The OpenClaw 4.20 update just dropped and it changes AI agents forever.
I've been through the whole release.
What's new, what got fixed, what this means for your business — this one is way bigger than most people are seeing.
Let me walk you through every change that actually matters.
Headline Change in the OpenClaw 4.20 Update: Kimi K2.6 Is Now Inside
Before 4.20, if you used Moonshot's Kimi model, you were on Kimi K2.5.
Now when you set up Kimi inside your OpenClaw, it uses Kimi K2.6 by default.
K2.5 is still there if you need it.
But K2.6 is the new one.
Why that matters
Kimi is open-source.
It's fast.
It's cheap to run.
And K2.6 is much better than K2.5 at thinking, replying, and handling tools — which is exactly what you need for agentic workflow.
If you're running your OpenClaw agent on Kimi, your agent just got smarter.
You didn't have to do anything.
Where K2.6 kicks in
Web search is included inside OpenClaw and uses Kimi K2.6 by default when you pick up the bundled Kimi setup.
Media understanding also uses it.
Token cost tracker knows Kimi K2.6 pricing too — so you can see what your tasks cost before you actually run them.
Thinking mode you can leave on
There's one more thing on K2.6.
You can actually keep thinking mode on all the time.
They call it "thinking keep all".
That means the model thinks before every reply.
It doesn't just guess.
It works out the answer first.
That matters because Kimi is one of the cheapest models you can run through OpenClaw.
And now it thinks through each reply like the bigger expensive models do.
So agents cost less to run — but give better answers.
🔥 Want to run Kimi K2.6 inside OpenClaw with the right setup? Inside the AI Profit Boardroom, I've got a full 30-day OpenClaw roadmap showing you exactly how to install it, wire it, and set up a communication system for follow-ups that happen without you. Plus four weekly coaching calls where you bring your setup and we fix it live. 2,800 business owners already inside. → Get access to the full roadmap
The /think Command Got Smarter
Before 4.20, if you told a model to think but that model couldn't think, stuff would break right in the middle of a task.
Not fun.
Now OpenClaw checks first.
It knows what each model can handle.
So no more random errors mid-job.
And when you turn thinking off, it stays off.
Before 4.20, some models would quietly keep thinking anyway.
That's gone.
Your agent is more reliable now.
It doesn't break halfway through a job.
The Biggest Fix in the OpenClaw 4.20 Update: iMessage Agents That Actually Send
This is the one I was most excited about.
If you've got a Mac on the newest macOS and you tried running an AI agent through iMessage using Blue Bubbles before — you know the pain.
Messages wouldn't send.
Or they'd take a full minute and just break.
You'd hit send and sometimes the message would vanish into nothing.
What the OpenClaw 4.20 update fixed
Three things.
One — they pushed the send time limit from 10 seconds to 30 seconds.
Two — they made the system use a private path on new macOS so text actually goes out.
Three — they removed the weird error that was breaking plain text on TAH.
What this opens up
Today, you can run an AI agent from your Mac that texts your customers over iMessage like a real person.
And messages actually land.
Think about what that means.
Local shop owner gets a text at 10pm about opening hours — agent replies straight away.
Coach gets a DM from a lead on the weekend — agent can book them into a call.
An agent can just handle it.
Tapbacks got fixed too
Those little iMessage reactions — the heart, thumbs up, that sort of thing.
Before, if your agent tried to react with any other emoji, the whole reaction would fail.
Now OpenClaw falls back to whatever's related.
Something always shows up.
Your customer still feels seen.
iMessage over SMS when both exist
If someone has both iMessage and SMS with the same number, OpenClaw now picks iMessage first.
Before it sometimes sent an SMS when the blue bubble was working fine.
Now it just does the right thing every time.
Cron Jobs Rebuilt in the OpenClaw 4.20 Update
Nobody's talking about this one.
But it's huge.
"Cron jobs" is just a word for scheduled tasks — the thing that lets OpenClaw do work on a timer without you touching it.
In the OpenClaw 4.20 update, this got rebuilt.
What was broken before
Jobs would say "delivered" but silently keep skipping sending.
Jobs set to no delivery would throw false errors.
