Anthropic's new top public model is 20% cheaper than Fable 5.1, but some old API calls now return 400 errors.

The Daily Bot

No. 028 · Tuesday, Nov 10, 2026

Note: The Daily Bot website is temporarily offline. Until it is back, this email carries every story in full. Source links below still go to the original posts, docs and changelogs.

Today's one job: if you call the Claude API, check your code for three changes that now return 400 errors before you switch to Claude Fable 5.2. Also inside: a fourth Grok Bot weekly limits reset, Grok Bot changelog notes through v0.79.3, private MCP servers over Tailscale, Bot refusals when a connector exists, ChatGPT dots on several computers, texting Hark, a founders briefing playbook, and the full digest.

In this issue: Claude Fable 5.2 · What's new (6) · How people are using it (1) · The digest by company · Known issues

Today's lead

Claude Fable 5.2 lands at $8/$40, with three API breaks to check (Anthropic · confirmed)

Claude Fable 5.2 landed Monday, Nov 9, on all platforms and on the API as claude-fable-5-2, at $8 per million input tokens and $40 per million output tokens. That is 20% below Fable 5.1's $10 and $50. If you build on the API, test before you swap the ID: the release notes list three changes that turn working Fable 5.1 code into 400 errors.

What changed

Around 2 PM ET on Nov 9, Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2. Anthropic says they are "the same underlying model." Fable carries cybersecurity and biology safeguards so it can go to all users; Mythos 5.2 goes only to organizations accepted into Anthropic's Cyber or Life Sciences Verification Programs.

Per the model overview, both have a 1M-token context window, 128K max output, always-on adaptive thinking, high as the API's default effort, and a June 2026 knowledge cutoff. Anthropic says outputs come "up to 30% faster" than the 5.1 models. An optional fast mode, which Anthropic says speeds token generation by up to 2.5 times, costs $16 and $80 per million tokens.

Claude Code v2.1.321, published to npm at 2:02 PM ET, makes Fable 5.2 "now the default Fable model."

Before vs after

Anthropic reports the 55% figure from "early testing across claude.ai, Claude Code, and Cowork," about 60% fewer in cybersecurity and 40% fewer in biology.

Why it matters

For API teams, the price cut is real but the migration is not a one-line change. The breaking changes sit in the release notes and docs, not the launch post. Our read: for subscribers, launch week is the cheapest time to test Fable against Opus 5.5 on your own work.

On coding, Anthropic reports 49.1% on its FrontierCode 2.0 evaluation at Xhigh effort, against 45.5% for Opus 5.5 and 32.4% for Fable 5.1. It also says Fable 5.2 at High effort scored about level with Opus 5.5 at Max "for roughly half the cost per task." These are Anthropic's own runs.

Gotchas

• Prefix edits now fail for everyone. Per the release notes, if your code edits the system prompt, tools or an earlier message and then sends back a Fable 5.2 or Mythos 5.2 thinking block from a later turn, the request returns a 400 error. The preserved thinking docs say to make integrations append-only.
• Thinking blocks are account-bound. They work only in the account that made them or a linked one, and switching a conversation back to another model drops Fable 5.2's thinking blocks silently.
• Computer use: computer_20251124 returns a 400 error except on Amazon Bedrock; use computer_toolset_20260801.
• Advisor tool (beta): a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor.
• Sampling and tool choice: the model page says to omit temperature, top_p and top_k, and that forced tool use (tool_choice any or tool) returns a 400 error.
• No default zero data retention. Both models require 30-day retention unless Anthropic expressly authorizes otherwise.
• Effort defaults differ. The API defaults to high; Anthropic's post says Medium is "the default in Claude apps and Claude Code."
• Cyber work still gets blocked. Fable 5.2's classifiers block high-risk cyber tasks and most research biology; blocked queries get a response from Opus 4.8 for cybersecurity and Opus 5 for biology, per the post.
• Limits drain faster. The Help Center says "Fable models use your usage limits faster than other Claude models."

Launch offers

Pro plans and standard Team seats get all Fable models included for up to 50% of weekly limits from Nov 9 through Nov 16 at 11:59 PM PT, plus one Fable reset that expires the same night. Max, Ultra, premium Team seats, and premium seats on legacy seat-based Enterprise plans get one to three resets by subscription length (less than six months, six to 12, more than 12), expiring March 8, 2027 at 11:59 PM PT. Resets go only to accounts with paid billing in place before Nov 9 at 10:00 AM PT, per the Help Center. The offers are not available on the Free plan, standard seats on legacy seat-based Enterprise plans, self-serve Enterprise plans, or usage-based Enterprise plans, and Fable limit resets are not available for HIPAA-ready Enterprise plans.

What to watch

Our read: watch the Claude apps release notes for when the 50% Fable allowance ends for Pro, the advisor docs for a wider list of accepted advisors on Fable 5.2 executors, and the system card follow-ups on Anthropic's note that both models voice suspicion of being evaluated more often than recent models.

Last checked: Nov 9, 2026, 3:58 to 3:58 PM ET What changed: refile for the Tuesday assemble. Date set to 2026-11-10, lead and launch-offer wording updated, fallback models named, and a cover added. Official pages, Help Center, release notes, model overviews, npm and the changelog were unchanged from the afternoon check.

