Aakash Gupta has a rule of thumb for Meta: “Something goes open source the week it stops being the best thing Meta sells.” It’s his bet, not a Meta statement. Meta itself says Muse Spark Open “shares the same base model as Muse Spark 1.3”, a model it still sells.

Whatever Meta’s motive, you now have an open model that does a lot of work, sold by many hosts. So ask the buyer’s question. If this does most of what you run, what is a closed vendor billing you for on the rest?

My answer: the hard tail. The catch is that nobody has published where your tail starts. Two disclosures first. I use Claude Code every day, so I sit on the closed side of this ledger. And I haven’t run Muse Spark Open. Everything below comes from Meta’s own pages, which I read in full, and from the posts I cite.

What the market now charges

Meta’s developer guide lists seven hosted-inference providers on launch day: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Same weights, seven sellers. The model is 816B parameters with 42B active per token and a 1M-token context.

On OpenRouter it lists at $1.25 per million input tokens and $4.25 per million output. Meta’s own paid Muse Wonder is $4 and $20. That’s 3.2x on input and 4.7x on output.

Don’t build a business case on those two lines. Different models, and one price is a host’s, not Meta’s. And as I argued in AI price per token lies, a cheap model that burns more tokens per job can lose to an expensive one. Cost per finished job is the number.

Meta compared itself to everyone except the closed models

The model card says the benchmarks compare Muse Spark Open “to other leading open-weight models.” I read every column. No Claude, no GPT, no Gemini. So the table can’t tell you what you give up by leaving a closed model.

It does show a split inside the open group, using Meta’s own figures, which I couldn’t find reproduced elsewhere. On MCP-Atlas, a tool-use test, Muse Spark Open scores 89.2 against 84.2 for GLM 5.3 and 82.3 for Kimi K3. On Terminal-Bench 2.1 it scores 76.4 against 83.9 and 85.0. Same table, Artificial Analysis Intelligence Index: 39 against 45 and 44. Gupta spotted the Terminal-Bench gap, which is why he reads “most capable US open-weight model” as a statement with “US” doing deliberate work.

So there is a ceiling for work an open model of this size can do. Where your work goes past that ceiling stays unmapped, and only your own cases can map it.

Route by task, and let code do the routing

Aaron Levie’s post on the release splits the work: frontier models for “orchestration and high end tasks,” tuned low-cost models for “repeatable workflows.” I agree with the architecture. Orchestration and hard reasoning go to the frontier model. Classification, extraction and standard replies go to the cheap one.

One reply under his post has the condition right. Karl wrote that if an employee must pick among three AIs just to send an invoice, “we’ve given them another job.” The router belongs in software, one rule per workflow.

Unit economics follow. You stop watching spend per vendor and start watching cost per finished job, per workflow. A workflow that moves to the cheap model drops out of your frontier bill, and the frontier invoice shrinks to the tail. If your contract prices per model, you can’t capture that. The clauses I listed for model routing cover this: price per unit of work, and the right to take your test cases with you.

Lock-in changes hands

Open doesn’t mean unbound. The Muse Model License, which I read in the text on the model card, allows use, fine-tuning, distillation and redistribution, with these conditions:

• A “Built with Muse Spark Open” notice in every product that uses the model or a derivative.
• “Derivatives” include any model trained, fine-tuned or distilled using the model or its outputs. Train a small model on Muse answers and it carries the notice too.
• If you sell AI models to third parties through an API and you plus affiliates top $20 million in revenue over twelve months, you sign a separate agreement with Meta first. Internal use and features of your own product are carved out.
• Meta’s usage policy is part of the license.

I found no monthly-user threshold in the text, a question that came up in the replies to Gupta’s post. On hardware, the card asks for a GPU setup with at least 700GB of VRAM for NVFP4. For most companies that means a host, so your lock-in becomes the host, your fine-tune and your test set. Meta also reversed course on open weights once already, which I covered in the Altman, Zuckerberg and Nadella post. The file you downloaded stays yours. The next version might not.

A test for this week

Pick your highest-volume workflow. Collect 20 real cases with known correct answers. Run them through the model you use now and through Muse Spark Open on one host. How many does each pass? What does each cost per finished case? Do they miss the same ones? Where they miss different ones, that’s your routing rule.

Building those cases is the slow part, and they stay with you through every model change after this one.

Send us your highest-volume workflow and we’ll set up the comparison