Fello AI's four-way test puts Meta's cheap Muse Spark near ChatGPT, but behind Claude

What the comparison found

As Fello AI reported in its four-way comparison, Meta's Muse Spark is now close enough to the big paid assistants to be worth a look. Muse Spark 1.3, Meta's cheaper model line, scores 48.1 on the Artificial Analysis Intelligence Index, an independent score that blends many tests into one number. That edges out GPT-6 Sol, OpenAI's mid-priced model, at 47.6.

It still sits well below the top. Claude Opus 5.5 leads the index at 57.6. Claude Sonnet 5.5 follows at 56.0, then Claude Fable 5.1 at 53.4, OpenAI's GPT-6 Astra at 52.7 and Google's Gemini 4 Argon at 52.6.

Google's newest Flash model, the lighter and cheaper Gemini 3.8 Flash, scores 40.9. Google's newest Pro model is still Gemini 3.1 Pro, at 29.7.

Meta's bigger models close the gap

Meta added a larger family called Muse Wonder on October 12, 2026. Muse Wonder Full scores 55.8 and plain Muse Wonder scores 52.4, which puts both above Muse Spark 1.3 and in the same range as the leading rivals.

There is a catch in where you can use them. Muse Wonder Full runs only inside Muse Code, Meta's terminal agent (a tool that carries out programming tasks from a command line). Muse Wonder is priced at $4 / $20 per million tokens on the API, the same as Claude Opus 5.5.

Muse Spark 1.3 remains Meta's cheap option, at $1.25 / $4.25 per million tokens. A token is a small chunk of text, and API pricing is what developers pay to use a model in their own software.

Coding is still where Meta trails

The sharpest gap shows up in coding. On Terminal-Bench 4.0, which tests how well a model handles programming jobs on its own, Muse Spark 1.3 scores 33.3%. Claude Sonnet 5.5 leads at 63.6%, Claude Opus 5.5 gets 59.6% and GPT-6 Astra gets 59.1%.

Cost complicates the picture. Artificial Analysis spent $3.26 per task running Astra and $5.98 running Opus 5.5. That means Astra reaches nearly the same coding score for about half the money, though Opus 5.5 is the stronger model overall.

Meta's own launch chart shows far higher coding numbers for 1.3. Those are the company's figures, measured against a rival model Anthropic has since replaced, so treat the independent result as the safer guide.

Where Muse Spark does well

Fello AI's strongest praise for Muse Spark is on health and images. On HealthBench Hard, an early Muse Spark scored 42.8, against 40.1 for GPT-5.4 and 20.6 for Gemini 3.1 Pro. The comparison also credits it with strong chart and figure reading.

The writers backed this with a hands-on test from April 2026, using a photo of an instant ramen nutrition label. Claude won, mainly for following instructions and giving a memorable verdict. Muse Spark and Gemini, though, both noticed details the other two missed: the noodles are likely deep-fried, and the packaging carries a Vegan Society logo. ChatGPT came last in that test.

Keep the age of that result in mind. It used the models available in April, not today's.

The free-app confusion

The most practical warning in the piece is about access. Meta's best Muse Spark numbers come from the paid API and Muse Code. It also put the model behind Muse, a personal agent that is free to start.

What Meta has never said is which version the free Meta AI app runs. So a benchmark score for Muse Spark 1.3 may not describe what you get when you open the app.

The other vendors have shifted too. Gemini 4 Argon has been open to Google AI Pro and Ultra subscribers since October 13. Free ChatGPT moved to GPT-6 Luna from October 8.