/output tokens, a frontier-tier record low
  • 887 tok/sec — independent throughput benchmark for Flash-Lite
  • 456B / 45.9B / 1M / 80K — MiniMax-M1 total params / active / context window / reasoning output
  • $534,700 — total RL training compute spend for M1

Compiled by AI Feed’s editor from verified web sources for 17 June 2025.