Give your agent a phone, a web host and a fast yes or noThis issue is landing later than usual, sorry for the wait. This week's releases let your coding agent tap through a real phone, replay a browser chore without the browser, sort messages in one call, publish an app with its database and search the live web. Five ways to use them, each with a prompt to paste. Reported we read the primary source and link it. Tested we ran it ourselves and say exactly how.
|
Warning · Check your share settings
If you shared an app from Grok Build
Security firm Claroc tried 43,551 Grok app addresses under grok.me. Of the 37,571 still online, 35,578 (94.7%) opened for anyone with no sign-in. Each app gets its own hostname, and passive DNS services record hostnames, so a link nobody sent you can still be found. Claroc says Publish proposed "Anyone with the link" in its account; the company told Ars Technica apps are private unless the creator shares them. Either way, check yours:
Check in two minutes 1. Open your app's link in a private window, signed out of everything.
2. If it loads, set access to Only you or Invited users.
3. If it has "Continue with Google", sign in with a spare Google account
and confirm it can't read, edit or delete your records. |
In a team app Grok built for Claroc, any Google account could read, edit and delete the records. Run the same check on any AI-built app that publishes to its own subdomain. Reported Source: Claroc research, Nov 10 · Ars Technica
|
|
Nov 9 Reported
Claude Fable 5.2
Anthropic's most capable generally available model now costs $8 per 1M input and $40 per 1M output tokens and writes up to 30% faster than Fable 5.1. Optional fast mode runs up to 2.5x faster at $16/$80. Anthropic still says to start with Opus 5.5 for most coding and step up to Fable for the hardest work.
Source: Anthropic announcement · Our page
|
|
Nov 10 Reported
Gemini 4 Carbon opens to every paying API developer
Nov 30 Last day of Carbon's $2/$10 intro price. After that Google charges $4/$20 per 1M tokens. Google's strongest model is in the Gemini API and AI Studio for all paying developers at an introductory $2/$10 per 1M tokens. Google says it ranks first on CursorBench 4.0. It also previews /swarm in Antigravity, which turns one prompt into a team of agents.
Source: Google DeepMind · Our page
|
|
Nov 12 Reported
Gemini 4 Hydrogen
Google's fast Gemini 4 model, flat $1.50/$7.50 per 1M tokens across a 1M-token context, in the API on every tier. Google reports 57% on Terminal-Bench 4.1 against 18.2% for Gemini 3.8 Flash, and tasks finished 2.5x faster.
Source: Google announcement · Our page
|
|
Nov 16 Reported Out this morning
Qwen4-Max
Alibaba's new flagship: a 2.4-trillion-parameter mixture-of-experts model that activates 71B per token, with a 1M-token context and image and video input. QwenCloud lists $1.50/$4.50 per 1M tokens and an OpenAI- and Anthropic-compatible API. Smaller open-weight Qwen4 models are promised, not out.
Source: Alibaba Cloud announcement · QwenCloud pricing · Our page
|
Each one runs on a specific upgrade from this week, with a prompt to paste. We haven't run these setups ourselves, so all five are Reported. |
01
Let Claude Code test your mobile app on a real phone
Reported
|
|
Result
A recorded run of your app on a real phone, with a pass/fail note for every step. |
Runs on Pinky, a free Apache-2.0 desktop app for macOS, Windows and Linux (public since Nov 6, pre-alpha, with new builds daily). Claude Code, Codex, Cursor or any MCP agent can operate a real Android phone (Android 10 to 17), an iPhone on iOS 17.4 or later, or an emulator. Payments wait for your click in Pinky's gateway, and every run is recorded as steps, screenshots and a screen recording. Source: Pinky README · ReleasesYou'll build A test pass of your Expo or React Native app on an actual device: the agent signs up, walks the main flow and writes down every step that broke, with a recording you can export as an HTML report. Paste into Claude Code, Codex or Cursor Use the pinky MCP tools. Work on my Android emulator (or the phone I have plugged in).
Open my app "<App name>".
Sign up with test+pinky@example.com, finish onboarding, create one item, edit it, then delete it.
After each step, note passed / failed / what looked wrong
(cut-off text, overlapping buttons, screens that take more than 2 seconds).
Don't enter payment details. If a step needs a password or a code, stop and ask me.
Write the results to pinky-report.md, failures first. |
- Download Pinky from the releases page. Plug in an Android phone over USB (Pinky brings its own adb) or start an emulator from Android Studio.