What to do: 1. Per Anthropic: start most coding and knowledge work on Opus 5.5 and move up to Fable 5.2 for the most ambitious, open-ended technical work. 2. API: swap the model ID to claude-fable-5-2, then run a multi-turn test before production, since editing an earlier turn and sending back a Fable 5.2 thinking block returns a 400 error (Anthropic release notes). 3. Computer use: move from computer_20251124 to computer_toolset_20260801; the old tool returns a 400 error on every platform except Amazon Bedrock (Anthropic release notes). 4. Pro or Team standard seat: pick Fable in the model picker before Nov 16 at 11:59 PM PT to use the launch week, which covers up to 50% of weekly limits (Claude Help Center).

Prompt (ours):

You are reviewing my Claude API integration before I switch it to claude-fable-5-2. Read the code I paste and list every place where it (1) edits the system prompt, tools or an earlier message after the first turn, (2) declares computer_20251124, (3) sets an advisor model other than Fable 5.2 or Mythos 5.2, (4) sets temperature, top_p or top_k, or (5) forces tool use with tool_choice any or tool. For each, quote the line and say what to change. Do not guess about code I did not paste.

(ours, untested)

Sources: - Introducing Claude Fable 5.2 and Claude Mythos 5.2 (Anthropic, Nov 9, 2026) - Claude Platform release notes (Nov 9, 2026 entry) - Claude Fable 5.2 model overview (Claude Platform Docs) - Claude Mythos 5.2 model overview (Claude Platform Docs) - Claude Fable 5.2 launch offers (Claude Help Center) - Preserved thinking (Claude Platform Docs) - Claude Code CHANGELOG.md, v2.1.321 - @anthropic-ai/claude-code on npm - Claude Fable 5.2 and Mythos 5.2 System Card (Anthropic) - CNBC: Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 ahead of IPO

What's new

Grok Bot resets weekly limits for all users, a fourth time (Grok Bot · confirmed)

If you ran out of Grok Bot usage over the weekend, you have a full allowance again. The official @bot account said at 2:17 PM ET on Monday: "We've reset weekly limits for all Grok Bot users."

What changed

This is the fourth all-user reset @bot has announced, after Aug 26, Sep 5 and Oct 2, again with no reason. At 2:32 PM ET, Lauren Tan (@poteto), who works on Grok Bot, quoted the @bot post and wrote: "reset rolling out!"

On Sunday at 8:17 AM ET, Elon Musk posted: "Grok Bot fixes for improved token efficiency and smarter behavior are underway". At 9:25 PM ET he added he had done a round table review with the engineering team, and "Fixes are being implemented overnight and you should notice a dramatic improvement by tomorrow."

At 9:58 AM ET, Sawyer Merritt posted that limits had reset "early for some users." The @bot post makes it everyone (reported for the earlier partial resets; confirmed for the all-user reset).

Why it matters

If Musk's overnight fixes are real, the same routines should use less allowance this week. Nobody at SpaceXAI has said what those fixes were. The Grok Bot changelog now lists "Latest v0.79.3" dated Nov 6. None of the Nov 3 to 6 entries mention token efficiency or an all-user weekly limit reset.

How to use it

1. Check Settings, then Usage & Billing, and note today's reading.
2. Run your normal week, then compare by Wednesday.
3. If usage still drops much faster than before, Musk asked people to reply to his Sunday post.

Look at my Grok Bot routines and tell me which three probably use the most usage each week, and how I could make each one run less often or do less per run without losing what I rely on. Don't change anything yet.

(ours, untested)

Good to know

• The post does not say whether your next scheduled reset date moved. A recheck of @bot originals after 2:17 PM ET found no later post with a reason or a next-reset date.
• It also does not mention Grok chat limits or paid usage credits.

What to watch

Watch for a changelog entry that names the efficiency fixes, and for slower burn by midweek.

Prompt (ours):

Look at my Grok Bot routines and tell me which three probably use the most usage each week, and how I could make each one run less often or do less per run without losing what I rely on. Don't change anything yet.

(ours, untested)

Sources: - @bot on X: We've reset weekly limits for all Grok Bot users (Nov 9, 2026) - @poteto on X: reset rolling out! (Nov 9, 2026) - @elonmusk on X: round table review with the Grok Bot engineering team (Nov 8, 2026) - @elonmusk on X: Grok Bot fixes for improved token efficiency and smarter behavior are underway (Nov 8, 2026) - @SawyerMerritt on X: early resets for some users (Nov 9, 2026) - @bot on X: We've reset usage limits for all Grok Bot users (Oct 2, 2026) - @bot on X: We've reset usage limits for all Grok Bot users (Sep 5, 2026) - @bot on X: resetting weekly usage limits for all users (Aug 26, 2026) - Grok Bot changelog (Latest v0.79.3, Nov 6, 2026)

Grok Bot changelog catches up with four entries through v0.79.3 (Grok Bot · confirmed)

You can now read what shipped in the Grok Bot 0.79 builds: the Bot can request a subscription refund, shop for a gift, book on Delta, American, Alaska, JetBlue or Frontier, and recheck a checkout total before it charges. The official changelog now lists "Latest v0.79.3" dated Nov 6, plus entries for Nov 3, 4 and 5. Until Monday afternoon it still stopped at v0.78.1 from Oct 27.

What changed

Four dated entries are now on the page. Nov 3 (v0.78.4, also 0.78.2 and 0.78.3) adds refund requests and, after you approve a cart, pay-on-its-own checkout. Nov 4 adds scheduled reorders with one approval that shows the total. Nov 5 (also 0.79.0 and 0.79.1) adds gift requests, airline booking that "follows that site's steps," a hidden-instructions notice, and ssh and psql. Nov 6 (v0.79.3) adds a theme button on Shared Pages, and "A busy or unreachable phone call now shows No answer instead of Didn't go through."

