Meta's Claude Code Killer is HERE!

Meta just launched MuseCode, their first ever AI coding agent, plus a brand new model called MuseSpark 1.2 to run inside it. And everyone's asking the same question, which is, is this a Claude code killer? Now I've read the full release. I've been through the benchmarks and we've already plugged it into our agent operating system and tested it.

So in this video, I'll give you the honest answer. I'll show you what the benchmarks actually say because Meta is being surprisingly upfront about what the model is and isn't. And I'll show you the three features that make MuseCode genuinely different from every other AI agent out there. I'll tell you what Meta is really planning here because once you see it, the whole release makes a lot of sense.

Plus near the end, I'll show you the one test we ran where Muse actually beat Coen318 because that result tells you exactly how to use this thing. So first, a quick version of what just dropped two things at once. So MuseSpark 1.2, that's the model, the brain, and then you've got MuseCode, and that's the agent built around it, the harness. And you can install it with one command inside your terminal, log in through your browser and you're up and running.

And if you can copy and paste one line, you can install this. Meta also expanded access at the same time. So you can run MuseSpark through MuseCode, through their model API or through OpenRouter. And that one matters a lot because OpenRouter is how a lot of us plug new models into our own systems the day they drop without setting up another account.

So if a terminal agent from Meta sounds familiar, here's the pattern. Anthropic has Cloud Code, OpenAI has their command line agent, Google has theirs, and now Meta has MuseCode. Every major lab is shipping the same kind of product right now, which is an agent that lives in your terminal, runs on their model and does long jobs on its own. And this is where the whole industry has landed.

The command line turned out to be the best home for AI agents because that's where an AI can actually touch files, run programs and finish real work. Instead of just chatting about it, I call what's happening, the harness war. And once you see it, you can't unsee it. For years, every AI lab fought one battle.

Who has the smartest model? That race is brutal and expensive and the lead changes hands every few months. So a second race is now opened up. The race to own the harness, the system your agents live inside, the thing that takes your goal, splits it into jobs, runs to work and brings it back.

Because here's what the labs have figured out. Brains get swapped all the time. A better model drops, people move on, but harnesses are sticky. And once your whole workflow lives inside one system, you stay even when a shiner brain comes out.

You just want the option to swap the brain and keep the system. MuseCode is Meta officially entering the war. So let's judge it properly. Brain first, the harness afterwards.

So they deserve very different scores. Start with the brain. MuseSpark 1.2 has some interesting benchmarks and there's something in Meta's own charts to tell you exactly what this model is. If you look at who they compared it against, it wasn't against Fable 5.

It wasn't against the biggest flagship models. Meta benchmarked MuseSpark against models like GPT 5.6 Terra from OpenAI, the smaller, faster, cheaper class of models. And that's not there. That Meta is telling you the truth about what they built.

MuseSpark isn't trying to be a frontier model. It's built for smaller weight class. The class where the fight is about speed and cost, not raw brain power. So Meta describes it as a moderate improvement over MuseSpark 1.1 tuned for the jobs coding agents get handed the most.

So long debugging sessions, multi-file changes, tasks that run way past a single prompt. And the pricing backs up the positioning. On the standard API output, tokens cost $4.25 per million with cashed input at 15 cents. That's cheap worker pricing.

That's a model price to run all day. Not a model price like a genius you can sell twice a week. And three things about the brain are worth knowing. One, it was trained inside the harness.

So MuseCode was in the training loop from day one. So when the model calls a tool or executes a plan inside MuseCode, it tends to work because the model grew up doing exactly that. Meta says it still generalizes to other coding agents too. So you can run it inside, for example, Cloud Code and other coding setups you already use.

And number two, it's built for long jobs. So it's trained on whole repository work, uses planning and context compaction to hold direction of hours and can fire off tool calls in parallel. So work keeps moving whilst other results are still pending. And then three, the context window is 1 million tokens.

Plain English, it can hold a whole project in its head at once, thousands of files, all the history without forgetting what it read at the start. For a model whose job is grinding through big piles of material, that's exactly the spec that you want. There's also an effort dial. So MuseSpark is a reasoning model, meaning it thinks before it answers.

And that thinking costs tokens. The effort command turns the thinking up or down. Simple job, dial it down, runs fast and cheap. Hard job, dial up.

You decide how much brain power each task get. So you never pay deep thinking prices for grunt work. Now does the brain hold up in practice? We tested it.

And honestly it wasn't actually that good. We ran it on real builds and compared it against GPT 5.6 against Fable 5, Quen 3.8 and MuseSpark pretty much lost every single head to head. Builds came out slow, laggy, missing detail. Quen 3.8 bit it pretty much every single time.

But here's a fair way to read that. We were testing it against frontier models and Mello never claimed that it was judged against its actual weight class in the benchmarks. So models like Terra are more competitive with it. So would you use this as your main builder?