- In Pinky, open Settings › Agents and click Connect next to your agent. By hand in Claude Code:
claude mcp add --scope user pinky -- pinkyctl mcp. - Use a test account. Agents can't type into password or one-time-code fields, and banking and wallet apps are hidden from them by default.
1 click to connect Claude Code, Codex or Cursor from Pinky's Settings › Agents, per its README. |
Why now Agents could already click through a web app in a browser. A phone app meant tapping through it yourself or writing Detox or Appium scripts. Now the agent that wrote the screen can tap through it on a device. Cost Free, open source and no account. It runs on your agent's plan. iPhones on iOS 17.4 to 26 need a one-time install from a Mac with Xcode, per the README. |
02
Turn a browser chore into a command that runs without a browser
Reported
|
|
Result
A saved command that fetches the same data over HTTP, using the browser only as a fallback. |
Runs on agent-browser 0.42 (Nov 9), Vercel Labs' browser CLI for agents, added saved flows (experimental). You record a task once. Each later run tries a direct HTTP request first, then the same request from the page, then a replay of the recorded clicks, and keeps whichever worked. In 0.42, runs send requests only to public addresses. Source: v0.42.0 release notes · Flow commandsYou'll build A one-line command that pulls the same data from a public site every morning (a supplier's stock list, a competitor's price table, a changelog) and returns only the fields you want. Paste into Claude Code, Codex or Cursor Use agent-browser (0.42 or later) and its saved flows.
1) agent-browser --capture flow open https://<public-site>/<search-page>
2) Search for "<term>", open the results, read name, price and stock for each item.
3) agent-browser flow save stock-check --param query="<term>"
4) Test it with another term:
agent-browser flow run stock-check --param query="<other term>" --limit 10
If the run comes back unverified, tell me which step failed and drive the page instead.
Then write scripts/stock-check.sh that runs the flow and appends the JSON to data/stock.jsonl. |
- Install with
npm install -g agent-browser, then agent-browser install for Chrome (first time only). It needs Node 22. - Saved flows keep no cookies or tokens, so start with public pages. A run on a site that changed is marked unverified rather than returning stale data.
- Once a manual run works, schedule the script with cron or a CI job.
3 tiers Direct HTTP, then the request from the page, then a replay of the clicks. The run keeps whichever worked. |
Why now Before, repeating a browser task meant the agent clicked through the page every time, spending minutes and tokens, or you worked out the site's hidden API yourself. Now the recording finds the request for you. Cost Free and Apache-2.0 licensed; it runs on your machine. Saved flows are marked experimental in the release notes. |
03
Sort every form submission into spam, sales or support in one call
Reported
|
|
Result
Every submission arrives labelled spam or not, with a category and urgency, and the unsure ones flagged for you. |
Runs on The Gemini Decisions API (Nov 11) and its new gemini-decision-1 model. It scores options instead of writing text: yes/no checks, multiple choice from 2 to 255 options, and rubric scores, up to 1,000 questions about one input. Every answer comes with a confidence from 0 to 1. Source: Google announcementYou'll build A triage step for your contact form or support inbox. Each message gets a spam check, a category and an urgency score, and anything the model is unsure about comes to you. Paste into Claude Code, Codex or Cursor Add a triage step to my contact form handler with the Gemini Decisions API
(google-genai SDK, api_version v1beta, client.decisions.create, model "gemini-decision-1").
Follow Google's Decisions API docs. Ask three questions in one request per submission:
is_spam: type "noul" (yes/no)
category: type "choice", criteria sales / support / bug_report / partnership / other
urgency: type "score", criteria ["Can wait a week", "Reply this week", "Reply today"]
Save each answer and its confidence on the submission row.
If is_spam is likely, archive it. If any confidence is below 0.7, tag it needs_review.
Otherwise route sales and partnership to my email, support and bug_report to the support queue.
Add a test with three sample messages. |
- Get an API key in Google AI Studio. Google offers a no-cost tier for prototyping.
- Start with the 0.7 cutoff, read a week of needs_review items, then move it.
- Google says integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM and others if you'd rather not call google-genai directly.