Why it matters

On Monday at 9:32 AM ET, user Sebastian Guevara wrote that he was on 0.79.3 "and the changelog still stops at 0.78.1 from oct 27." Our Friday wrap had listed 0.79.0 to 0.79.2 as files on a download server, not a release. The page now names 0.79.0 and 0.79.1, not 0.79.2. A Nov 5 fix says "A one-off routine turns off after it runs instead of firing again next year."

How to use it

1. Open Settings and check that your app reads 0.79.3 or later.
2. Ask your Bot to request a refund on one subscription charge and read the list of ways to ask before you let it send anything.
3. If you use one-off reminders, set one and confirm it turns itself off after it runs.

Look at my Grok Bot routines and find any one-off reminders that are still turned on after they already ran. List them with their names and dates, and don't delete or pause anything until I say so.

(ours, untested)

Good to know

Phone calls show up only as card fixes (Nov 5 and 6). There is still no line announcing calls themselves, which @bot announced on Nov 4. Treat the dates as build dates. We have not run the refund or one-off-routine flows ourselves.

Last checked: Nov 9, 2026, 6:30 to 6:34 PM ET. What changed: first entry.

Prompt (ours):

Look at my Grok Bot routines and find any one-off reminders that are still turned on after they already ran. List them with their names and dates, and don't delete or pause anything until I say so.

(ours, untested)

Sources: - Grok Bot changelog (Latest v0.79.3, Nov 6, 2026) - @pargara_ on X: on 0.79.3 while the changelog still stopped at 0.78.1 (Nov 9, 2026) - @grokreleases on X: Grok Bot 0.79.3 summary (Nov 9, 2026)

Tailscale says Grok Bot can reach private MCP servers on your tailnet (Grok Bot · confirmed)

If your team runs an MCP server (a tool server an AI agent can call) on a private network, Tailscale says your Grok Bot can now use it without you opening it to the internet. Tailscale posted at 3:37 PM ET Monday: "@bot can now connect to private MCP servers on your tailnet without exposing them to the public internet."

What changed

Its blog post says: "This flow is already in production: Grok Bot can use OAuth apps to connect to private Model Context Protocol servers on your tailnet, without exposing them to the public internet." The Bot joins as "a device associated with your user identity," so your rules, logs and device limits apply.

The changelog on Oct 27 said the plugin "shows Not available yet on accounts that can't use it," and on Nov 3 a private Tailscale MCP server "now says to connect your tailnet from the Tailscale plugin, then try again."

Why it matters

The old options were to expose an internal server or skip it. The Bot can sit behind the same rules as a laptop, with less access. On Premium or Enterprise it gets "a custom posture attribute."

How to use it

Tailscale says: "Connect Grok Bot to your tailnet using the Tailscale plugin in the Grok Bot desktop app." A sample policy opens only tag:mcp-servers on port 443.

1. In the Grok Bot desktop app, open Connect Apps and add the Tailscale plugin.
2. Sign in to Tailscale and approve, so the Bot's computer joins your tailnet under your identity.
3. Limit the Bot's device to the MCP servers it needs, with a posture attribute on Premium or Enterprise or a tag on other plans.
4. Add the private MCP server in Grok Bot and ask the Bot to list its tools to confirm it connects.

I just connected Tailscale. List the hosts on my tailnet you can reach, then try the MCP server at <host name> and tell me which tools it offers. Don't call any tool that changes data yet.

(ours, untested)

Good to know

@bot has not posted an original about this. Confirmation is from Tailscale, the changelog, and Lauren Tan (@poteto) reposting it. Some accounts may still see Not available yet. We have not connected a tailnet ourselves.

Last checked: Nov 9, 2026, 6:30 to 6:34 PM ET. What changed: first entry.

Prompt (ours):

I just connected Tailscale. List the hosts on my tailnet you can reach, then try the MCP server at <host name> and tell me which tools it offers. Don't call any tool that changes data yet.

(ours, untested)

Sources: - Tailscale blog: Grok Bot has a Tailscale integration, and now you can, too (Nov 9, 2026) - @Tailscale on X: @bot can now connect to private MCP servers on your tailnet (Nov 9, 2026) - Grok Bot changelog (Nov 3 and Oct 27 Tailscale plugin entries) - @poteto on X: repost of the Tailscale post (Nov 9, 2026)

Grok Bot staff confirm bots refuse tasks when a connector exists (Grok Bot · confirmed)

If your Grok Bot started refusing work around Nov 6 and blamed an "access restriction," your setup is fine. Cursor staff confirmed Monday that the team knows and is fixing it. Local commands still work. You can get most posts out if the Bot drafts them.

What changed

Since about Nov 6, users say Bots stop browser work on Slack, GoHighLevel, Meta Ads, Xero, WordPress and X when a connector exists, even an unusable one. One X report says the connector "can’t create posts with images (it has no media upload), so the task just stops."

At 9:16 AM ET Nov 9, staff member Colin replied and in the X posting thread: "The team is aware of an issue where Grok Bot refuses actions when a connector exists (even for actions that can’t be taken via that connector). The team is working on this."

At 10:01 AM ET Colin added: "We’ve become aware that bots have been refusing more work than they should over the last few days, and the team is working on changes to improve this." A change made Bots "much more cautious about reaching into another program’s own data," and "regular commands still work."

Why it matters

Browser routines can stall at a draft. Users spent days on settings that were never the problem.

How to use it

1. Check Connect Apps for a connector for that service.
2. Ask which tool it would use, and if it can finish:

Before you refuse, tell me in one sentence which tool you would use for this task, whether that tool can actually do it, and what exactly is stopping you. Don't take any action yet.

