Meta's Muse Code Prices a Coding Agent at a Quarter of Claude Opus 5, Not First Place
Meta shipped its third model in four months on August 5, and for the first time it comes bundled with a product you'd actually run against your own codebase: Muse Code, a terminal coding agent built on the new Muse Spark 1.2 model. The model itself isn't frontier. It scores 54 on the Artificial Analysis Intelligence Index, tied with GPT-5.5 and Grok 4.5, a real step behind Claude Opus 5's 61. What changes the conversation is the price tag: $1.25 per million input tokens and $4.25 per million output, against Opus 5's $5 and $25. That's not a rounding difference. It's a 4-6x discount on a model that's close enough to matter for a lot of what a solo operator actually asks a coding agent to do all day.
I've been skeptical of every "good enough at a discount" pitch in this category, because most of them turn out to be good enough at a discount and also annoying to use. So the interesting question here isn't the benchmark chart, it's whether that gap is worth testing on a real repo instead of dismissing because Muse Spark 1.2 isn't the top model on the leaderboard.
What actually shipped
Muse Code is Meta's first terminal-based coding agent, in beta, installable with a one-line curl command on macOS and Linux. It runs a main agent loop alongside persistent async background agents that stay active for the whole session instead of spinning up per task, and it keeps a local event log of every model call, tool run, and edit so it can resume a crashed session exactly where it stopped. It ships default skills too: /plan turns a task into an approval-gated plan, /grill stress-tests that plan, and /goal works toward a stated objective. None of that existed when Meta's April piece covered Muse Spark going proprietary; back then there was a model and an API, not a product you'd point at a terminal.
Muse Spark 1.2 is the model underneath it, and it's Meta's third release since it dropped the open-weight Llama approach in April: 1.0 scored 43 on the Artificial Analysis Intelligence Index at launch, 1.1 scored 51 last month, and 1.2 now scores 54. Meta and Muse Code were co-trained together, with training that specifically targeted long-horizon coding: whole-repository generation, large end-to-end projects, and what Meta calls auto-research. According to Artificial Analysis, most of the jump from 1.1 to 1.2 is concentrated in agentic evaluations rather than raw reasoning: GDPval-AA v2, a benchmark for realistic agentic knowledge work, rose 260 Elo points, while scientific reasoning benchmarks like SciCode actually dipped slightly. This is a model getting better at doing things, not necessarily at knowing things.
Where it actually lands on the leaderboard
The plain version: Muse Spark 1.2 (xhigh) sits second on Terminal-Bench 2.1, Meta's own agentic-coding benchmark, ahead of GPT-5.6 Terra and Grok 4.5 but behind Claude Opus 5. Artificial Analysis's independent run put the model's Terminal-Bench 2.1 score at 80%, up from 1.1's 78%. On the broader Intelligence Index, Muse Spark 1.2 lands at 54, statistically tied with GPT-5.5 (55) and Grok 4.5 (54), and meaningfully behind the current frontier cluster: Claude Opus 5 at 61, Claude Fable 5 at 60, GPT-5.6 Sol at 59, and Kimi K3 at 57. Worth flagging, since Meta ran its own benchmarking and shared results with Artificial Analysis ahead of the public release: Meta's own methodology notes that its evaluation harness and system prompts may not be tuned to third-party models' strengths, so treat any Meta-reported competitor number as directional rather than gospel.
The number that stands out more than the ranking is cost efficiency. Artificial Analysis calculates Muse Spark 1.2 costs about $0.40 per Intelligence Index task, among the cheapest in its performance tier; only Grok 4.5 and a mid-tier GPT-5.6 Sol configuration beat it. That's the real story buried under the leaderboard position: this isn't a model racing for first place, it's a model racing for best price-per-capability.
The pricing math that matters more
Here's the number I'd actually sit with if I were choosing a daily-driver coding agent: Muse Spark 1.2 costs $1.25 per million input tokens and $4.25 per million output tokens (with cache hits discounted further to $0.15 per million). Claude Opus 5 costs $5 and $25 for the same two numbers. That's a 4x gap on input and a 6x gap on output, and output tokens are usually where a coding agent burns its budget, since generated code, diffs, and explanations all count as output.
Run the arithmetic on a realistic day. If your coding agent generates something like 500,000 output tokens across a working day (not an outrageous number if you're running multi-file refactors or test generation in a loop), that's $2.13 on Muse Spark 1.2 versus $12.50 on Opus 5. Do that five days a week and you're looking at roughly $11 versus $63 for the month, just on output. Input costs compound the same way once you're feeding an agent a large repo's context on every call. None of this is exotic; it's just what "running an agent against a codebase all day" actually costs once you stop treating tokens as free.
The realistic use case, minus the hype
I wouldn't point Muse Code at my hardest debugging session or a gnarly architecture decision. A 7-point gap on the Intelligence Index and a real deficit on GDPval-AA v2 against Opus 5 (1631 Elo versus 1852) means there's a ceiling here, and if you're paying for intelligence on the task that actually needs it, pay for it. But most of what fills a solo operator's day isn't that task. It's boilerplate, it's test generation, it's routine refactors, it's the fourth CRUD endpoint this week that looks like the first three. That's exactly the kind of high-volume, lower-stakes work where a quarter of the price for maybe 85% of the ceiling is a genuinely different trade than the one the leaderboard implies.
The honest failure mode of "close enough at a discount" pitches is usually workflow friction, not raw capability: worse tool-calling reliability, clunkier context handling, an agent that technically completes the task but takes three extra turns to get there. Muse Code is beta, it's brand new, and I haven't run it against a real production repo yet, so I'm not going to claim it's frictionless. That's exactly why this is a test-it-yourself claim, not a take-my-word-for-it one.
What I'd actually do
Don't replace your frontier model. Split your workload instead. If you're running Claude Opus 5 or GPT-5.6 Sol for every single coding task regardless of stakes, you're paying frontier prices for tasks that don't need frontier intelligence, and that's a habit, not a decision. I'd install Muse Code this week, point it at a low-stakes slice of a real project (test generation, a routine refactor, some boilerplate you'd otherwise hand-roll), and track two things: does it actually save money at the token rate difference, and does the beta-grade tooling cost you more time than the price gap saves. Keep Opus 5 or GPT-5.6 Sol for the architecture calls and the debugging sessions where being wrong is expensive. That's not a verdict on which model is "better." It's just matching the tool's price-to-capability ratio to the actual stakes of the task, which is a cheaper habit than defaulting to the most expensive option out of routine.
Author
Lukas
@lukcombinator