$0 for output tokens. Input is $0.10 per 1M tokens, per Google's Nov 11 post. |
Why now Classifying with a chat model meant parsing generated text and paying for output tokens on every message. This returns probabilities your code can threshold, and Google bills no output tokens. Cost $0.10 per 1M input tokens and $0 for output, per Google. The shared input is billed once per request however many questions you ask. |
04
Put the app your agent built online, with a database, behind a private link
Reported
|
|
Result
A live app with its own database at a private link, with versions you can roll back. |
Runs on Spacefast (Nov 9), Automattic's hosting for what your own agent builds. It is not an AI builder: Claude Code, Codex, Cursor or ChatGPT publish to it through a plugin, MCP server, CLI or API. Every Space comes with a database, file uploads, versions you can roll back and comments pinned to the page. Spaces are private by default. Source: Spacefast · Spacefast llms.txt · Automattic announcementYou'll build A small app (a sign-up sheet, a team booking tool, a client feedback page) live at a private link your testers can open and comment on, with its data stored in the Space's own database. Paste into Claude Code, Codex or Cursor Publish this project with Spacefast.
Read https://spacefast.com/skill.md first and follow it.
Store form submissions in the Space's built-in database, not localStorage.
Keep the Space private. Create one share link for my testers and give it to me.
Turn on comments. Then give me the claim link so I can keep the Space. |
- Connect your agent: the official Claude, ChatGPT/Codex and Cursor plugins, or the local MCP server with
npx -y spacefast mcp. Or publish a folder yourself with npx -y spacefast publish ./dist. - Claim the Space. Unclaimed Spaces normally expire 33 hours and 20 minutes after the last publish that changes files.
- When testers comment on the page, ask your agent to read the comments and fix each one.
33h 20m How long an unclaimed Space stays up after your last publish. Claim it to keep it. |
Why now A vibe-coded app used to stop at localhost, or at a deploy guide with a separate database and hosting account. Here the agent publishes it and the database comes with the Space. Cost Personal is $0 with $1 of usage a month and no card. Go is $5 a month with $5 of usage included. Built-in app sign-in isn't here yet; Automattic says it's coming in the next few weeks. |
05
Give your app's AI answers that cite this week's web, with one API call
Reported
|
|
Result
Answers in your app that cite pages from domains you picked, published in the last 30 days. |
Runs on The Web Search API on Amazon Bedrock AgentCore (Nov 11). One InvokeWebSearch call returns up to 25 ranked results from an index Amazon runs, each with a text extract plus title, URL and publish time when available. It's signed with your AWS credentials, so there's no API key and no gateway to set up. Source: AWS announcementYou'll build A "what's new" answer in your app: the user asks, your server searches only domains you trust from the last 30 days, and your model answers with the URLs it used. Paste into Claude Code, Codex or Cursor Add a server-side web search tool with the Amazon Bedrock AgentCore Web Search API.
Use boto3 1.43.127+ (client "bedrock-agentcore", region us-west-2), invoke_web_search with
webSearchId "web-search.v1", maxResults 8,
filters.domainFilter.include from config/trusted-domains.json,
filters.publishedDateFilter.publishedAfter = 30 days ago (RFC 3339, UTC).
Pass title, url, publishedAt and text of each result to the model,
and require a cited URL for every claim. Cache results per query for 1 hour.
Log every query so I can see the bill coming. |
- Try it in AWS CloudShell first:
aws bedrock-agentcore invoke-web-search --web-search-id web-search.v1 --search-query "latest EKS features" --max-results 5. - Give your app's IAM role one permission:
bedrock-agentcore:InvokeWebSearch. - It runs in six Regions: N. Virginia, Ohio, Oregon, Ireland, Tokyo and Sydney. Queries can be up to 200 characters.
$7 per 1,000 queries, with no gateway charges, per AWS's Nov 11 post. |
Why now Web search on AgentCore used to need a gateway, a search target and an MCP client. Now your own code makes one signed call, and each request carries its own domain and date filters. Cost $7 per 1,000 queries and no gateway charges, per AWS. The cache line in the prompt is there because every repeated question is another query. Self-hosted Claude Managed Agents Companies can now run Anthropic's whole Managed Agents service in their own cloud, but it's a private beta you request through an Anthropic account team. If you build alone in Claude Code, nothing about your week changes. Reported Source: Anthropic Checked · Nov 14 to 16
50 models in our LLM API pricing table, each re-read on the vendor's own page |
This week we added Claude Fable 5.2 at $8/$40 and, this morning, Qwen4-Max at $1.50/$4.50. Our Sunday recheck of every tool pricing page we list found no price changes. Anything we get wrong goes on the corrections page. LLM API pricing comparison · Corrections No sponsor this week. Sponsor an issue. No affiliate links in this issue. |