(ours, untested) 3. For X posts, one user's stopgap is to let the Bot write the text and give you an X compose link, then press Post yourself; for images, attach the media by hand. It works for text, but pictures still need you in the loop.

Good to know

• No staff workaround or date. Watch for a follow-up, a changelog line, or paused routines finishing.
• Cloud-only coding jobs are a separate thread; staff said Nov 6 "this is how Grok Bot is set up to work right now".

Last checked: Nov 9, 2026, 10:33 to 10:36 AM ET. What changed: Tuesday cut; stopgap fixed; no staff fix after 10:30 AM ET.

Prompt (ours):

Before you refuse, tell me in one sentence which tool you would use for this task, whether that tool can actually do it, and what exactly is stopping you. Don't take any action yet.

(ours, untested)

Sources: - Cursor forum: Grok Bot refuses browser tasks across multiple websites, citing an access restriction (staff reply Nov 9) - Cursor forum: Grok Bot forces the X connector (no image upload) and refuses to post to X via browser since Nov 6 (staff reply Nov 9) - Cursor forum: Bots refusing local commands after update 0.79.2 (staff reply Nov 9)

ChatGPT dots can now use more than one of your computers (OpenAI · confirmed)

@OpenAIDevs posted Monday at 3:01 PM ET: "You can now connect multiple local computers to the same dot and choose which one it uses for a task." @OpenAI reposted it.

What changed

In a follow-up in the same thread, @OpenAIDevs offered three things to try: "ask your dot to test a change in Xcode on a second Mac and send you screenshots of how it looks"; "run the same test on your laptop and on a machine with an older SDK" and "compare the two results without changing your everyday setup"; and "Have your dot click through a flow on a second computer and capture where things break," which "keeps your own screen and cursor free while you work on the fix."

Why it matters

Neither the official posts nor the Learn pages we checked describe automatic machine picking.

How to use it

Per Learn: 1. "Open your dot’s profile in the ChatGPT app on that computer." 2. "Under Computers, find the computer you’re using and select Allow." "select Allow access to connect it." 3. "To connect another computer, repeat these steps in the ChatGPT app on that computer."

Good to know

• Learn: "Each thread uses one computer at a time."
• Help: "Changing your dot’s selected computer doesn’t move an existing task."
• Per Meet dots, "must be online with the ChatGPT app open." Per Learn, Offline "doesn’t revoke the connection"; use Revoke access.
• Per Help, connecting a computer to Codex or enabling Work Sync doesn’t give your dot local access.
• Neither post lists plans. Help lists eligibility: Pro users 18 and older outside the EEA, Switzerland, and the UK; Business Premium in supported ChatGPT regions, subject to workspace eligibility; and eligible Enterprise, Edu, and Healthcare workspaces. During the admin preview from November 4 through November 18, 2026, those workspace users can use dots if an admin enables access for their role. Starting November 18, dots are on by default unless an admin opts out. Workspaces with UAE data or inference residency are ineligible. Help also says Pro and Business Premium are rolling out gradually.

What to watch

samir (Dots & ChatGPT, per his bio) wrote: "all part of a bunch of improvements for how dots interact with desktop codex that we are working on." Watch What's new.

Last checked: Nov 9, 2026, 4:13 to 4:14 PM ET. What changed: r2.

Prompt (ours):

Start a separate thread on my desktop computer. Pull the latest code, run the build and the test suite there, and send me the failing tests. I'll keep working on my laptop.

(ours, untested)

Sources: - @OpenAIDevs on X: connect multiple local computers to the same dot (Nov 9, 2026) - @OpenAIDevs on X: things to try with a second computer (Nov 9, 2026) - @OpenAI on X: repost of @OpenAIDevs (Nov 9, 2026) - samir (OpenAI) on X: rolled out multiple computers for dots (Nov 9, 2026) - ChatGPT Learn: Connect computers and apps to your dot - ChatGPT Learn: Meet dots - OpenAI Help Center: Getting started with your dot

How to text Hark over iMessage or RCS, and what it can't do yet (Hark · reported)

You can now hand Hark a task by text. Since Monday, Hark answers over iMessage and RCS, so you can ask it to book, buy or look something up from your phone's own messaging app, and the conversation lands in the same thread you see in the Hark Pro app. If you already use Hark, ask it in chat to set up texting.

What changed

Hark's company account posted at 11:04 AM ET Monday: "Starting today, you can iMessage with Hark." It added: "Book appointments. Make purchases. Save money. All with Hark’s secure logins and payments fully intact." and "It’s the same Hark you know, now in iMessage and RCS." A follow-up post says "Everything you do on the web or app works over text: payments, computer use, logins, and more." Brett Adcock, Hark's founder, posted the same news moments later with a 36-second company demo video, in which someone texts Hark to send flowers to a coworker and picks an arrangement from a page Hark builds.

Before vs after. Until Monday, Hark ran inside its own apps and web app: Hark's announcement still says "Hark Pro’s native apps on iOS and Android will always offer the richest experience on mobile," and points to the Mac app for desktop. Now the same Hark also lives in your text thread, and per @hark_labs, "All of your texts will also appear in your existing thread in Hark Pro, so you can continue the conversation whenever and however you want."

