AI gets cheaper. The review still matters.

Good morning! Today’s AI news has a useful theme: the machinery is getting more capable, but the job still needs a clear finish line. A cheaper model can help with the inbox; it cannot decide what “done” means for your business.

Three Tuesday highlights:

Also inside: three tools for document workflows and a funding round for AI in wealth management. The paperclips have their own agenda.

OpenAI introduced GPT-6.1 Terra on October 26. It targets visual reasoning and computer tasks, from interpreting documents to navigating software. It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.1-terra.

Standard API pricing is $1 per million input tokens, $0.05 for cached input and $5 for output. OpenAI says those prices are 50% below GPT-5.6 Terra and GPT-6.1 Sol. The company also reports faster task completion and stronger results on computer-use evaluations. Those are vendor measurements, rather than a guarantee about your particular workflow.

The accompanying system card matters as much as the launch post. Its evaluations use research or API environments that can differ from production, including tools, instructions and reasoning settings. OpenAI treats Terra as Critical in cybersecurity and High in biological and chemical capability under its own preparedness framework. Launch details · System card.

Who benefits? Operations teams processing recurring paperwork, founders building document products and developers automating established desktop routines have a concrete reason to experiment. Start with a task whose result you can check without asking the model whether it did a good job.

For an invoice, “done” might mean every required field is present, the arithmetic reconciles and each value points back to the original page. For a sales proposal, it might mean the requested sections exist and every commercial claim has an approved source.

Choose a batch that includes an awkward scan and a missing field. Clean examples make good demonstrations; awkward ones reveal whether your workflow actually handles an ordinary Tuesday.

Before expanding access, decide who reviews exceptions and what the agent may change. A bounded pilot with a visible audit trail beats a grand launch followed by detective work.

On October 26, Anthropic filed an S-1 registration statement for a proposed public offering. It intends to list on Nasdaq as “ANT.” Share count and price range have not been determined, and the registration was not yet effective in its announcement. This is a filing, not a completed fundraising. Source: Anthropic.

An October 26 OpenAI case study describes CodeRabbit using GPT-6.1 Sol to investigate security issues across repositories and dependencies. Selected findings can be checked in an isolated sandbox, with proposed fixes sent for human review. This is evidence of an existing workflow, not a newly announced independent benchmark. Source: OpenAI.

OpenAI released data covering 165 work activities across nearly 100 countries on October 26. It spans early 2025 through the third quarter of 2026. The sample covers individual ChatGPT plans, excluding enterprise messages and Codex; usage patterns should not be mistaken for measured productivity gains. Source and data: OpenAI.

Parses PDFs and other files into structured formats, including Markdown and JSON, with support for layout, tables and OCR. Useful for engineers building document pipelines. The code is free under MIT; hosting costs and individual model licenses are separate. Official project.

Combines document extraction with routing, validation and integrations. Best for operations teams seeking a managed workflow. Start with $50 in free credits; the published Starter offer then lists $100 monthly for 100 credits. Growth and Enterprise require quotes. Check block consumption against your document mix. Current pricing.

A configurable interface for labeling data and checking model predictions. Helpful for teams building a reference dataset or correction queue. Community Edition is free and open source; Starter Cloud and Enterprise are paid. Advanced reviewer assignment belongs to paid editions. Edition comparison.

Use a small batch of approved sample invoices. The first version can live in a spreadsheet; your goal is a reliable handoff, not an elaborate platform.

1. Define the record. List supplier, invoice number, date, currency, subtotal, tax and total. Add source filename, page and review status. Agree on the format before processing anything.

2. Extract with evidence. Ask your approved AI tool for one row per invoice and a supporting quote for every field. Missing or unclear values should remain blank and receive a reason.

3. Apply simple checks. Flag duplicate invoice numbers, absent currencies and totals that fail arithmetic. Keep the original documents beside the extracted records so a reviewer can inspect the evidence.

4. Review and measure. Have a person approve every row in the pilot. Record correction time and recurring failure types. Expand only when the accepted records save time after review.

Give each reviewer a short, written correction guide: whether to preserve date formats, how to handle credit notes and where to record conflicting totals. Consistent review rules make the initial pilot useful even when the model changes later.

Keep payment execution outside this exercise. Extraction and approval are separate jobs, even when the assistant offers to be helpful.

Money raised: $70 million, Series B. Valuation: Undisclosed in the company announcement.

Announced October 26, the round was led by EQT Ventures, with Charles Schwab and existing investors participating. Nevis says total funding is now $110 million. Its platform combines customer records, meeting intelligence, financial planning and operational workflows for wealth managers. The new capital will support agent development and expansion across that software stack. Company announcement · Dated reporting.

Our read: A narrow industry with repetitive paperwork offers a clearer starting point than a promise to automate every department. Integration depth and review discipline will decide whether the platform earns its place.

Today’s reminder: efficiency needs a target. Before an agent starts, specify the deliverable, the stopping point and the one thing it absolutely should leave alone.

Which document task steals your morning? Reply with one example, and forward today’s briefing to the colleague who owns the review queue.