Meta's Muse Spark 1.3 matches GPT-5.6 Sol for under a third of the price. The cheaper tier has a catch.
Muse Spark 1.3 scores 48 to Sol's 47 at $1.25 / $4.25 against $4 / $20, and runs three times faster. Meta's $0.10 tier trains on your data.
GTM Architect & Growth Operator · Now · 16 September 2026
TL;DR · Key insights
- On Artificial Analysis's Intelligence Index v4.3, Meta's Muse Spark 1.3 scores 48.17 and GPT-5.6 Sol scores 47.06. About a point apart, which I treat as level. On the Coding Agent Index they sit at 54 and 55.
- Muse Spark 1.3 lists at $1.25 / $4.25 per million tokens on Meta's Model API. Sol lists at $4 / $20, a promotional price. Meta is about 3x cheaper on input and 4.7x on output.
- It is also much faster: 225.9 output tokens per second against Sol's 64.7, and it cost Artificial Analysis $2,000 to run the full index against $3,465 for Sol.
- Meta also sells a contributor tier at $0.10 / $0.20, a further 12 to 21 times cheaper. Meta's own page says that tier is used to improve its products. The standard tier is not.
Fifth in the ranking Artificial Analysis published with its Intelligence Index v4.3 is not an Anthropic or OpenAI model. It is Meta’s Muse Spark 1.3, at 48.17, a point above GPT-5.6 Sol.
That alone would be worth a look. The price is what makes it a comparison.
| Info | Muse Spark 1.3 | GPT-5.6 Sol |
|---|---|---|
| Input / output per 1M | $1.25 / $4.25 | $4 / $20 |
| Cached input per 1M | $0.15 | $0.40 |
| Cheaper tier | $0.10 / $0.20, trains on your data | None |
| Intelligence Index v4.3 | 48.17 | 47.06 |
| Coding Agent Index | 54 (in Muse Code) | 55 (in Codex) |
| Output speed, tokens/s | 225.9 | 64.7 |
| Cost to run the full index | $2,000 | $3,465 |
| Context window | 1M | 1,050,000 |
Prices from Meta's Muse Spark model page and OpenAI's GPT-5.6 Sol model page. Index to two decimals from Artificial Analysis's v4.3 announcement; speed and run cost from its model pages; Coding Agent Index from its GPT-6 Astra note. Read on 16 September 2026.
Level on capability
On Intelligence Index v4.3 the two are 1.11 points apart. In the explainer on the index I said to treat anything under a point or so as a tie, and this is right on that edge. On the Coding Agent Index they swap. Sol scores 55 in Codex and Muse Spark 1.3 scores 54 in Meta’s own harness, Muse Code.
So I read them as level. Once capability is level the decision moves to everything else, and on everything else the two are far apart.
Not level on price or speed
Meta’s standard tier lists Muse Spark 1.3 at $1.25 input and $4.25 output per million tokens. Sol lists at $4 / $20. For the same capability, that is about 3.2 times cheaper on input and 4.7 times on output.
Artificial Analysis’s blended price lands in between. It assumes seven cached tokens and two fresh input tokens for every output token, and on that mix Meta costs $0.78 per million against Sol’s $3.08, about a quarter.
Actual spend points the same way, less steeply. Running the full index cost $2,000 on Muse Spark 1.3 and $3,465 on Sol.
It is also faster. Muse Spark 1.3 generates 225.9 tokens per second and Sol 64.7, so the cheaper model runs at three and a half times the rate.
Sol’s $4 / $20 is a promotional price, and OpenAI’s own discount implies $5 / $30 without it, at which point the gap would be 4 times on input and about 7 times on output.
The details of Sol’s promotional wording are in Sol against Sonnet 5.
The contributor tier
Meta sells a second model ID, muse-spark-1.3-contributor, at $0.10 per million input tokens, $0.002 cached and $0.20 output. That is another 12.5 times cheaper on input and 21 times on output than Meta’s own standard tier.
The difference between the tiers is one line on Meta’s model page. Inputs to the standard muse-spark-1.3 are marked as not used to improve Meta’s products, while inputs to the contributor tier are marked as used for exactly that. So part of the price is paid in data. Whether that is acceptable depends on what you send.
Who should switch
| Situation | Model | Why |
|---|---|---|
| New work, free to choose a vendor | Muse Spark 1.3 | Level capability at roughly a quarter of Sol's cost, and 3.5x the speed. |
| Latency-sensitive agent loops | Muse Spark 1.3 | 225.9 against 64.7 tokens per second compounds across every step. |
| Stack built around OpenAI's tools | GPT-5.6 Sol | Moving harness, prompts and integrations costs more than a model price. |
| Non-sensitive bulk work, cost decides | Muse contributor tier | $0.10 / $0.20, if you accept Meta using the inputs to improve its products. |
| Sensitive or contractual data | Standard tier only | Meta states the standard tier is not used to improve its products. |
On capability they are level, so price, speed and data terms decide.
For Sol against the model above it in OpenAI’s lineup, see GPT-6 Astra against Sol, and for Sol against Anthropic’s nearest tier, Opus 5 against Sol.
What would change my mind
A cost-per-completed-task comparison on the same harness. Muse Spark 1.3’s Coding Agent Index score was measured in Muse Code and Sol’s in Codex, and the harness matters as much as the model.
Sol’s price after 21 November. A cut would narrow the gap; the implied undiscounted price would widen it.
Meta changing the contributor tier’s terms. The whole case for that tier rests on one line on one page.
Questions people asked
Is Muse Spark 1.3 better than GPT-5.6 Sol?
On capability, no. They are level at 48.17 against 47.06 on the Intelligence Index v4.3 and 54 against 55 on the Coding Agent Index, and Muse Spark 1.3 is much faster and much cheaper.
How much cheaper is Muse Spark 1.3 than GPT-5.6 Sol?
About 3.2 times on input and 4.7 times on output: $1.25 / $4.25 against $4 / $20. The full index cost $2,000 to run on Muse Spark 1.3 and $3,465 on Sol.
What is the Muse Spark 1.3 contributor tier?
A second model ID at $0.10 / $0.20 per million tokens that Meta’s page describes as used to improve its products. The standard tier is described as not used to improve them.
Should I use Muse Spark 1.3 or GPT-5.6 Sol?
On the numbers, Muse Spark 1.3 on the standard tier. Stay on Sol if your stack is built around OpenAI, and use the contributor tier only for inputs you are comfortable sharing.
If you are weighing a vendor change for a production workload, the contact page is the fastest route to me.