No, absolutely not. You wouldn't use it for that. And I'll show you in a minute where it actually did impress us because there is one spot and that's the brain. So now let's talk about the harness.

And this is where the release gets genuinely interesting because Muse code has three features that are different from what's out there. And each one fixes a real problem that people hit with agents every single day. Feature one is fan out with automatic isolation. So you can give Muse code one big job and it splits the work across multiple sub-agents automatically.

In Meta's own demo, they handed it a game project with six things to fix and at one prompt, the parent agent spawned six workers, one per task running in parallel. And here's the part that's new. Each worker gets his own isolated copy of the project automatically. So in plain English, if you picture one document and six helpers, instead of all six scribbling on the same page and wrecking each other's work, every helper gets their own photocopy.

They make changes on their own copy. Then you review each one and you keep what you like. Your original never gets touched until you say so. And you set up none of this because Muse code automatically delegates to subject agents automatically on its own.

So why should you care if you're not a coder because agents stepping on each of its files is the number one reason multi-agent setups usually fall apart. Two agents touch the same folder. They overwrite each other and your work is gone or it's wrecked or it's broken in one single piece of code. So Meta theoretically have solved that at the system level made it the default.

Every agent platform is going to want to do this. Now obviously you can do this with Cloud Code already, but it is a useful feature. And a lot of people ask me about orchestrating sub-agents. So that's why I mentioned it and why you can do it.

Now, quick pause here. If you're running a team of agents and that sounds powerful, but you don't know where you'd start, this is exactly what we've built for you inside the AirProfit boardroom. The day Muse code dropped, which is today, we plugged it into our agent operating system, the dashboard where all your agents run in one place. So your Cloud Code, Hermes, OpenCore, any other agents you use and now Muse code as well.

They're all connected. They're all able to pass work to each other. You get the full agent operating zip file, a video tutorial covering the Muse code setup and a 30 day roadmap for putting agents to work in your business. Daily step-by-step tutorials.

And right now those cover delegating jobs to Muse as a sub-agent. The exact workflow I'll show you shortly for coaching calls every week, where you can share your screen and we fix your agent set up live and over 3,800 business owners inside plenty already running multi-agent stacks. And some had never touched AI before joining. Link in the comments description or go to the airprofitboardroom.com feature.

Number two, the event log inside Muse code. And this one answers the question you're already thinking, which is if six agents are running at once, how do I know what they actually did? Because Muse code logs everything. So every agent it spawns, every action, every steer, every cancel, every decision, all of it lands in a plain text file on your own desk that you can search and replay.

Think of it like a diary. The agents can't lie inside. So if something looks wrong, you don't just guess, you look it up and see exactly which worker did what and when. Meta built the whole agent around this idea.

They call it transparent. Bility, sorry. And it's clearly aimed at the biggest fear people have around agent swarms, handing work to a team you can't actually see. And that same log powers a command called resume.

So if your session crashes, your laptop dies, you close the lid and go to bed. Next session, Muse code reads the log and picks up from the last recorded step. Again, you can do this sort of stuff inside core code, but I just want to talk you through the features inside Muse code as well. So you understand it properly.

You don't have to re-explain anything. It already knows where it was. And if you've ever lost an hour of agent work to a crash, you know exactly why that matters. And feature number three is bundled skills.

So Muse code ships with a small set of built in playbooks. And two of them are ideas I really like. So one is called taste. It's an anti-slop filter, a checklist of visual defaults.

The agent is banned from using. So what it builds stops looking machine made. I'm sure you already know what that means, right? So the same layout on every AI built tool we've all been there.

And this skill exists purely to kill that. The other one is called grilling and this means you can basically hand it rough idea. And instead of running off to build it, it interviews you one hard question at a time until the idea holds up. I think this is great because a lot of people have shiny object syndrome.

They struggle to explain exactly what they want in clear terms. And this just helps you break it down in a way where the agent does a heavy lifting for you. It matters on demo. The planning skill read the real project files and caught a rule mistake before a single line was written.

An agent that pushes back before it builds and ask your questions before wasting an hour going the wrong way. And that's the difference between the tool and an actual system. And all four of these skills only fire when you call them. So they don't just jump in uninvited and the screen shows you exactly which turn a skill was loaded on.

So you always know when a playbook is shaping the work. So that's the update. Weaker brain than the frontier by design, genuinely strong harness. Now, what is Meta actually planning here because this release tells you a lot if you read between the lines.

So Meta knows that they probably can't win the frontier brain race right now, maybe in the future, but their answer is to win somewhere else. So own the harness, make the workflow sticky and improve the model with the people using it. And that last part is built right into the pricing. The default tier is called the contributor tier.

