Skip to content
RET Consulting Contact me
Blog

AI Digest

What shipped in AI, most mornings, in the time it takes to read a message. Each brief answers the same three questions: what happened, why it holds up, and what it changes for a product team.

The rules of a brief

  • Under 80 words
  • One named source
  • Reviewed by hand
  • Nothing on quiet days

Briefs

  1. Open source

    MiniMax opens H3 video weights, but not to the EU

    Weights for MiniMax H3, a 33B omni-modal video and audio model, went up on Hugging Face on 3 August under the MiniMax H3 Community License. Artificial Analysis ranks it first for video editing, making it the strongest open-weight video model published so far. The licence excludes the EU, UK, US and South Korea from its applicable territory, so a European team cannot self-host it without separate authorisation, and the hosted API stays the only route.

    Source: MiniMax on Hugging Face

  2. Regulation

    US frontier model review framework goes to the labs

    The White House met OpenAI, Google and Anthropic in early August to brief them on a finished voluntary framework, under which agencies get up to 30 days of early access to a frontier model before it reaches other partners. It follows the June executive order and stops short of licensing. Model availability dates in a delivery plan now carry a policy dependency.

    Source: CNBC

  3. Open source

    DeepSeek ships a new V4 Flash under an MIT licence

    The V4-Flash-0731 checkpoint landed on Hugging Face with MIT-licensed open weights, re-post-trained at the same architecture and size as the preview build, and ahead of the earlier V4 Pro preview on several agent benchmarks. For teams that need weights they can host themselves, the open tier keeps closing on the hosted frontier.

    Source: The New Stack

  4. Business

    OpenAI cuts GPT-5.6 Luna by 80 per cent

    On 30 July OpenAI reduced the price of GPT-5.6 Luna by 80 per cent and Terra by 20 per cent, weeks after the Sol, Terra and Luna family shipped. The cheap tier is now cheap enough that routing high-volume, low-judgement steps away from the flagship changes a project's economics. Any cost model written before that date is worth re-running.

    Source: OpenAI

  5. Models

    Gemini 3.6 Flash cuts output tokens by 17 per cent

    Google introduced Gemini 3.6 Flash alongside 3.5 Flash-Lite and a security-tuned 3.5 Flash Cyber. The workhorse model reports 17 per cent fewer output tokens than 3.5 Flash, up to 65 per cent on some coding benchmarks, at a lower cost per output token. Shorter answers save twice over: fewer tokens billed, and less to read before the work is usable.

    Source: Google

  6. Products

    Copilot gains Word, Excel and PowerPoint agents

    The July release of Microsoft 365 Copilot lets an @mention pull the Word, Excel or PowerPoint agent into Copilot Chat, producing a document, a spreadsheet or a deck without leaving the chat. GPT-5.6 and Claude Sonnet 5 also became selectable, with Claude Fable 5 in Cowork. Picking a model per task is now an everyday choice for office users, not an admin setting.

    Source: Microsoft 365 Copilot blog

  7. Models

    Claude Opus 5 lands at the price of Opus 4.8

    Anthropic shipped Opus 5 on 24 July, close to Fable 5 on coding and knowledge-work evaluations at half the price, with input and output pricing unchanged from Opus 4.8. A Fast mode runs about 2.5 times the default speed at twice the base price. For a team already budgeted on Opus, that is a capability step with no new line in the contract.

    Source: Anthropic

Briefs are drafted with an AI research workflow, then reviewed by hand. Nothing goes live without that review, and no brief goes out on a day with nothing worth reporting.

Written by Redha Etbaz · Independent product owner and business analyst Back to Field notes