The announcement lists what Hark does in a text thread:

• Status. Read receipts, a typing indicator, and tapbacks to acknowledge asks. When Hark starts Handoff, its cloud computer, it shows "an animated sticker that reads “Using Computer,” “Cancelled,” and “Completed.”" Abidur Chowdhury, whose X bio lists him as Hark's Head of Design, wrote that the team chose the sticker "to communicate the status of the task instead of sending extra messages."
• Rich media. "Your orders for food and products appear as images of a checkout cart, so you can audit a purchase before approving it." Flights show in full, "Long summaries arrive as beautifully formatted PDFs," and "comparisons will appear as clean tables." Trip plans arrive as Pages you open on your phone.
• Logins and cards. To sign in somewhere or add a card, "Hark will text you a Secured by Hark link to enter the information into your encrypted vault."

Why it matters

Our read: the useful part is not the chat, it is the review step moving into the thread you already check. A checkout image you can approve from a text, and a status sticker that tells you Handoff is working, mean you can start an errand on the go without opening the app. That is the case Hark itself makes: the announcement pictures firing off "a text on the subway" and having "a burrito waiting at your door by the time you arrive home."

How to use it

Per Hark's announcement and Terms:

1. In Hark Pro chat, ask Hark to set up texting, or "use a Panel posted in Home that offers the option to sign up." New users "can also register for texting during sign up."
2. Text Hark a task. When it goes to work on its cloud computer, look for the Using Computer sticker, then Completed or Cancelled.
3. Before you approve a purchase, open the checkout image and check the items and total.
4. When Hark needs a login or a card, use the Secured by Hark link it texts you. Don't type passwords or card numbers into the thread.
5. To stop, Hark's Terms say "You can stop all text messages from Hark at any time by removing your mobile phone number from your Account in your account settings or by telling Hark".

This prompt is ours and untested:

From now on, text me updates on errands. Before you buy anything, send me the checkout image and wait for me to reply yes. If you need a login or a card, send me a Secured by Hark link instead of asking me to type it here.

Good to know

• Outside the US, it may not be there yet. Two users pointed that way on Monday. Sparky300 (@sparky300) replied to Adcock: "Quick heads up to everyone. This new feature is only available to USA" (user report). David Rau (@drauflux) asked: "Please make it available outside of the US" (user report). We found no availability statement from Hark in its announcement or posts.
• RCS is a little behind iMessage. Asked whether Android gets it, Chowdhury replied: "Yes! Available on RCS/Google Messages, it’s missing a few nice to haves vs iMessage but is very complete and a pleasure to use too!" He did not say which features.
• No group chats, per one user. Randy Bigbie (@rbigbie) wrote: "Right now it doesn't support group text messaging. It also doesn't reply in the group that it doesn't support it." (user report). We did not reproduce this.
• What is encrypted. The announcement says vault details go through the secure link and "We never store your sensitive information in plain text." It does not say whether the text thread itself is end-to-end encrypted, which is one more reason to keep passwords and card numbers out of it.
• Message rates. Hark's Terms say texts include "replies to your requests, task updates, and account notifications" and "Message and data rates may apply."
• Same idea, different rival. ChatGPT's dots got texting over iMessage and RCS on Oct 28; see our story.

What to watch

Adcock posted at 11:20 PM ET Sunday: "Hark has 4 important product updates coming this week, first one tomorrow morning" (founder claim). Texting arrived Monday morning. Hark has not said whether it is one of the four. Watch @hark_labs for the updates Adcock teased, for an availability note outside the US, and for which RCS features catch up with iMessage.

Last checked: Nov 10, 2026, 6:29 AM to 6:29 AM ET. What changed: restored the Adcock source link in What to watch.

Prompt (ours):

From now on, text me updates on errands. Before you buy anything, send me the checkout image and wait for me to reply yes. If you need a login or a card, send me a Secured by Hark link instead of asking me to type it here.

(ours, untested)

Sources: - Hark: Hark arrives on iMessage and RCS (Nov 9, 2026) - @hark_labs on X: Starting today, you can iMessage with Hark (Nov 9, 2026) - @hark_labs on X: Already using Hark? Just ask it to text you (Nov 9, 2026) - @hark_labs on X: texts appear in your Hark Pro thread (Nov 9, 2026) - Brett Adcock on X: Today Hark is launching on iMessage (Nov 9, 2026) - Abidur Chowdhury on X: the animated status sticker (Nov 9, 2026) - Abidur Chowdhury on X: RCS is missing a few nice to haves (Nov 9, 2026) - Hark Terms of Service: SMS Terms (updated Oct 27, 2026) - Sparky300 on X: only available to USA (user report, Nov 9, 2026) - David Rau on X: make it available outside of the US (user report, Nov 9, 2026) - Randy Bigbie on X: no group text messaging (user report, Nov 9, 2026) - Brett Adcock on X: 4 important product updates this week (Nov 8, 2026)

How people are using it

Grok Bot for founders: one Bot, one daily briefing, nothing sent (Grok Bot · reported)

You can hand your morning catch-up to one Grok Bot that reads your mail, calendar and files and writes the same four-part briefing every day, without letting it send anything. On Nov 1, Hong Kong AI trainer Sam Wong wrote that he tried a crowd of Bots first, cut them to that one Bot, and kept it (his post). It fits founders and solo operators who get more from one fixed document than from many pings. It's his own account, and we haven't run it.

Not tested: we haven't run his setup on our own mail and calendar.

Who runs it

Wong describes himself as "a Hong Kong AI adoption trainer who has trained 11,000+ people." He uses Grok Bot two ways. On stage, "a group of role bots shows a team of AI agents working together live." At his desk, "I started with a team of bots and cut it to one bot that writes one daily briefing."

