The first model Meta ever charged for is priced to win agent work, not coding. At $1.25 and $4.25 Muse Spark 1.1 is the cheapest near-frontier model from a US lab, and the benchmarks it wins tell you exactly what it is for.
Meta opened a paid API on July 9, the first time it has charged developers directly for one of its own models. Muse Spark 1.1 comes in at $1.25 input and $4.25 output per million tokens, roughly a quarter of what OpenAI and Anthropic ask. But the interesting part is not the discount. It is that the benchmarks Meta chose to win are all tool-use and agent evals, and the ones it loses are coding. The rate card and the scorecard are telling the same story. Here is what the price buys, and where it does not.
Image source: Meta AI Blog
Meta's meter, line by line
| Line | Price |
|---|---|
| Input | $1.25 / 1M |
| Cached input | $0.15 / 1M |
| Output (and reasoning tokens) | $4.25 / 1M |
| Web search grounding | $2.50 / 1K queries |
| New-account credit | $20 free |
USD per million tokens unless noted. The context window is 1M tokens. Cached input reads at $0.15, an 88 percent discount off the input rate. One thing to price in: Muse Spark 1.1's Thinking mode bills its reasoning tokens at the $4.25 output rate, so a heavy chain-of-thought call costs like a long answer, not a cheap one.
Meta put a price tag on a model for the first time
For years the Meta model story was Llama, downloaded free and run on your own hardware. Muse Spark 1.1 breaks that. It ships only as a hosted, metered API from Meta Superintelligence Labs, and it is the first time the company has asked developers to pay per token for a Meta model. Mark Zuckerberg marked the moment by posting to X for the first time in about three years, calling it a strong agentic and coding model at a very low price and pitching the rate as roughly 25 percent of what Anthropic and OpenAI charge.
That quarter figure is not marketing rounding. On input it is exact: $1.25 against the $5 that both GPT-5.5 and Opus 4.8 charge is precisely a quarter. On output Muse Spark undercuts even that. Its $4.25 is 17 percent of Opus 4.8's $25 and 14 percent of GPT-5.5's $30. So the honest phrasing is a quarter of rivals' input price and closer to a sixth of their output. For an agent workload, which reads far more than it writes, the blended saving lands near the quarter mark Meta quoted.
What a quarter of the rate looks like on a real bill
Per-token rates are abstract, so price a concrete workload. Take an agent that runs 100M input tokens and 20M output tokens a month, an input-heavy shape that matches how tool-use models actually get used: lots of context and tool results going in, terser answers coming back. No cache. Here is where each model lands.
| Model | Input / output | 100M in / 20M out |
|---|---|---|
| DeepSeek V4 Pro | $0.44 / $0.87 | $61 |
| Meituan LongCat-2.0 | $0.75 / $2.95 | $134 |
| Meta Muse Spark 1.1 | $1.25 / $4.25 | $210 |
| GPT-5.6 Luna | $1 / $6 | $220 |
| Grok 4.5 | $2 / $6 | $320 |
| Claude Sonnet 5 (intro) | $2 / $10 | $400 |
| GPT-5.6 Terra | $2.50 / $15 | $550 |
| Claude Opus 4.8 | $5 / $25 | $1,000 |
| GPT-5.5 | $5 / $30 | $1,100 |
Against the frontier the gap is stark. A team paying $1,000 a month for Opus 4.8 on this workload drops to $210 on Muse Spark, and one on GPT-5.5 goes from $1,100 to the same $210. That is the 4.8x and 5.2x cut Meta is selling, and on an agent shape the numbers hold. What it is not is the floor. Look at the top of the table and two Chinese models sit well below it, which is the part of the story Meta's framing quietly skips.
