Three releases from the past ten days map neatly onto ZetsuEDU's framework. One lets your coding agent create media from the terminal. One makes screen-reading, computer-using agents cheaper to build. One puts invoices, payment links, and purchases inside an agent that asks before it spends. Pick the test that matches a job you already do.
The releases below were announced between October 19 and October 26, 2026. The tests are proposed builder tests, not results from ZetsuEDU testing. Benchmarks and performance claims are company-reported.
1. Gemini API CLI for agent-made media (Create)
Google released an experimental preview of the Gemini API CLI on October 19. The command is gemini-api, a zero-dependency binary for macOS, Linux, and Windows. It is a separate tool from the older Gemini CLI coding agent, and the GitHub repository is open source under Apache 2.0.
The point is to give any agent with shell access, such as Claude Code, Codex, Cursor, or Antigravity, one predictable way to call Gemini. Commands cover text generation, images, speech, music, video, file and YouTube analysis, transcription to formats including SRT captions, and background managed agents such as Deep Research. The README lists the default models: Gemini 3.8 Flash for text and analysis, Nano Banana 2.1 for images, Gemini Omni 1.1 Flash for video, Lyria 3.5 for music, and Gemini 3.8 Flash TTS for speech.
Two design choices matter for agent workflows. Media commands write the file to disk and print only its path, so your agent does not fill its context with base64 data. And every command describes itself offline through --help, --usage, and --schema, with --dry-run to validate a call without touching the network.
AI short-film makers, landing-page builders, and anyone who already asks a coding agent to build pages should care. Asset creation can become one more step in the same agent run instead of a separate trip to five apps.
A concrete builder test
Install the CLI, set GEMINI_API_KEY, and add the README's short instruction block to your AGENTS.md or CLAUDE.md. Then give your coding agent one real brief: build a landing page for a product you sell, with a hero image, a 15-second voiceover, and a short background track, all saved into ./public/.
Record how many commands the agent ran, how many failed and why, whether it used --dry-run before spending, how many assets you kept without regenerating, and the total API cost. Keep it if the agent produced usable assets with fewer manual steps than your current tool-hopping process.
Limits to check
This is a v0.4.x experimental preview, and Google says the command surface is still changing, so pin a version before you script around it. The CLI itself is free, but every call bills against your Gemini API key at that model's rate. The announcement lists no prices, so check the Gemini API pricing page for the video and music models before batching. Files uploaded through its Files API commands are kept for 48 hours. Install scripts pipe a remote script into your shell, so read the script first or use go install.
2. GPT-6.1 Terra for cheaper computer-use agents (Build)
OpenAI introduced GPT-6.1 Terra on October 26. It is a fast, lower-cost model aimed at visual reasoning, computer use, and everyday writing. OpenAI's example workloads include filling out forms, reading incoming invoices and entering them into accounting software, and interpreting technical drawings.
It is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. The model page lists text and image input, text output, a 1,050,000-token context window with up to 922,000 input tokens, and 128,000 max output tokens. Tool calling requires the Responses API, where supported tools include computer_use, mcp, web_search, and hosted_shell. The same day, GitHub made it generally available in Copilot for Pro, Pro+, Max, Business, and Enterprise users, and AWS added it to Amazon Bedrock.
OpenAI reports that, at medium reasoning effort, Terra scores 15 percentage points higher than GPT-5.6 Terra on its Agents' Last Exam computer-use benchmark and finishes tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol. It also reports 50.3% on Terminal-Bench 4.0 at max effort, up from 21.5% for GPT-5.6 Terra. These are company-reported numbers. OpenAI itself says Sol and Astra remain the better choice for the most complex coding tasks.
Builders automating repetitive screen work, such as data entry, form filling, or checking listings across sites, should care. A cheaper computer-use model can turn a demo that was too expensive to run daily into something worth keeping on.
A concrete builder test
Pick one boring task you repeat every week, such as copying 20 invoice or order details from PDFs or screenshots into a spreadsheet or back-office form. Run it through the Responses API with Terra at medium effort, then again at high effort. Keep the same 20 items as your baseline set.
For each run, count fields entered correctly, items where the agent stopped or asked for help, wall-clock minutes, and total cost from your usage dashboard. Adopt it only if accuracy holds on all 20 items and the cost per item beats the time you would spend doing it yourself.
Costs and limits to check
Standard pricing is $1 per million input tokens, $0.05 cached, and $5 per million output tokens, with Batch and Flex at half price. Prompts over 272,000 input tokens are billed at 2x input and 1.5x output for the whole request, so long screenshot histories can get expensive fast. Check the pricing page for any per-call fees on computer-use and search tools. OpenAI says the model meets the Critical cybersecurity threshold under its Preparedness Framework and keeps the safeguards built for GPT-6 Astra. Run computer-use agents in a sandboxed browser or test account, never on a logged-in admin session.
3. Claude Duet with Stripe and Shop Pay for approval-gated money tasks (Monetize)
Anthropic opened a research preview of Claude Duet on October 21. It is a separate app on iOS, Android, and claude.ai where Claude works as a proactive agent. It runs on Claude Opus 5.5, uses its own virtual machine and browser so your devices can be off, checks your connected accounts every few hours, and sends a morning brief.
The monetization angle is in the integrations. Anthropic lists checking QuickBooks or Stripe for unpaid invoices and drafting reminders, plus setting up and managing a Shopify store. Stripe announced the same day that Duet includes Stripe MCP, which can send invoices and payment links, manage products and prices, analyze revenue, and issue refunds. Duet can also buy things with Link, Stripe's wallet, which pays with a virtual card after you approve the spend, and it can check out with Shop Pay from Shopify.
The guardrails are the part to study. Duet shows you exactly what it will do and waits for approval before it sends a message, makes a payment, signs a document, or deletes anything. Trading securities and moving money between accounts are blocked even with approval. Anthropic also describes a classifier designed to reject sharing credit card numbers and government ID.
Solo creators and small business owners who chase invoices, sell digital products, or run a small store should care. The approval pattern is also a useful model for any money-touching agent you build yourself.
A concrete builder test
Use a Stripe sandbox, not live mode. Create three test products, such as a template pack, a guide, and a coaching call, plus five test invoices with two marked overdue. If you have Duet access, connect Stripe there. If not, connect the Stripe MCP server to the agent you already use with OAuth and a sandbox.
Ask the agent to create a payment link for each product, find the overdue invoices, and draft reminder messages. Count wrong amounts or products, actions that ran without asking you, and minutes saved against doing it by hand. Keep it only if every write action waited for approval and no amount was wrong.
Limits to check
Duet is waitlisted and starts with US Pro and Max subscribers only. It has no separate price but shares your plan's usage limits, which is why it checks accounts every few hours rather than constantly. Anthropic says it may still misread a thread or misjudge what is urgent, and it cannot yet show you its browser. On the Stripe side, Stripe's MCP docs say refunds and outbound payments need human confirmation, and that from December the server will stop accepting full-access secret keys and restricted keys without the Agent tag. Migrate to OAuth or agent API keys by November 30, 2026.
ZetsuEDU takeaway
Run one test, not three. Create: have your coding agent generate a page's assets with the Gemini API CLI and count the manual steps. Build: move one weekly data-entry task to GPT-6.1 Terra and measure accuracy and cost per item. Monetize: put products and overdue invoices in a Stripe sandbox and check that every money action waits for you.
Adopt a tool only when your own numbers beat your current workflow. If an agent touches money, approval before every write is the minimum bar.