Jobs that ran every hour would pile up and eat your memory.
All fixed.
What's new
The runtime state got split into its own file so your job list stays clean.
And the big one — they fixed main session delivery.
That means when a scheduled job wakes up and wants to reply in a chat, it lands in the right chat.
If you set up a job that posts to multiple channels, it checks the setup when you save it.
So you can't accidentally build a broken job that fails at 3am while you're asleep.
Real-world example
Say you run a local gym.
You want an agent to text every new sign-up on day 3 with a pep talk and again on day 7 asking how the first week went.
With the new cron in 4.20, you set it up once — it runs forever.
If something breaks, you get a real error.
Not fake success.
Same for coaches.
Client signs up on day one — welcome message.
Day three — check-in.
Day seven — review call booking on autopilot.
While you focus on the actual coaching.
Security Locked Down in the OpenClaw 4.20 Update
OpenClaw 4.20 blocks a bunch of sneaky attacks.
SSRF guard
All that means is a bad request can't trick your server into doing something silly.
Workspace ENV injection blocked
If someone sends a weird file into your project folder, it can't silently change your setup.
Agent can't rewrite its own config
This one's important.
Before 4.20, the AI agent could technically rewrite parts of its own config.
It could change its own permissions or the tools it has access to.
Pretty scary.
In 4.20, the model can't touch sandbox settings, trust rules, or MCP server configs.
Even if it tries to sneak an edit through, it gets blocked.
Why this matters
If you're running an agent that talks to customers or handles files, one bad input could let someone mess with your setup.
Now it fails safe.
Device pairing tightened
If you connect a phone or a laptop to your OpenClaw setup, that device only sees its own pairing info.
It can't peek at other devices or approve them.
And if it fails to connect, you get a real error telling you why.
No more useless "auth failed" messages.
🔥 Want a safe production setup for OpenClaw agents? The AI Profit Boardroom has the full OpenClaw security and deployment playbook. Plus our member map so you can find OpenClaw users near you. Someone is always online. → Join the Boardroom here
Memory and Cost Tracking in the OpenClaw 4.20 Update
This is where things were getting messy before.
The memory pile-up bug
If your cron jobs and sessions ran a lot, old sessions would pile up.
Eventually your gateway would run out of memory and crash.
In 4.20, there's a built-in cap and an age-based cleanup.
Old sessions get pruned automatically.
So your gateway stays healthy even if you run hundreds of jobs per day.
Cost tracking is honest now
Before, if a session got saved multiple times, the costs would get counted twice — or 10 times.
Your numbers were basically wrong.
Now they snapshot the costs properly.
When you check what your agent spent last month, the number is actually real.
/new and /reset actually work
Before, when you started a fresh session, the same old model and provider choices would stick around.
You'd think you were on Kimi K2.6 but you'd be stuck on some fallback from three mistakes ago.
Now those commands clear out the stale stuff while keeping the choices you made on purpose.
Sounds small.
It isn't.
If you run one agent for customer replies and another for content, you want each session to start clean.
Now it does.
Personality Files Work Better in the OpenClaw 4.20 Update
OpenClaw has two files called SOUL.md and IDENTITY.md.
These are where you tell your agent who it is — its tone, its style, its quirks.
Before, the newer OpenAI models like GPT-5 and Codex were kind of stiff about these files.
They'd read them.
But replies still felt like a corporate support robot.
In 4.20, they made GPT-5 and Codex pick up that personality way better.
Your agent actually sounds like the character you wrote.
The OpenClaw team said it should feel "more like your weirdly capable little friend and less like a polished customer support automation".
Those are their exact words.
Why it matters
If your agent handles DMs, messages, texts, or leads — it has to sound like you.
Not a generic chatbot.
This update makes that way easier on GPT-5 and Codex.
More Channel Updates in the OpenClaw 4.20 Update
A few more changes worth calling out.
Auto-reply got smarter
Before it was one-size-fits-all.
Now it behaves differently based on where the message is.
In a direct chat, it writes a helpful reply.
In a group, it can stay quiet unless someone tags it.
Your agent doesn't spam group chats but still helps people who message you one-on-one.