It wins the agent evals and loses the coding ones, on purpose
Read Meta's benchmark charts and a pattern jumps out. Every eval Muse Spark 1.1 tops is about using tools. It beats Opus 4.8 and GPT-5.5 on MCP Atlas, a scaled tool-use test, and on JobBench, which measures professional tool use, it clears Opus 4.8 by more than six points. On Humanity's Last Exam with tools it leads the field at 62.1. Then look at the losses. They are all pure coding and computer use: SWE-Bench Pro, DeepSWE, Terminal-Bench, OSWorld. The model is not trying to be the best coder you can rent. It is trying to be the cheapest thing that can reliably drive a stack of tools, and it priced itself to match.
| Benchmark | Muse Spark 1.1 | Best rival |
|---|---|---|
| MCP Atlas (tool use) | 88.1 | Opus 4.8 82.2 |
| JobBench (pro tool use) | 54.7 | Opus 4.8 48.4 |
| Humanity's Last Exam, tools | 62.1 | Opus 4.8 57.9 |
| SWE-Bench Pro (coding) | 61.5 | Opus 4.8 69.2 |
| Terminal-Bench 2.1 (agentic code) | 80.0 | GPT-5.5 83.4 |
| OSWorld-Verified (computer use) | 80.8 | Opus 4.8 83.4 |
Keep the usual asterisk in mind: every one of these numbers is Meta's own, run on Meta's harness, and nobody outside the company has reproduced them yet. DataCamp's early rundown notes the model ships closed-weight with no published architecture, so the scorecard is Meta's word until someone reruns it. Vendor benchmarks flatter the vendor. But the shape of the claim is at least internally consistent, and it is easy to sanity-check against your own tool traces. If your agent spends its day calling APIs and stitching results, this is the profile you want. If it spends its day closing SWE-bench-style tickets on its own, Opus 4.8 and GPT-5.5 still win the quality line, and you are trading capability for the lower rate.
The price floor is still in China
Meta's pitch is that Muse Spark undercuts the American frontier, and it does. What it leaves out is that the actual price floor is still in China. Meituan's LongCat-2.0 charges $0.75 and $2.95, landing the same agent month at $134 against Muse's $210. DeepSeek V4 Pro runs a fraction of both, near $61, and V4 Flash is cheaper still. So the precise claim is that Muse Spark 1.1 is the cheapest near-frontier model you can buy from a US company, which for a lot of teams with data-residency or vendor constraints is the sentence that matters. For teams that can point traffic anywhere, the Chinese open-weight coders were already sitting below this, and Muse Spark does not change that math.
Two catches before you switch
First, geography. Muse Spark 1.1 launched as a US-only public preview with no EU access and no international timeline. If your team or your users sit in Europe, this is a model to watch, not one to build on this week. Preview also means the pricing and the availability are provisional until Meta ships a stable release, so do not wire it into a budget you cannot revise.
Second, the paper trail is thinner than the launch noise suggests. Meta's own blog and developer pages describe the model but do not print the numbers; the $1.25 and $4.25 come from Meta's API console as reported by Reuters and Bloomberg, and every tracker repeats the same pair. Meta has also not disclosed the architecture, the parameter count, or whether it is a mixture-of-experts, and it has published no batch-pricing tier. None of that blocks you from using it, but it means the spec sheet has holes an official page would normally fill, and you should treat the numbers as reliable-but-secondhand until Meta puts them on its own docs.
Point tool-use traffic at it, keep coding on the incumbent
If you run an agent or tool-use workload on Opus 4.8 or GPT-5.5 and the bill is the pain, Muse Spark 1.1 is the clearest 4 to 5x cut on offer from a US provider this month, and the benchmark profile says you keep the capability where it counts for that job. Route non-critical tool traffic to it, diff a week of runs against your incumbent, and watch whether it drives your specific tool stack as cleanly as the MCP Atlas number implies.
Hold in three spots. If your workload is coding-first, Opus 4.8 and GPT-5.5 still win the SWE-bench line, so pilot before you migrate. If you can send traffic to Chinese models, LongCat-2.0 and DeepSeek V4 are cheaper still. And if you are in the EU, you cannot use it yet. Drop your real input-output mix into the cost calculator and compare it head to head on the pricing page, because on an agent shape a quarter of the rate is a big enough number to be worth a week of testing.
Sources
- - Meta AI announcement: ai.meta.com
- - Pricing report: AI Weekly
- - The Decoder (pricing and price war): the-decoder.com
- - DataCamp (benchmarks and specs): datacamp.com
- - Related: Grok 4.5 pricing and the coding-agent price war
- - Related: Meituan LongCat-2.0, the 75-cent frontier coder
- - TokenCost pricing page: tokencost.app/pricing