Meta’s Muse Spark 1.1 is one of the cheapest frontier models you can run, and Meta built it to be an agent rather than a general-purpose chat model. On raw benchmarks it sits mid-pack. Put Muse Spark in front of real software with a job to finish and Meta reports it takes first place. If you already pay for GPT or Claude by the token, this is the release that changes what you pay, not what you can do.
What changed from Llama to Muse Spark
Meta’s flagship was Llama for years, and Llama had two things going for it: open weights, and general purpose. Muse Spark goes the opposite way on both. It is closed. You use it free inside Meta’s AI app, or you pay for it through Meta’s API.
Rebuilding the entire training stack from scratch on new architecture got Meta the old Llama 4 result with ten times less compute. Intelligence Index, an independent scoring system, puts Llama 4 at 18 and Muse Spark at 52, about a year apart. Version 1.1 is the upgrade that just landed: the first Muse model with a paid API, a jump in coding and tool use, and the safety numbers below.
Why Muse Spark is good at agent work
Muse Spark carries a one million token context window and can run copies of itself. A token is about four characters, so a million of them is roughly 750,000 words: every job sheet, quote and complaint email your team wrote last quarter, held in front of the model at once instead of the early half falling out while it works on the rest. The copies are how Muse Spark handles the size. A main agent splits a big job into pieces and hands them to sub agents, parallel copies working at the same time. Sub agents work the same way in Claude Code.
Computer use is the part that decides whether any of this reaches your business. Muse Spark is trained to operate software through a browser, on desktop or mobile, and it picks how: write a quick script when the work is bulk, click through the interface when clicking is faster. That is the difference between a temp who fills in the same form 400 times and one who works out that 400 forms want a script. If your operation lives inside a utility interconnection portal or a warranty claims site with no API, that choice is most of the job.
Muse Spark 1.1 vs Opus 4.8 and GPT-5.5
| Test | Muse Spark 1.1 | Opus 4.8 | GPT-5.5 |
|---|---|---|---|
| Humanity’s Last Exam (broad knowledge and reasoning) | 62.1 | 57.9 | not published |
| MCP Atlas and Job Bench (using outside tools) | first place, per Meta | not published | not published |
| Agentic finance tasks (money jobs run end to end) | wins, per Meta | behind, no figure given | behind, no figure given |
| SWE-bench Pro (fixing real software bugs) | 61.5 | 69.2 | 58.6 |
| Terminal Bench (command line work) | 80 | ahead, no figure given | ahead, no figure given |
Two of those five tests measure whether a model can pick up an outside tool and finish something, and those two are where Muse Spark takes first place. They are also the two that look like work you would hand it. SWE-bench Pro is the row to read if you are paying someone to build software on this model: 61.5 edges past GPT-5.5 and lands just under eight points behind Opus 4.8. Anthropic’s Fable 5, the model that was pulled after three days, scored 80.3 on the same test. Muse Spark 1.1 orchestrates and decides at frontier level, and writes hard code at mid-tier.
The safety result is the biggest technical leap in the release, and it arrives as two numbers that are not the same number. StrongReject is a jailbreak test, and its attack success rate counts how often someone slips a banned request past the guardrails. Muse Spark’s StrongReject rate fell from 25% in the previous version to half a percent in 1.1. Meta’s broader framing is that fewer than one in ten single-turn attacks get through, against roughly one in five for GPT-5.5. Those two figures sit twenty times apart because they measure different things: half a percent is the StrongReject score, one in ten is Meta’s own headline across single-turn attacks generally. Quote the test. A former OpenAI policy researcher summed the release up: Meta used to be the lab that leaned towards capability over safety, and that spot now belongs to xAI alone.
Muse Spark pricing, and when to switch
Muse Spark 1.1 runs at $1.25 per million input tokens and $4.25 per million output tokens, roughly a quarter of what the flagship OpenAI and Claude models charge.
| Per million tokens | Muse Spark 1.1 | Flagship OpenAI or Claude model |
|---|---|---|
| Input | $1.25 | about $5 |
| Output | $4.25 | about $17 |
The right hand column is that quarter claim worked backwards, not a published price list, so read it as the ratio and check your own invoice.
Meta built the API as a drop-in replacement, compatible with the OpenAI and Anthropic SDKs, so a developer already building on GPT or Claude switches with a one-line change. That matters to you in one way. Nobody can tell you the switch is a big project.
The work most field-services teams hand to an agent is tool calling and clicking: pull the job from the CRM, check the engineer’s calendar, send the quote, chase the invoice on day 31. None of that is hard code, and all of it is the profile Muse Spark is priced for. Pull last month’s token bill for that automation and divide it by four. If the number that comes back is not worth the switching work, stay where you are.
