HAI Daily

Decision Models Arrive, Anthropic's Free OSS Scanner and China's Plan for New Quality Productive Forces

Microsoft launches Decision-1, a fast scoring model built on Qwen, TypeSafe raises $870 million for its non-text model Jev, Anthropic offers free AI vulnerability scans to open-source maintainers, and China sets out how to develop new quality productive forces.

By H.AI Tools Team3 stories
Products in this issue: Claude, Snyk, GitHub Copilot, Qwen Chat, ChatGPT
In this issue

The short version

  • Microsoft-Decision-1 and TypeSafe's Jev return structured decisions instead of text. Decision-1 costs $0.042 per million input tokens, with output free.
  • Anthropic's OSS Scanner offers free AI vulnerability scans to maintainers of widely used open-source projects. Penetration testers confirmed 85 of 97 serious findings in a sample.
  • China's guidelines on new quality productive forces put science and technology first and name robotics, new energy and other emerging industries.
  • Use decision models for high-volume routing and classification, and keep a general model for anything that needs explanation.

October 10 introduced a new kind of AI model: one that does not write text at all, but picks an answer and returns it to software. Anthropic made AI security scanning free for key open-source projects, and China published broad guidelines on its next phase of economic development. Each section explains what was announced and how to use it.

01

Decision models: Microsoft-Decision-1 and TypeSafe's Jev

What was launched

Microsoft released Microsoft-Decision-1 on October 9, a model post-trained from Alibaba's (阿里巴巴) open Qwen3.5-9B for structured decisions. Instead of generating prose, it scores a fixed set of options in a single pass and returns the result in a structured format. Microsoft reports the highest average accuracy across 36 benchmarks and says it is about 35 times faster than GPT-6 Sol. It costs $0.042 per million input tokens, with output free, through Microsoft Foundry and OpenRouter; the weights are not released. Microsoft plans to adapt the approach to other base models. These figures are Microsoft's own, and its documentation advises customers to check calibration on their own data.

The same day, TypeSafe AI raised about $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with DCVC and reportedly Sequoia. Its model Jev, launched in September, also returns typed decisions for software to act on rather than text. TypeSafe says a third of Fortune 500 companies use it, but it has not named customers, and that figure has been questioned.

Where decision models fit

Many steps in software need a choice, not an essay: is this email a complaint, which team should handle this ticket, does this transaction look risky, is this request allowed. Using a general chatbot for these is slow and expensive, and its answers vary in format. A decision model returns one of your defined options quickly and cheaply.

To try one, pick a classification step you already run with a general model or by hand. Write the options clearly, collect 200 past examples with known answers, and measure accuracy, speed and cost against your current method. Pay attention to the cases it gets wrong and add a "needs human review" option for low-confidence results. Keep a general assistant such as ChatGPT or Qwen (通义千问) for steps that need reasoning or an explanation.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

Anthropic's free OSS Scanner

What Anthropic offers

Anthropic opened OSS Scanner on October 8, a free service that uses its models to scan open-source projects for vulnerabilities on a regular schedule. Core maintainers of widely used, security-critical projects can enrol by submitting a pull request to Anthropic's repository, under criteria similar to Google's OSS-Fuzz. Anthropic says its models had found over 29,000 candidate vulnerabilities in open-source software over the previous six months, and about 6,000 had been reviewed and graded by people. In a check by experienced penetration testers, 85 of 97 high- and critical-severity findings across 48 projects met the standard for coordinated disclosure. The reports are generated entirely by AI, and Anthropic says it will keep improving the system.

Coverage notes that the bottleneck is now fixing, not finding: far fewer issues have been patched than reported, and some maintainers say severity ratings were inflated. Anthropic is also funding open-source security organisations to help.

What maintainers and users should do

If you maintain a widely used project, enrolling gives you findings you would struggle to get otherwise. Prepare first: set up a private way to receive security reports, decide who triages them and agree a target time to fix critical issues. Treat each AI finding as a lead to confirm, not a verdict.

If you depend on open-source software, which nearly every team does, keep your dependencies up to date and scan them with a tool such as Snyk. Coding assistants such as GitHub Copilot or Claude can help apply upgrades, but run your tests before merging. The AI Security category lists more tools.

03

China's guidelines on new quality productive forces

What the guidelines say

The Communist Party Central Committee and the State Council published guidelines on October 9 on developing "new quality productive forces", the government's term for growth driven by innovation rather than investment in traditional capacity. They make a large rise in total factor productivity the core goal and call for innovation in science and technology, industry, ways of developing, institutions and talent. Tasks include strengthening original and disruptive research, upgrading traditional industries, growing strategic emerging industries such as next-generation IT, new energy, new materials, robotics, biomedicine and aerospace, and developing the low-altitude economy and commercial space on the condition of safety.

What it means for businesses

These guidelines set the direction for the more specific plans that followed in October, including the industrial foundations plan and the guidelines on fusing science with industry, covered in our October 22 issue and October 23 issue. For companies working in or with China, the message is consistent: support goes to measurable productivity gains from technology. Frame AI projects around productivity metrics such as output per hour, defect rates or energy per unit, and record a baseline before you start.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

What to try next

Put one change into practice

Move one high-volume classification step to a decision model and measure accuracy on 200 past cases. If you maintain an important open-source project, set up a private security reporting channel before enrolling in a scanner. Record a productivity baseline before starting any AI project.

Explore Chatbots, AI Security and Code Generation to build your shortlist. Our Claude review covers Anthropic's assistant.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper