All Issues
Jul 27 - Aug 23, 2026

AI Weekly: Astra Solves Unsolved Maths, EU AI Act Kicks In

Models & Releases

4 stories

Meta Returns to Open Weights with Muse Glimmer

  • Meta released Muse Glimmer on August 10: a 30-billion-parameter agentic model under Apache 2.0 licence, small enough to run on a single consumer GPU with 24 GB VRAM.
  • It uses DFlash speculative decoding for 3.1× faster inference and is distilled from Muse Spark, making it the first open-weight model in Meta’s Muse line and a reversal of the closed-API strategy seen with Muse Spark 1.1 (Jul 12 edition).
  • Zuckerberg used the launch as a policy pitch for open-source AI leadership, and Meta simultaneously announced plans to release Muse Spark 1.2 weights — signalling a broader return to the open-weights strategy.

Gemini 3.7 Flash: Coding Workhorse at Half the Price

  • Google released Gemini 3.7 Flash on August 13 — just three weeks after 3.6 Flash — calling it its most intelligent workhorse model yet for coding, agentic workflows, and document processing.
  • API prices are halved through the end of 2026, and the model now powers Gemini Spark consumer agent, enabling it to consolidate files, draft emails, and update status documents autonomously.
  • DeepSeek also shipped V4-Pro GA (August 12–13) in the same window, with vendor-reported benchmark gains of up to 49.9 percentage points — though no independent replication has confirmed those figures yet.

GPT-5.6 Luna Drops 80% — and DeepSeek V4-Flash Goes Open

  • OpenAI cut GPT-5.6 Luna prices by 80% on July 30 (to $0.20/$1.20 per million tokens) and GPT-5.6 Terra by 20%, responding to competitive pressure from cheaper frontier alternatives just three weeks after the models launched.
  • The same week, DeepSeek released open weights for V4-Flash (July 31), continuing its hybrid open/closed strategy before the flagship V4-Pro reached GA on August 12–13.
  • OpenAI also retired o3 from ChatGPT on August 26 after its 90-day sunset, completing the model generation turnover to the GPT-5.6 lineup.

People & Business

3 stories

Anthropic Breaks Silence on Open Weights, Frames China Risk

  • On July 27, Anthropic CEO Dario Amodei published a formal position paper clarifying that the company has never advocated a ban on open-weight models — breaking a silence that had left Anthropic as the last major frontier lab not to sign the ‘Open Weights and American AI Leadership’ coalition letter.
  • Rather than banning open weights, Amodei called for mandatory pre-release safety testing, chip export controls, and distillation restrictions, citing the 2026 Intelligence Community Threat Assessment on China’s military AI progress.
  • The position generated significant pushback: critics noted Anthropic made a structurally similar argument itself six weeks earlier, and signing the letter while adding conditions would have been a more direct path — an episode that reveals the fracture lines in the industry’s open-source consensus.

ChatGPT Work Gets Admin Plugin for Enterprise Management

  • OpenAI launched an Admin plugin for ChatGPT Work on August 25, letting IT admins manage workspace activity, access controls, and usage data through a natural-language conversation interface.
  • The plugin integrates with existing enterprise identity systems and lets admins take supported administrative actions — such as adjusting permissions or pulling usage reports — without leaving the chat interface.
  • The launch accompanies the final retirement of o3 from ChatGPT (August 26) and the Assistants API, marking the completion of OpenAI’s generational model turnover in its enterprise surface.

Policy & Ethics

3 stories

White House Finalises Voluntary Frontier AI Safety Tests

  • The Trump administration finalised a voluntary cybersecurity testing framework for frontier AI models on August 3, offering the government up to 30 days of pre-release access to covered models — explicitly not a licensing regime.
  • OpenAI, Anthropic, Google, and Meta were invited to a White House meeting on August 4 to discuss the framework, which focuses on measuring offensive hacking capabilities before public deployment.
  • The framework lands directly alongside OpenAI’s own August 18 pacing paper, which independently proposed mandatory external evaluation at cyber-critical capability thresholds — suggesting convergence between lab self-governance and government frameworks.

Anthropic Adds Invisible Text Watermarks to All Claude Models

  • Anthropic announced on August 11 that it is embedding invisible, machine-readable watermarks into text generated by all new Claude models worldwide — not just in the EU — triggered by the EU AI Act’s new transparency rules.
  • The watermarks survive copy-and-paste and appear even when Claude has only lightly edited a human text; a companion C2PA signed-metadata system applies to generated files and images.
  • Anthropic confirmed that translations produced by Claude carry a watermark (every word is chosen by the model), and that models launched before August 2 are under a transition period to add the capability — raising questions about detection limits and creative-industry liability.

Products & Hardware

1 stories

Research & Resources

1 stories