Meta quietly released its strongest AI model yet on Tuesday, and almost nobody outside the developer world noticed. Muse Spark 1.3 matches the biggest models from OpenAI and Anthropic on independent benchmarks, while costing less than half as much per task. That matters even if you’ve never touched an API, because it changes what “free AI” can do.
What is Muse Spark 1.3?
Muse Spark is Meta’s flagship AI model, the same family that powers parts of Facebook, Instagram, and the company’s own coding assistant. Version 1.3 is the biggest jump Meta has made since the line launched in April, according to Mark Zuckerberg, who announced the release on X with a post that got nearly half a million views.
The headline: on the Artificial Analysis Intelligence Index, an independent ranking that scores how smart models actually are, Muse Spark 1.3 scored 61. That puts it in a dead heat with GPT-5.6 Sol at max settings, Grok 4.6, and Claude Opus 5. Two months ago, Muse Spark 1.2 could only manage a 57. That’s a big move for any model, let alone one from a company that was mostly an also-ran in the frontier race a year ago.
So what changed under the hood? Meta says the new version is better at long, complicated tasks, especially coding and agent work. It also got a personality upgrade: the model now asks clarifying questions when your prompt is vague, asks for help when it’s stuck, and confirms before doing anything consequential. That sounds small, but it’s actually the fix everyone’s been begging for. Most models just barrel ahead and burn tokens on a wrong path. This one stops and checks.
How good is it actually?
Benchmarks from the company that built the model are always a little suspect, so let’s look at the independent numbers first. The Artificial Analysis index isn’t run by Meta, and it puts Muse Spark 1.3 right next to the most expensive models on the market:
| Model (reasoning mode) | Intelligence score | Cost per task |
|---|---|---|
| Muse Spark 1.3 (xhigh) | 61 | $0.55 |
| GPT-5.6 Sol (max) | 61 | $0.95 |
| Grok 4.6 (high) | 61 | $0.94 |
| Claude Opus 5 (high) | 61 | not disclosed |
Same intelligence, roughly half the price. On Meta’s own 11-test benchmark card, the new model wins or ties OpenAI and Anthropic on 5 of the 11 rows, including Terminal-Bench 2.1, a coding test where it matched GPT-5.6 Sol at 88.8 and beat Claude Opus 5’s 86.7.
It’s not perfect everywhere. Claude Opus 5 still leads on a few agentic tests like OSWorld and JobBench, and one independent metric (AA-LCR) actually slipped from 83% to 79%. Meta’s numbers are also self-reported, so take the wins with a grain of salt. But the direction is unmistakable: Meta is now genuinely competing at the top.
The price is the real story
Here’s the part that should make you sit up. Muse Spark 1.3 costs $1.25 per million input tokens and $4.25 per million output tokens through the API. Compare that to what OpenAI and Anthropic charge for their top models, which run at multiples of that rate.
And then there’s the contributor tier. If you let Meta use your prompts to improve the model, prices drop to $0.10 per million input tokens and $0.20 per million output tokens. That’s a 90% discount for what most hobbyists and solo founders are doing anyway. Zuckerberg called the pricing “almost too cheap to meter,” and for once the hype is close to the mark.
Why does the price matter to a beginner? Because the cost of AI gets passed down. When a frontier model gets cheap, the tools built on top of it get cheap too. That’s how we got free tiers of tools that used to charge $20 a month, and it’s why your next AI tool will probably cost less than the last one.
How to try Muse Spark 1.3
You don’t need to be a developer to poke at this thing. Right now the model is live in two places:
- Muse Code, Meta’s coding assistant, which you can think of as their answer to Cursor or Claude Code
- The Meta Model API, if you want to build something with it
Bloomberg reported the update will also roll out to Meta AI and to users of Instagram and Facebook soon, so you’ll probably meet it in a chat window before the year is out without doing anything special.
For non-coders, the honest advice is: wait for the chat rollout, or play with it through Muse Code if you’re curious about AI-assisted coding. The model won’t replace your tools overnight, but it’s worth trying before you pay for another subscription.
Open weights are coming
Here’s the part that has the open-source community excited. Zuckerberg promised that open weights versions of Muse Spark are coming soon, along with a bigger model internally nicknamed “Watermelon.” Open weights means the actual model gets released for anyone to download, run, and build on.
If that happens, it’s a genuine shift. An open model with frontier-level performance would give people a real alternative to paid subscriptions, and it would let anyone with a decent computer run a top-tier model privately (Meta already has a history here, since its Muse Glimmer model runs on regular hardware). We covered what open-weight models like GLM 5.3 mean for normal users, and Muse Spark could make that story even bigger.
I’d hold off on celebrating until the weights actually drop, though. “Soon” is doing a lot of work in that sentence, and Meta has been burned by open-source releases before. Still, it’s the most promising sign in months.
Who should switch?
Let’s be direct about this. If you already pay for Cursor, Claude, or ChatGPT’s pricier plans and you mostly use AI for coding, Muse Spark 1.3 is absolutely worth a test drive, right up there with the other free coding AI options we’ve tested. The agentic coding improvements and the cheap contributor tier make it a legit alternative, not a compromise.
If you just use ChatGPT for writing, planning, or everyday questions, you probably won’t feel the difference for a while. The chat rollout hasn’t happened yet, and your current tools work fine. You’ll benefit in the long run through cheaper tools and more competition.
One warning: don’t switch your whole workflow on launch week. The model has a higher abstention rate than its predecessor, meaning it sometimes refuses to answer rather than guess. That’s safer, but it can be annoying mid-task.
Bottom line
Muse Spark 1.3 is the first time Meta has genuinely stared down OpenAI and Anthropic on raw capability, and it’s doing it at half the price. Independent benchmarks put it in the top three models in the world, and the open weights promise could make it a genuine free alternative in the coming months.
Try it in Muse Code if you code. Watch for the chat rollout if you don’t. And when open weights drop, that’s the moment this gets really interesting for everyone.
