Meta Muse Spark 1.3: Cheap Frontier Tokens and Long Context, Tested (2026)

Meta shipped Muse Spark 1.3 on September 2, 2026, and the interesting thing about it is not the intelligence rank, it is the price. The model lands inside the top ten of the Artificial Analysis Intelligence Index while charging 1.25 dollars per million input tokens and 4.25 per million output. Set that against GPT-6 Astra at 10 and 50, and the gap is not incremental, it is an order of magnitude on output. When a model this cheap posts near frontier scores and near perfect retrieval across a million token window, the question stops being whether it is the smartest model available and becomes which parts of your workload you can stop overpaying for.

Attribute	Muse Spark 1.3	GPT-6 Astra	Claude Opus 5
Vendor	Meta	OpenAI	Anthropic
Context window	1,000,000 tokens	1,000,000 tokens	large, model dependent
Input price per 1M	1.25 dollars	10 dollars	frontier tier
Output price per 1M	4.25 dollars	50 dollars	frontier tier
Contributor tier input	0.10 dollars (data shared)	not offered	not offered
Modalities	text, image, video	text, image	text, image

Prices above are list rates at launch. Treat the Opus 5 column as the frontier reference point rather than a single published number, because Anthropic prices Opus tiers by mode and the comparison that matters here is the shape of the gap, not the decimal.

The pricing is the headline, so read it carefully

Two dollars and change per million tokens combined is cheap enough to rewrite the economics of any high volume task: classification, extraction, summarization, first pass drafting, retrieval reranking. Meta also published a contributor tier at 0.10 dollars per million input tokens, and that number deserves a hard look before anyone wires it into a pipeline. The 0.10 rate requires you to grant Meta permission to train on your prompts and completions, and it comes with lower rate limits. For a public content workflow that is a defensible trade. For anything touching customer data, proprietary code, or contract text, it is a compliance problem wearing a discount, and you should price your real usage at the standard 1.25 and 4.25 rates, not the contributor rate you cannot legally use.

The cost story gets sharper on a per task basis. On Artificial Analysis cost per task measurements, the higher effort Muse Spark 1.3 configuration lands around 0.84 dollars per task, among the lowest in the field, partly because Meta claims the model uses roughly 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2 to reach the same result. Fewer tokens to finish a job compounds with a low per token rate, and that compounding is where the real savings live, not in the sticker.

Long context retrieval is the standout result

The benchmark where Muse Spark 1.3 genuinely separates is long context retrieval. On a million token needle test the model scores in the high 90s, reported around 98.5 and 98.1 on the public runs, while several competitors fall into the 60s and 70s as the context fills. That matters because most models advertise a million token window and then quietly lose the ability to find a specific fact once you actually pack the window. A large context you cannot reliably retrieve from is a marketing number. A large context with high 90s retrieval is an architecture you can build on.

Concretely, this is the profile you want for retrieval augmented generation over big corpora, for feeding an entire repository or a long contract into a single call, and for any workflow where the alternative is chunking and hoping your reranker surfaces the right passage. If your current stack chunks documents mainly because your model forgets the middle of a long input, Muse Spark 1.3 is worth a controlled test against that exact failure mode.

Where the frontier still wins

Cheap tokens and strong retrieval do not make a model an agent. Meta wins every coding row in its own launch scorecard, and vendor scorecards are exactly the numbers to distrust, because the vendor picks the tasks and the harness. On independent agent benchmarks the picture is more sober: Claude Opus 5 in its higher effort mode still leads on four of six of the agentic evaluations that stress multi step tool use and long horizon planning. Muse Spark 1.3 is competitive, not dominant, on the work where a model has to hold a plan across many steps and recover from its own mistakes.

That split is the actionable part. Retrieval and single pass generation reward the cheap, high recall model. Autonomous multi step agent loops, where one wrong tool call cascades, still reward the model with the best planning and self correction, and today that is the frontier tier. The mistake would be to read one strong intelligence rank and route your agent traffic to the cheapest option, then wonder why task completion dropped.

A simple routing rule

The way to use a model like this is not to switch everything to it. It is to split your traffic by job.

Send high volume, low risk, retrieval heavy, and single pass work to Muse Spark 1.3 and capture the order of magnitude cost cut where it is safe. Keep long horizon agent loops, high stakes reasoning, and anything where a single error is expensive on the frontier model. A router that classifies the incoming request and picks the tier is a few hundred lines of code, and at these price gaps it pays for itself almost immediately on any workload with real volume. We walk through the token math behind that split in our breakdown of prompt caching and throughput economics across Opus, GPT, and Gemini, and the same logic that made Kimi K3 attractive against GPT-6 Astra on cost applies here, only more so given Muse Spark 1.3 is cheaper still.

FAQ

Is Muse Spark 1.3 open weights? No. Despite Meta's history with open models, Muse Spark 1.3 is served through the API. The contributor pricing tier is a data sharing arrangement, not a weights release, so do not plan on self hosting it the way you would an open weights model.

How does the price compare to GPT-6 Astra? Muse Spark 1.3 lists at 1.25 and 4.25 dollars per million tokens against Astra at 10 and 50. On output tokens that is roughly a twelve times difference, which is why the routing decision is worth making deliberately rather than defaulting everything to the frontier.

Can I trust the coding benchmarks? Treat Meta's own scorecard as a claim, not a verdict. It wins coding rows that Meta selected. On independent agent evaluations Opus 5 still leads several categories, so validate on your own tasks before you move agent traffic.

What is the contributor tier catch? The 0.10 dollar input rate requires granting Meta permission to train on your prompts and completions, plus lower rate limits. For most business data that permission is a non starter, so budget at the standard rate.

What is the best first use case? Long context retrieval and high volume single pass tasks. That is where the combination of near perfect million token recall and a very low per token price produces immediate, low risk savings.

When this applies to your stack

If you are running retrieval over large document sets, processing long contracts or repositories in single calls, or paying a frontier rate for high volume extraction and classification, Muse Spark 1.3 is a live candidate to cut that specific line item hard while you keep the frontier model for agent loops and high stakes reasoning. The work is not swapping providers, it is building the routing layer that sends each request to the right tier and measuring the completion rate so a cost cut does not quietly become a quality cut.

That routing and evaluation layer is exactly the kind of applied AI infrastructure we build. If you want help designing a model router, an evaluation harness that catches quality regressions before they ship, or a long context retrieval pipeline that actually uses the window you are paying for, Contra Collective can put the measurement and the plumbing in place so the savings are real and the quality holds.