Disclosure: Wong sells corporate AI training, and all three posts we cite end with a pitch for it, including private Grok Bot classes. We found no tie to SpaceXSI or Cursor in the posts.

What he tried first

His Oct 16 setup post laid out the first version. "On 15 October I moved my recurring work off my laptop." Only one Bot, Coach, talked to him, and "Every other bot is a desk." A desk did "one job on a schedule and writes the result to a shared Drive folder." He listed six: Email, Calendar and chase, Radar, Workshop prep, Content and Event photos.

He set four rules before any Bot ran:

1. "Bots never send."
2. "Bots do not touch my calendar."
3. "A bot's report is a lead, not a fact."
4. "Anything that needs my laptop stays on my laptop."

His reason for the single talker: "I have ADHD. Ten bots pinging my phone would be ten more things I ignore."

What he cut, and why

It didn't hold. In the Nov 1 post, Wong wrote that each Bot "was useful on its own. Together they made more messages for me to read, not fewer." Coach's button questions went nowhere: "I did not answer one of them." He trimmed the team on Oct 17, and "22 October: I paused every bot except one." In his words: "there was so much noise. For my own desk, I am not a fan of multi-agents."

The one Bot he kept

The Bot "works on the cloud computer while my laptop is closed. It reads my mail, my calendar and my files." Each morning it writes one document, and "A short evening update follows." The sections never change:

1. Needs me: "at most three items, each with a recommendation and options A, B or C."
2. For my assistant: "draft text, reminders and follow-ups." His main AI assistant on his Mac picks these up.
3. FYI.
4. Closed since yesterday.

It also "looks for money items: bills, renewals and unpaid invoices." It "messages my phone only if a deadline is less than 24 hours away or money is at risk, and at most once a day."

The limits: "It never sends anything." "It never writes to my calendar." "It never deletes or moves mail." "It never pays." And "Its briefing is a lead, not a fact."

Still running? The trial "ran for a week and ended on 29 October. I kept it." He wrote "Every deadline reached me in time" and named one miss: "the target was one briefing a day, and I still get a morning and an evening document." His Nov 7 post covered teaching a VC firm's team, not his own desk. We found no later update on the Bot.

Why it matters at work

Our read: the hard part of a Bot setup is often attention, not capability. Six useful desks still added up to more to read, and Wong ignored the questions meant to help him. One document in a fixed order, capped at three decisions, is something a busy founder actually opens. He gave no time-saved figure, and his only result is that deadlines reached him in time.

Set it up

His four steps, from the Nov 1 post: "Pick the one document you want each morning." "Fix its sections and their order." "Write the list of things the bot must never do." "Decide who checks the bot's report before anything happens."

How to do that in Grok Bot is our part, built from his posts and Cursor's help pages:

1. Start with one Bot. Give it read access to mail, calendar and files only. His advice: "Add a second bot only when the first one is reliable and boring."
2. Write the never list into its instructions. Use his four: no sending, no calendar writes, no deleting or moving mail, no paying.
3. Add Ask first rules as a backstop. "An action that matches one of your Ask first rules always gets a card", and "If two rules conflict, Ask first wins" (Approvals help). Our read: add one rule for sending email and one for payments.
4. Make it a routine. Ask in chat for a morning routine. Pick a safe input before you press Test, since "A test does real work" (Routines help).
5. Read it for a week, then cut. Wong ran his trial for a week before deciding. Our read: if you skip a section all week, drop it.

Starter prompt (ours, untested)

Every weekday at 7:30 AM, read my email, my calendar and my files. Write one document with four sections in this order: Needs me (at most three items, each with a recommendation and options A, B or C), For my assistant (draft text, reminders, follow-ups), FYI, Closed since yesterday. Also list bills, renewals and unpaid invoices you find, each with the email or file it came from. Only message my phone if a deadline is under 24 hours away or money is at risk, at most once a day. Never send, pay, delete or move mail, or write to my calendar.

The four sections, the money scan, the phone rule and the never list are from his post. The time, the weekday schedule and the source for each money item are ours.

Good to know

• A lead, not a fact. Wong's rule from October: "If a bot says an invoice is unpaid, I check the bank before I chase anyone." Our read: keep the source link on every money item so the check takes seconds.
• Every Bot shares the computer. He wrote that "All your bots share one computer. Same files, same browser logins." His advice: "Don't put client files there that only one job needs."
• Where it runs. He wrote that "Grok Bot computers run in the United States today", citing Grok Bot's security docs. Check your own data rules first.
• Approval cards from routines expire. A card raised by background work "expires after 30 minutes" (Approvals help). Our read: a draft-only briefing should rarely raise one.
• Runs cost usage. "Each run spends usage, so a routine that runs often spends more than one that runs rarely" (Routines help). Two documents a day spend more than one.

What to watch

• One document or two. Signal: Wong updating the post after he folds the evening update into the morning one.
• Team numbers. His Nov 7 post said "No time saved is claimed yet" and that he will "time the weekly job before and after". Signal: before and after times added to that post.
• When to add a second Bot. Signal: a follow-up saying what he added once the first was "reliable and boring".

Prompt (ours):

Ours, untested. Every weekday at 7:30 AM, read my email, my calendar and my files. Write one document with four sections in this order: Needs me (at most three items, each with a recommendation and options A, B or C), For my assistant (draft text, reminders, follow-ups), FYI, Closed since yesterday. Also list bills, renewals and unpaid invoices you find, each with the email or file it came from. Only message my phone if a deadline is under 24 hours away or money is at risk, at most once a day. Never send, pay, delete or move mail, or write to my calendar.

