Yuval Avidani
Author
"Meta, the company that gave away models for free for years, just asked me for a credit card." That's how I'd sum up this week in one sentence. I'm talking about Muse Spark 1.1, Meta's new model, which I tested myself through the Meta AI app on mobile - and my first impression was simply wow. Insanely fast, Hebrew that sounds natural, and answers that flow almost faster than I can read them. But as always in our newsroom, after the "wow" come the numbers - and that's exactly where the story gets interesting.
The "free" Meta started charging money
For years Meta was the open-source player - the Llama series was shared with open weights, for free, and any developer could download it and run it on their own server. Something changed. On July 9, 2026, Meta launched Muse Spark 1.1, and this is the first time ever that Meta is charging businesses for access to its model. The price: $1.25 per million input tokens, $4.25 per million output tokens.
Let's pause on one word: token. A token is a small piece of text - roughly three to four characters in English, sometimes a single letter in Hebrew - that the model reads and writes unit by unit. Think of it like the bricks that make up every sentence. When talking about a model's price, you're counting how many bricks went in and how many came out.
It's important to understand that this isn't just a chatbot. It's a multimodal model (understanding text, images, video and audio together) with an emphasis on "agentic" capability. An agent is a model that doesn't just answer, but plans steps, activates tools, and writes and runs code for us - like an employee who receives a task and carries it out from start to finish, not just an advisor giving advice. This family is developed by Meta's superintelligence lab, Meta Superintelligence Labs, led by Alexandr Wang. The first version, Muse Spark, came out back in April 2026, and the next generation is already in training under the codename Watermelon.
What does "fast" even mean? There are two different numbers here
When I said "insanely fast," I meant the feeling. But it turns out a model's "speed" isn't one number but two, and it's worth separating them:
- Tokens per second (TPS) - how many words the model fires at us after it's already started talking. This is the "speaking pace."
- Time to first token (TTFT) - how many seconds we wait in silence until the first word appears. This is the "thinking time" before it opens its mouth.
The fun part is that a model can be excellent at one and bad at the other. According to Artificial Analysis (a leading independent measurement body in the field), Muse Spark 1.1 fires output at a pace of about 114 tokens per second - above the median of its category (69.8), meaning it's genuinely fast. But its TTFT is about 21 seconds, because it's a "thinking" model that runs a full chain-of-thought silently before answering. In the app, which apparently runs a lower thinking effort, this wait is barely noticeable - which is why my impression was one of continuous speed.
We measured it: Muse Spark 1.1 versus the leaders
Here's the comparison you asked for - the output pace of the leading models, all from Artificial Analysis measurements, all on each company's official API. Note: Muse Spark 1.1 is fast, but it's not the fastest on the field - GPT-5.6 Terra beats it with 140 tokens per second. And when GPT-5.6 Sol runs on special Cerebras hardware, it even reaches about 750 tokens per second. So "the fastest I've felt" is true to the sensation, but on the chart it's a respectable middle spot, not first place.
Where Muse Spark 1.1 truly shines is in the ratio between speed, price and intelligence. On the Artificial Analysis Intelligence Index it scores 51 - lower than Fable 5 (60), Opus 4.8 (56) or Sonnet 5 (53), but higher than all Google models. And its price is significantly lower than the entire group. Think of it like a car: not the fastest and not the most luxurious, but the ratio between performance and price is among the best on the market right now.
Support for Israel and Hebrew
This is the part that concerns us directly. In June 2026 Meta officially launched Meta AI in Israel with full Hebrew support, right inside WhatsApp and on the meta.ai website. You can talk to it in natural Hebrew, generate and edit images from text, and upload a photo for analysis - without leaving the app we all already use daily.
And indeed, in Hebrew it feels like the best I've encountered so far. It's important to remember why this is hard: Hebrew "costs" models roughly 30 to 45 percent more tokens than English for the same sentence, because the tokenizer breaks Hebrew words into smaller pieces. A model that feels smooth and fast in Hebrew did a good job both in the breakdown and in the pace. In my view, this is one of the reasons the impression is so positive.
Hallucinations: when the feeling meets the number
I noticed while using it that the model sometimes makes things up - giving facts that sound confident but are wrong. This is called a hallucination: a situation where the model fills a knowledge gap with a guess that sounds convincing, like a student who doesn't know the answer but responds with full confidence. And here's the beauty of verification - Artificial Analysis measured Muse Spark 1.1's hallucination rate at 38 percent, among the highest in this generation. In other words, my impression wasn't a coincidence, it was measurable.
This also connects to its overall profile: it actually leads in agent and tool benchmarks, but lags behind Opus 4.8 and GPT in pure coding benchmarks (on SWE-Bench Pro it scores 61.5 versus Opus 4.8's 69.2). It's also "chatty" - consuming more tokens than average to reach an answer, which in practice raises the cost and lengthens response time.
My take
In my view, Muse Spark 1.1 is Meta's first serious entry into the big league, and it's impressive: fast, cheap, multimodal, and simply pleasant to work with in Hebrew. But it should be said honestly - it's not the king. It's fast but not the fastest, smart but not the smartest, and it fabricates facts 38 percent of the time, so every important answer needs verification. Also - as of now there's still no open access to the API (it's on the Meta Model API waitlist), and in the app I didn't see connections to external tools and skills like competitors have.
Bottom line: Meta has proven it can build a fast, cheap model that feels excellent, especially in Hebrew. The real question isn't whether it's impressive - it is - but whether speed and low price beat accuracy when you really need to trust the answer. When a model answers you quickly and smoothly, do you fact-check it, or do you believe it?