It's rate limited by tokens in a rolling five hour window instead of a request count. And that means you can run it all day. And Meta says usage on that tier may be used to improve their products. So just be aware that the deal is generous limits in exchange for usage and data helping train the next version.

Some people are comfortable with that. Some people are not. You pick whichever one you prefer. The fastest way to build a great coding agent is with the people use every day.

So that's their plan, I guess. Ship the harness, get it into daily use and let real world usage make Muse Spark 1.3 and 1.4 better. Now, one honest warning that comes with that. If you're handling private client data material, don't run it on the contributor tier, right?

Swap to the standard model through the model API, or you can just use open router where it's no more pay as you go. And Meta is even starting to accept zero data retention requests. So you want to know which lane you're in before you feed it anything. Now here's a test result I promised because it shows you exactly where the model fits today.

We built Muse code into our agent operating system and ran Muse as a sub-agent inside Hermes. We gave it a delegated task, a learn task, and a read this guide sort of pull out what matters and turn it into a skill type of task where the agent can reuse that skill that it saves for later. We did this with Hermes and then we ran the exact same task with Quern 3.8 side by side, same prompt. Muse actually replied faster, faster than Quern 3.8 and the skill it built was genuinely decent in depth, well-organized.

It pulled the right information from a guide, built a clean table and got specific. So this was all whilst Quern 3.8 was still sitting there thinking. So the model that loses head to head builds one, the delegated task. And that's the whole story of Muse Spark in one result.

It's a worker, it's the fast cheap teammate you can hand one clear job to whilst your smart model does the heavy thinking. For example, your main agent plans your content week whilst a Muse worker might read a long industry report and hand back a one page summary. Small job, but it's done fast and it's done cheap. Now let's zoom out because the old way versus a new way here matters more than this one single release.

The old way of judging an AI was that you drop one question and you ask yourself, okay, is this the smartest model? But if you look at the way agents are going, you don't really need to rely on frontier models. You don't need frontier models for everything. So the new way asks two questions, which is how good is the harness and what job is this brain right for?

So under those rules, this release is actually worth your attention because Muse code is one of the best designed agent harnesses anyone has shipped. Parallel workers with automatic isolation, a full playable and replayable log, crash recovery, playbooks are pushed back before building and Muse Spark run as a worker instead of a lead builder. It's fast, it's cheap, it holds a million tokens and it beat Quen 3.8 on response speed in our delegation test. So is it a Claude code killer?

No, Claude code still has the far stronger brain behind it. And for real building work, that gap is huge, right? But Muse code is a serious harness. It brought a few genuinely new ideas to the table and the automatic work isolation is pretty useful.

Now you could do that with ultra code, but obviously that can get quite expensive with Claude or you could use, for example, Google's managed agent API, but again that can get quite expensive and it's very, very slow to use. So they're building the system now and betting the brain catches up through exactly the usage deal I described. Watch what Muse Spark 1.3 looks like after months of contributed data. I think that's a real play.

And here's what this means for you. The people winning with AI right now aren't waiting for one perfect model. They're building a system where models are parts, a smart cheap brain for hard thinking, cheap, fast workers like Muse for the grunt work and a harness connecting all of them together. Now when a new model drops as well, they plug it in the same day just like we did whilst everyone else is still reading reviews.

And I know what some of you are thinking, terminals, sub-agents, command lines. I'm not a coder. This isn't my world, but here's the truth. The install is one line you copy and paste.

The login is literally a browser click. From there you talk to the agent in plain English. You describe the job, it does the work. And the members getting the most from our agent community and just agents in general aren't engineers.

They're agency owners, they're freelancers, they're e-commerce sellers who learned one skill, which is describe a job clearly. And some of you also thinking by the time I learn this tool, a new one drops and I'm behind again. But that's exactly why you want to learn the harness instead of the tool. The harness doesn't change that much, right?

Cloud code doesn't change that much. Muse code probably won't change that much at all. But if you learn the system once every new release becomes a free upgrade instead of a fresh headache. And everything from this video is already sitting inside the airport boarding ready for you.

So Muse code is plugged into our agentic operating system. So you can run Muse Spark as an agent next to your cloud code, Hermes, OpenCore and every other agent. And you get the agent operating system file, the video walkthrough for this and daily updates every time we improve the system. So when Muse Spark 1.3 lands, your setup is already current.

The 30 day roadmap takes you from step by step from literally nothing to working agents, handling real jobs in your business, saving you time and bringing in leads. And you get four weekly coaching calls where you can ask about your Muse or multi-agent setup live and get it fixed on screen. Plus you get a prompt library covering agent delegation and member maps. So you can meet people near you running these exact setups and over 3,800 business owners inside link in the comments, description or go to the airportforboarding.com to get access.

Thanks for watching. I'll see you in the next one.