Meta opened its first paid developer API on July 9, and it led with price. Muse Spark 1.1 lists at $1.25 per million input tokens and $4.25 per million output on the new Meta Model API — roughly a quarter of what a frontier tier like GPT-5.6 Sol charges ($5/$30), and comfortably under Claude Sonnet 5 ($2/$10). New accounts get $20 in free credits. The coverage wrote itself: Meta undercuts Anthropic and OpenAI by 75%.

The sticker is real. Whether it lowers your bill is a separate question, and the answer is not "yes" — it's "measure."

Your bill is tokens × price, and Meta only cut one of those#

A per-token rate is half of a cost. The other half is how many tokens the model spends to finish your task, and that half is a property of the model's behavior, not its price sheet.

A cheaper model that reasons longer, retries more often, or needs more few-shot scaffolding to hit the same quality can burn more tokens per task — and a task that costs 3x the tokens at a quarter of the price is not cheaper. This is the trap every "we switched and saved 75%" post walks into: it compares price per million and forgets to compare tokens per task.

The only number that pays your invoice is cost per completed task. Price per million token is the number the vendor controls; tokens per task is the number your workload controls. Optimize the one you can measure.

The output gap is the one that bites agents#

Read the two columns, not the headline. Muse Spark undercuts on input by two-to-four times. On output — $4.25 versus Sonnet 5's $10 versus GPT-5.6 Sol's $30 — the ratio is different at every tier, and output is where agent loops live. A reasoning-and-tool-calling agent generates a lot of tokens: plans, tool arguments, self-critique, final answers. If your workload is output-heavy, you capture the smaller half of the discount, not the "quarter of the price" headline.

Founders who priced this on the input rate alone will owe finance an awkward correction. Run your actual input/output mix through both rate cards before you promise a number.

What you're actually buying isn't the price#

The price is the reason to try Muse Spark. The reasons to keep it are two capabilities Meta put first.

Computer use is first-class. The model decides, step by step, whether to write an automation script or click through a UI directly, and it emits batches of actions rather than one click at a time. If your product drives a browser or an app, that's a genuine differentiator, not a benchmark line.

It manages its own context. Muse Spark actively works a 1M-token window — retrieving from much earlier in a task and compacting older steps while keeping what later work needs. For long-horizon agent runs, that's the difference between a model that degrades as the transcript grows and one that stays on task. (If you're wrestling with this yourself, see how to manage context in a long-running agent.)

Route, don't switch#

The right move with a cheaper capable model is never a migration. It's a routing experiment.

Take a representative slice of your agent's real tasks. Send them to Muse Spark 1.1 alongside your current model. Measure two things: completion rate and cost-per-completed-task. Keep the frontier model on the hard tail where quality pays for itself, and let the cheap model absorb the volume it can actually complete. This is the same discipline behind an LLM cascade versus a router — the winner is decided on your data, not Meta's press release.

The $20 in free credits exists precisely so you can run that experiment before you owe anyone a dollar. Spend it on measurement, not on a benchmark you'll never run in production.

Prices and availability are as of the July 9, 2026 US public preview and change often; verify the current Meta Model API pricing page before you commit.