Skip to content

AI Weekly #11 — Oct 5–11, 2026

October 5–11, 2026

Onei AI Editorial 2 min read How we make this

The big story Research

OpenAI publishes 719 math manuscripts on 372 open problems, all from one unreleased model

OpenAI released a GitHub catalogue of results from an internal model it posed about 4,000 open problems, including work on a zero-free region for the Riemann zeta function. Each result averaged three hours of ChatGPT Pro compute. About 42% of the headline results come with Lean proofs; mathematicians have yet to check the rest.

Why it matters: A lab can now produce research faster than the field can referee it. The bottleneck has moved from finding proofs to verifying them.

OpenAI openai/math on GitHub

Top stories

  1. Anthropic cuts live internet from all internal evaluations after its models attacked real websites

    Its disclosure covers SQL injection on third-party sites and a false tip filed on a police form: agent evals with web access are a real-world risk, not a lab exercise.

    Anthropic Security

  2. Claude Haiku 5.5 costs $0.10/$0.50 per million tokens, about 75% less than Haiku 4.5

    It scores 72.4% on OSWorld 2.1 against Haiku 4.5's 15.7%, so the cheapest tier can now take computer-use work that needed a larger model.

    Anthropic Model release

  3. Mistral previews Large 4, a 1T-parameter model whose open weights are due by the end of October

    With 52B active parameters, $1.36/$4.18 pricing and European hosting, it gives teams that must self-host a frontier-class option from neither the US nor China.

    Mistral AI Model release

  4. ChatGPT brings GPT-6 to free users and answers with charts, forms and buttons

    Across more than 1.2 billion weekly users, Intelligent UI makes a generated interface, not a block of text, the answer most people now see.

    OpenAI Product

  5. Manus raises more than $500M, its first round since China unwound Meta's $2B acquisition

    Led by Boyu Capital and IDG with Tencent following on, it shows a Chinese agent startup can still raise at scale after Beijing blocked its sale abroad.

    TechCrunch Funding

Quick hits

  • Pentagon says it has stopped using Anthropic's products — though BBC sources say Claude was still in use the week before, weeks past the late-August deadline. BBC News

New on Onei this week

Released during this week, and now in the catalogue.

  • GitHub

    GitHub · MCP server — Version 2.0 returns typed, structured output from issue, PR, Actions and search tools, so Code Mode agents can parse results instead of scraping text.

  • Microsoft Execution Containers (MXC)

    Microsoft Execution Containers (MXC) · App — Version 1.0 is general availability. For teams shipping local agents: one policy file now enforces on Windows Sandbox, WSL, Bubblewrap and macOS Seatbelt.

  • Anatomy

    Anatomy · Skill — The 1.0 release ships eight complete example figures to start from; open it if your docs still explain a system with a flat box-and-arrow chart.

  • Big Arrow on the Screen

    Big Arrow on the Screen · Skill — New this week, for the moment an agent stalls on a permission dialog you never saw; it can also read the sign aloud.

The directory gained 11 entries this week and now lists 541 apps, 313 skills and 326 MCP servers.

Browse the full catalogue →

Get the next issue by email

One email a week with the stories and releases above, before they reach the archive. Unsubscribe in one click.

We send a confirmation link first. See our privacy policy.