Setup wizard cleaned up
The first-time setup used to feel like a sketchy terminal thing.
Now it's a proper walkthrough with a loading spinner.
You're not just staring at a blank screen.
Clear headings, clear warnings, anyone can follow it.
Telegram agents
Polling used to stall out sometimes.
Now it has a longer grace period before OpenClaw thinks the connection is broken.
Stateless reactions work better too.
If your agent replies and you set it to remove the thinking emoji after — it actually does that now.
Matrix fix
You can allow-list and change allow-list settings without restarting the whole channel.
You can add or remove who can message your agent on the fly.
Real pain point before.
Manus live preview
Your agent now shows a live draft preview as it writes.
People in the chat see the thinking happening.
Then the final reply lands in place.
Feels way more natural.
Discord fixes
/think only shows options that actually work for your model.
No fake choices.
And if a channel has missing info like slash commands, they don't crash anymore.
Video notes + links to the tools 👉 here
What OpenClaw 4.20 Actually Means for Your Business
Let me zoom out.
Five things from this update matter for business owners:
1. Kimi K2.6 is now inside OpenClaw
You have a smarter, cheaper model out of the box.
Thinking mode makes it more accurate for the same cost.
2. iMessage agents are real now
Huge channel for business owners who get texts from customers, especially locals.
3. Scheduled tasks finally work
You can trust your agent to do work on a timer and actually deliver.
No silent failures.
4. Security is locked down
You can hand the agent to a team and not worry about them accidentally breaking the AI — or the AI breaking itself.
5. Personality picks up on GPT-5 and Codex
Your agent sounds like you, not a generic chatbot.
If you want to see how this pairs with other agent stacks, check out my Hermes Agent Workspace breakdown.
Caveats: Be Honest About the Rough Edges
Most AI agent tools are pretty rough.
They break.
They hallucinate.
They run up weird numbers.
OpenClaw 4.20 is still going to be buggy in some places.
I've got to admit that.
When I've checked out Hermes, it still feels a lot smoother than OpenClaw in some spots.
But I think both tools are getting better.
For me personally, I switch between them.
I like to try everything, then decide which one is the best for the specific job.
Sometimes if you update your setup, that can break things and you have to restart the whole thing.
Just be aware of that.
I'd recommend setting up an OpenClaw backup first so you can restore it later if you have any problems.
If you're deciding between the two, my Hermes vs OpenClaw post covers the real trade-offs.
How to Get the OpenClaw 4.20 Update
Simple.
Just type update inside your OpenClaw and it should update to the latest version.
Takes under a minute on most setups.
OpenClaw 4.20 Update FAQ
What's the biggest change in the OpenClaw 4.20 update?
Kimi K2.6 is now the default Kimi model inside OpenClaw, and iMessage agents via Blue Bubbles finally work.
Those two together change what you can build.
Does the OpenClaw 4.20 update cost extra?
No — same pricing, same licence.
Can I still use Kimi K2.5 after the OpenClaw 4.20 update?
Yes, K2.5 is still available if you need it.
K2.6 is just the new default.
Will the OpenClaw 4.20 update break my existing setup?
Sometimes an update can break custom tweaks and you have to restart.
Back up first.
How do I install the OpenClaw 4.20 update?
Type update inside OpenClaw.
What if I run into bugs on the OpenClaw 4.20 update?
Restore your backup and wait for the next patch.
Or drop into the community and ask — someone's always online.
Related Reading
Learn how I make these videos 👉 here
Get a FREE AI Course + Community + 1,000 AI Agents 👉 here
🔥 If you're on the fence about automating your business with AI, this is the window. Tools like OpenClaw are getting easier every month. The people who learn this in 2026 will run circles around competitors. Inside the AI Profit Boardroom, you get step-by-step videos on your first OpenClaw agent, how to use cron jobs for follow-ups, how to write your SOUL.md so your agent sounds like you. Four coaching calls a week. 2,800 members. Prompt library. 30-day roadmap. Member map. → Get access here
That's the OpenClaw 4.20 update — Kimi K2.6 inside OpenClaw, smarter thinking, iMessage agents that actually work, cron jobs that deliver, tight security, better personality on GPT-5 and Codex.