Claude adds Projects, Bonsai shrinks AI: Sept. 18

Claude can now split a project across several AI workers, while a smaller Bonsai model makes private local processing more practical and Google opens verified UN figures to automated research. In 30 minutes, you can put approval limits around parallel AI work, choose one private task for a local test, and tighten how reports collect evidence.

Claude Projects coordinates several AI workers

Anthropic released a redesigned Projects feature in beta on September 17. The official announcement says one Claude Project can create several threads, meaning separate work sessions, while a coordinator assigns jobs and combines the results. It is starting with selected Pro and Max subscribers who use Claude Code cloud sessions and have no existing web or desktop projects.

The business gain is less waiting on one long job. You could give one thread a sales-page draft, another the supporting research, and a third the quality check. Shared memory keeps the brief and earlier decisions available across those threads. The cost risk rises too: Anthropic says every thread is a full Claude Code session, so a project can reach its usage limit faster.

Do not hand it three versions of the same page. Overlapping work can create a merge conflict, which means two sessions changed the same material and a person must choose what survives. Use the control pattern from the AI content approval workflow, and compare the wider tool tradeoffs in ZCode versus Claude Code.

Bonsai 2 puts a 27-billion-parameter model in 5.9 GB

PrismML released Ternary Bonsai 2 27B on September 17 with a 5.9 GB footprint. A parameter is one of the internal values an AI model uses to produce an answer; 27 billion usually demands much more memory. The release notes say the model accepts text and images, has an Apache 2.0 license, and runs on NVIDIA graphics cards or Apple devices through custom software.

This matters when documents should not be sent to an outside AI service, or when repeated cloud charges make a high-volume task expensive. PrismML reports that its compressed model retained 98.2% of the full version's combined benchmark score. Treat that as a maker-reported test, not proof that it will understand your invoices, product sheets, or brand rules correctly.

Pick one private, reversible job such as classifying old documents. Install nothing on a staff computer yet. Ask whoever manages your technology to run 25 known examples locally, record wrong answers, and compare the same files with your current approved process. The earlier Bonsai dispatch item explains the local-computing case; this release earns a new test, not an automatic switch.

Google opens verified UN data to AI research

Google and the United Nations launched UN System Data Commons on September 17. The Google announcement describes a searchable collection that joins UN statistics across subjects and locations. It also supports Model Context Protocol, or MCP, a standard connection that lets an AI assistant retrieve information from another system.

For a business owner, the useful change is traceable evidence. A research assistant can fetch an official population, health, education, or infrastructure figure without copying rows among several spreadsheets. That can strengthen market briefs and location reports, but it does not make every number suitable for a sales claim. Geography, measurement date, and the UN agency behind the figure still need a human check.

Start manually at the UN data site. Ask one plain-language question, open the underlying dataset, and record the source agency and date beside the answer. Only after five correct checks should you connect it to an automated report. Keep the final draft behind the review step in the weekly marketing report build, or use the guarded pattern in the n8n AI agent setup.

On the bench

• Qwen3.8-Omni-Flash: its official page is dated September 14, outside this dispatch window. Test the audio and video claims only after checking the exact API price for your region.
• Claude Projects on local computers: Anthropic says local tools and code are coming, but gives no release date. Keep cloud-only material non-sensitive for now.
• UN dataset coverage: Google says the project aims to include 80% of UN system statistical datasets by 2027. Recheck coverage before depending on a narrow industry measure.

Get the next build from the AI tools runbooks.