(ours, untested)

Sources: - Sam Wong: How I use Grok Bot, a team of bots on stage, one bot at my desk (Nov 1) - Sam Wong: Grok Bot in Hong Kong, the rules I set first (Oct 16) - Sam Wong: Grok Bot private training for a VC firm's Hong Kong team (Nov 7) - Cursor Help Center: Routines - Cursor Help Center: Approvals and Auto-review

The digest

Grok Bot and Cursor

• Commands hanging on an "asleep" computer: edit startup files on your Mac; don't Reset. After Monday's item. Community Support Engineer Colin wrote at 5:10 AM ET that “Asleep” on the card is normal idle: it "wakes within a second or two whenever it’s needed". For one user, a common cause is "a line in the computer’s startup files that launches a program in the background." Update and Reset "bring the same file back," so have the bot read .profile, .bash_profile and .bashrc in /home/box with its file reader, not a shell command; copy the file to your Mac, comment out that line, copy it back, click Update Bot's Computer once, then run echo ok. 24/7 programs pause during sleep.
• If your bot's computer is stuck and support says it's on their side, hold off on Reset. Cursor staffer mohitjain told one user at 5:14 AM ET that their computer "is stuck on a server that hit a hardware problem, which is also why Update just loops." He asked them to "hold off on Reset so you don’t lose your files and logins" (post). In a second thread at 5:11 AM ET, mohitjain told another user that two failed Resets "didn’t wipe anything" and to "hold off on Reset and Update for now" while the team recovers it through the email support ticket (post). Both are single-user threads; staff gave no count or timeline.
• Grok Bot's Slack app is failing every DM for one user; the app chat still works. mohitjain said at 5:13 AM ET that the error is "on our side, not your Slack setup," that the bot "hits an error while preparing each Slack reply, before it ever reads your message," and that "the bot still works normally in the Grok Bot app chat" (post). No need to reinstall. One report so far.
• GitLab is missing from Cursor Automations' Add Trigger menu; duplicate an old automation instead. Colin confirmed at 5:12 AM ET that the new Automations page "currently leaves GitLab out of the Add Trigger menu," and existing GitLab automations "aren’t affected" (post). His workaround: duplicate an existing GitLab automation, edit the GitLab trigger already in the copy (for example merge request opened), update the prompt and save. It only helps if you already have one GitLab automation.
• Orange needs-input dot stuck on Cursor's Agents list after a subagent approval. Colin said at 5:11 AM ET that after you approve a subagent's request, "the “waiting on you” state isn’t always cleared." That stuck state also hides the working indicator (post). The fix that worked for another user: wait for the subagent to finish, then open each subagent once from the parent chat and click back. Passed to the team.
• Claude Fable 5.2 is now available in Cursor. At 2:52 PM ET Monday, @cursor_ai posted: "Claude Fable 5.2 is now available in Cursor!" It called Fable 5.2 "the top model on CursorBench at 60%" at max effort, and said it "scores 56.9% on CursorBench at default effort, up from 49.2% for Fable 5.1." We have not run CursorBench ourselves.

Anthropic

• One sentence changes how Fable 5.2 refactors. Anthropic's Prompting Claude Fable 5.2 guide says the model "responds closely to what the prompt says about who owns the code it works on." It suggests this line when you want structural cleanup: "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." In Anthropic's small coding tests, the model merged repeated code in about half of tries with it and under a tenth without. The costs: about 1.5 times the output tokens, unrequested changes, and existing feature behavior changed in one try in six to ten. Review the diff, not the summary, and leave it out when you own the code. (confirmed)
• Anthropic's model guide: start on Opus 5.5, not Fable 5.2. On Nov 9 at 4:03 PM ET, @ClaudeDevs posted a new guide, Which Claude model should you build with? by Addy Osmani. Its short answer: "start most work on Opus 5.5 at its default effort level, medium." It points simple high-volume work and subagents to Haiku 5.5, scoped coding and documents to Sonnet 5.5, and the hardest technical projects to Fable 5.2. It suggests trying a different effort level before switching models. One cost trap: the prompt cache belongs to the model that wrote it, so the first turn on a new model pays the cache-write price on the whole conversation. The guide suggests running /compact first in Claude Code. (confirmed)
• Fable 5.2 designed three working gadgets, per Anthropic. On Nov 9, Anthropic published Testing Claude on real-world electronics design. It says Fable 5.2 designed a USB speakerphone, a voice-assistant hub and a handheld virtual pet from short prompts, covering schematics, board layout, enclosures, firmware and parts lists. Diode Computers, which has a commercial relationship with Anthropic, built and tested the prototypes, and Anthropic says they "worked without any fixes to their circuitry." Anthropic reports 69% on PCB Gym, an internal benchmark Diode ran, versus 25% for Fable 5. The limits it names: the devices have not had electrical validation, antenna tuning, thermal testing or certification, and humans still built and tested them. (confirmed)
• Claude Code 2.1.322 adds fast mode on Vertex. Claude Code 2.1.322 is now npm latest and next, published Nov 9 at 8:58 PM ET on the npm registry (GitHub release Nov 9 at 9:29 PM ET). Its single changelog line: "Added fast mode on Vertex in the global, us and eu regions, falling back to standard speed if Vertex refuses it." If you run Claude Code through Google Cloud Vertex AI in other regions, this changes nothing for you, and a refused request falls back to standard speed rather than failing. npm stable is still 2.1.311. (confirmed)

