GPT-6.1 Terra Cuts Your Agent Bill in Half

Today I've got OpenAI's half-price model for agents that click and type, GitHub's new way to stop picking models by hand, the end of the 5-hour wall on Claude's top plan, one robots.txt group that...

Happy Tuesday,

Today I've got OpenAI's half-price model for agents that click and type, GitHub's new way to stop picking models by hand, the end of the 5-hour wall on Claude's top plan, one robots.txt group that tells four AI companies not to train on your site, and a giant open-weight model you can rent for under a dollar.

Let's get into it.

In This Email

» The Big Thing: GPT-6.1 Terra cuts the price of computer-use agents in half

» Worth Your Time: Copilot Fusion, Claude Ultra, a robots.txt line for AI training, and Mistral Large 4

» Quick Hits: agents inside GitHub Actions, and free file editing in Gemini

» Skip It: Gemini 4 Carbon

» Here's My Take: rerun your real tasks before you trust any chart

🚨 The Big Thing: Your Agent Bill Can Shrink This Week

GPT-6.1 Terra Is Built for Computer Use at Half the Price

OpenAI released GPT-6.1 Terra on Oct 26. It's a fast model built for computer use, visual reasoning and writing: filling out forms, reading invoices and typing them into accounting software, working through browser jobs.

What it costs: $1 per million input tokens, $0.05 cached and $5 output. OpenAI says that's 50% below both GPT-5.6 Terra and GPT-6.1 Sol.

What OpenAI says it can do: 50.3% on Terminal-Bench 4.0 at max effort, more than double GPT-5.6 Terra's 21.5%. Those are OpenAI's numbers, not an independent test. OpenAI also says that for the hardest coding work, GPT-6.1 Sol and GPT-6 Astra are still the better pick.

Where you get it: Plus, Pro, Enterprise and Edu users have it in ChatGPT Work and Codex, and developers can call it in the API as gpt-6.1-terra.

Why you should care: if you run form-filling, invoice entry or browser jobs on a pricier model, this is the cheapest test you'll run all month. Here's how I'd do it.

Pull 10 real tasks. Grab 10 jobs your agent actually ran last week, including two that failed.
Change one line. In your existing OpenAI call, swap the model name:
model="gpt-6.1-terra"

Re-run and log. For each task, write down pass or fail and the cost. If you want your agent to keep score for you, paste this:

Run these 10 tasks with gpt-6.1-terra and then with our current model. Make a table with task, pass or fail, tokens used and cost for each model. Flag any task where Terra failed and the current model passed.

Switch only if it holds. If the pass rate matches and the bill drops, move that job over. Keep your current model on anything Terra fails.

→ Read OpenAI's Terra announcement here

⚡ Worth Your Time: Four Changes You Can Use Monday

1. Copilot Now Picks the Model for You

GitHub made Copilot Fusion generally available on Oct 28, on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains. You pick an effort level (Standard, High or Max) and Fusion decides, per prompt, whether one model answers, a cheap model drafts and a stronger one steps in, or one model drafts and another reviews.

GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the API cost per task. You're billed for each model it uses, with 30% off model costs on paid plans for a limited time.

Try it: open the model picker, choose Copilot Fusion, set it to Standard, and only step up to High or Max when a task needs it.

2. Claude's $200 Plan Drops the 5-Hour Limit

Anthropic renamed Max 20x to Ultra ($200 a month, no price change) and Max 5x to Max ($100). Verified Ultra accounts no longer have a 5-hour limit. Weekly limits stay the same, at 10 times Pro, so your total doesn't grow. You can just spend it in one long session.

Check yours: go to Settings, then Usage. If you don't see a Current session bar, your account is verified. Newer accounts keep a temporary limit until Anthropic verifies them automatically. The pricing page also now lists monthly Claude API credits on Max and Ultra.

3. One Robots.txt Group Now Speaks to Four AI Companies

On Oct 30 OpenAI said it will honor a shared robots.txt signal for AI training, so you can state your preference once instead of listing every crawler. It comes from an IETF draft written by an Anthropic standards lead. It's an individual draft, not an official standard yet.

Per The Verge, Anthropic already honors it, Google and Microsoft plan support next year, and OpenAI says "in the coming months." Your existing per-crawler rules keep working.

Copy this into your robots.txt to say no to AI training everywhere except a folder you're happy to share:

User-Agent: __usage_ai-training
Disallow: /
Allow: /press-kit/

Swap /press-kit/ for your own folder, or delete that line. One catch from the draft: a User-Agent: * group does not count for this signal, so it needs its own group.

4. Mistral Large 4 Is Now Open-Weight

On Oct 29 Mistral released the weights of Mistral Large 4, a 1-trillion-parameter multimodal model with low, medium and high reasoning modes. Mistral says it leads open-weight models on Terminal-Bench 4.0 at 41%. The API costs $0.70 per million input tokens and $2.10 output, and it's also on OpenRouter, Microsoft Foundry and NVIDIA NIM.

The catch: the license is a modified MIT. You can use, fine-tune and sell products built on it, but running it as a hosted model service needs a separate commercial license. Weights are on Hugging Face.

🛠️ Quick Hits

Agents inside GitHub Actions. GitHub Agentic Workflows went GA on Oct 28: describe a chore like issue triage in a Markdown file and it runs as an agent in Actions, read-only by default. The default Copilot engine needs a Copilot license.

Free Gemini edits your files in the chat. Every Gemini app user can now create and edit Docs, Sheets, slides and PDFs right in the chat. Most file types are web-only for now; PDFs also work on mobile.

🙅 Skip It: Gemini 4 Carbon

Google calls Gemini 4 Carbon its strongest model. Close read: it's going to a small group of cyber defenders first, then trusted partners, with no date for developers or businesses. The quoted $2/$10 price is introductory and rises to $4/$20 after. You can't use it Monday, so don't plan around it.

💬 Here's My Take

Every launch this week says the same thing: the work is getting cheaper.

But a cheaper model that fails one task in five isn't cheaper. That's why I keep pushing the 10-task test. It takes an hour and tells you more than any launch post.

So pick one job this week, run the test, and only switch what passes. And if you publish anything online, add those three robots.txt lines while you're at it.

Join the Free AI Academy →

Free, no card. Guides, tools and the people running them live here.

That's the Debrief. Which job are you re-running on Terra first? Reply and tell me, I read them all.

See you Friday with a free tool from the Academy.

From Drew and The AI Debrief team

P.S. Run the 10-task test and send us your before and after numbers. Hit reply, we read every one.

P.P.S. New here, or missed one? Every past edition is in the archive.