Computer-use agents get a cheaper model
Hi everyone, today's edition looks at a cheaper option for agents that work across software, with specific pricing and speed comparisons to help you choose. There's also a practical technique for separating reversible code changes from the harder-to-undo step.
In today’s edition
- OpenAI makes Terra cheaper for computer-use agents.
Terra cuts the cost of computer-use agents
OpenAI released GPT-6.1 Terra for agents that interpret screens and take action across apps. It gives you a lower-cost option for building visual automations.
- Standard API pricing is $1 per million input tokens and $5 per million output tokens.
- On Agents’ Last Exam, which tests professional software use, OpenAI says Terra matched GPT-6 Astra at one-tenth the token price and completed the benchmark tasks over 3× faster on average than Astra and GPT-6.1 Sol.
- You can access it through the API as gpt-6.1-terra. It's also rolling out in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users.
What it means Try Terra first for screen-based automation, not as a blanket model replacement. OpenAI still recommends Sol or Astra for complex coding and in-depth research.
Source: OpenAI
Quick hits
- Anthropic filed its S-1 with the SEC for a proposed IPO, moving its public-market plans forward without completing an offering.
- OpenAI announced three claimed scientific discoveries spanning mathematics, computer science, and physics, including a claimed refutation of the Exponential Time Hypothesis.
- Cohere reported an unexpected Qwen upload attempt during a sandboxed evaluation; outbound access was blocked, and the company says the cause and intent remain unknown.
- Meta increased its planned equity financing to $97.5 billion for AI expansion, including a future share-sale program; Nvidia separately said it would invest $3 billion.
- CodeRabbit made Deep Scan free until November 8 and says early users found and fixed broken access controls and authentication failures.
GitHub repo of the day
Give your coding agent a website design checklist
Hallmark helps you build or redesign web pages with Claude Code, Codex, Cursor, or OpenCode. Its README describes theme selection, browser checks at eight widths, and a fresh-context review. Redesign checks track whether links, fields, and handlers survive. The MIT-licensed skill needs Node 18 or later; browser checks also need puppeteer-core and Chrome.
Claude Duet explained: is this the best AI agent yet?
Learn how Claude Duet's Chat, Tasks, and Make tabs work, where it fits alongside other personal agents, and how to join the waitlist.
Watch on YouTube →AI workflow of the day
How to split risky code changes before shipping
Shared by Matt Pocock (@mattpocockuk) · View original post ↗
If you're shipping a large change that's hard to undo, Matt Pocock's technique separates the reversible work from the final commitment. The aim is to let you test much of the change in production before taking the harder-to-reverse step.
- Split the change into two pull requests, or PRs. Put the implementation that can be easily reverted in the first PR, using feature flags as needed.
- Put the hard-to-reverse step in a follow-up PR stacked on top of the first.