OpenAI

• ChatGPT Enterprise is on AWS Marketplace. ChatGPT Enterprise, including Chat, ChatGPT Work, Codex, and dots, can now be bought through AWS Marketplace, per OpenAI's November 9 post: "Today, ChatGPT Enterprise is available through AWS Marketplace." Eligible spend can apply toward AWS commitments. OpenAI says AWS customers get the same models, features, and functionality available directly from OpenAI, with new features released at the same time. Eligible customers can also get ChatGPT for Healthcare and ChatGPT for Financial Services through Marketplace. The post is dated November 9; it does not give a clock time. (confirmed)
• Known issues: Costs API exports had wrong data; API errors resolved. OpenAI's status page posted incorrect cost data in Costs API exports at 11:42 PM ET Monday: "We have identified an issue causing incorrect cost data in Costs API exports published between November 6 and 9." It marked the incident resolved Tuesday at 2:02 AM ET: "All affected services are operating normally again." Our read: that note doesn't say whether downloaded exports changed, so pull November 6 to 9 exports again before you reconcile. Monday's API errors (6:08 to 6:50 PM ET) and ChatGPT conversation errors (1:16 to 2:08 PM ET) are resolved. The U.S. business Mastercard incident is still monitoring, last updated Friday at 8:29 PM ET. (confirmed)
• Codex CLI 0.169.0-alpha.7 is a pre-release with no notes. OpenAI published Codex CLI pre-release 0.169.0-alpha.7 on GitHub on Monday at 4:55 PM ET. Its notes say only "Release 0.169.0-alpha.7." A separate pre-release, 0.167.0-alpha.4.1, followed at 6:45 PM ET with the same kind of one-line note. The npm registry now points the alpha tag at 0.169.0-alpha.7 (npm time 5:05 PM ET Monday). The latest tag is still 0.167.0, from Thursday, November 5, at 8:46 PM ET. Version tags alone don't show new capabilities; watch the GitHub notes or a new stable tag. (confirmed, pre-release)
• OpenAI's Control Room can stop a research job's model output within 35 seconds. OpenAI described Control Room on Monday, "a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise." It "halts model output within 35 seconds of a decision to intervene to stop or pause a job," and fails closed: "all inference requests will be denied until Control Room resumes normal operation." Fail-closed covers frontier model research workloads now; the rest follow "over the next few days." @OpenAI posted it at 5:01 PM ET. Our read: in the Sep 20 DNS incident from our containment explainer, "the run did not stop automatically as expected" and was killed 2.5 hours later. The post cites it as something "caught by a monitor." (confirmed)

Hark

• iOS 2.9.0 is "laying the groundwork"; the Mac download changed too. Apple's lookup at 6:06 AM ET Tuesday shows version 2.9.0 (released Mon Nov 9, 2:00 PM ET), notes "Laying the groundwork for some exciting releases. Stay tuned.", 2,842 ratings, average 4.72. Monday's issue had 2.8.1 with 2,800. The Mac DMG, checked at 6:06 AM ET by HTTP headers (not an install), now shows Last-Modified 9:42 PM ET Monday and 20,696,177 bytes, up from 20,465,674. Google Play's listing still shows 1.0.0-81, "Updated on Nov 8, 2026". (store listings)
• Hark's tip for Projects users: pin a Panel. Kevin Kavanagh (@krkav) wrote at 12:04 PM ET Monday that he keeps separate Projects "for my condo board volunteer work, one for a never-ending email cleanup, one for dumping paid Microsoft 365, one for Amazon. Each keeps its own memory, files and reminders." (user report) Hark's account quoted him at 8:05 PM ET: "10/10 already. For anything you check often, add a panel on Home so you can see it at a glance." (company post)
• Day-one texting replies include usage and sign-in complaints. Replying to Adcock's launch post, Vince (@VinceOnAI) wrote at 9:43 PM ET Monday: "One message and the entire monthly credit is gone." (user report) Earlier, at 12:28 PM ET, 0xSantos said that after logging out to add a phone number, Hark "doesn’t seem to accept my Authenticator code or even my recovery codes" (user report). Hark replied: "We've DMed you." We reproduced neither.

Last checked: Nov 10, 2026, 5:52 to 6:53 AM ET across the five desks. What changed: Tuesday issue. Standalones held from Monday afternoon and evening, plus Tuesday morning digest lines.

Known issues and workarounds

• Grok Bot computer hangs on an "asleep" computer (Grok Bot, staff guidance): have the bot read .profile, .bash_profile and .bashrc with its file reader, comment out the background launch line on your Mac, copy it back, click Update Bot's Computer once, and skip Reset.
• Bot computer stuck on a server hardware problem (Grok Bot, staff guidance, single-user threads): hold off on Reset and Update while support recovers it through your ticket.
• Grok Bot Slack app fails every DM for one user (Grok Bot, staff say it is on their side): the app chat still works, no need to reinstall.
• GitLab missing from Cursor Automations' Add Trigger menu (Cursor, confirmed by staff): duplicate an existing GitLab automation and edit its trigger.
• Orange needs-input dot stuck on Cursor's Agents list (Cursor, passed to the team): wait for the subagent to finish, then open each subagent once from the parent chat and click back.
• Bots refuse tasks when a connector exists (Grok Bot, confirmed by staff, fix in progress): see the full story above for the prompt and the X posting stopgap.
• Costs API exports for Nov 6 to 9 had incorrect data (OpenAI, resolved): our read is to pull those exports again before you reconcile.

You're getting this because you subscribed to The Daily Bot. Reply to this email with corrections or tips; a person reads every reply. How we label: confirmed means we checked a primary source or staff post; reported means a credible claim we haven't verified; "Our read" is our own analysis.