Muse Spark 1.3: Pricing, Specs, and the Contributor Tier

Meta shipped Muse Spark 1.3 on 2 September 2026 without a blog post. There is no launch announcement on Meta's AI blog — the most recent Muse Spark entry there is still 1.1 from July. The release landed as a documentation and model-catalogue update, which is why the interesting details are buried in a pricing table rather than a keynote.

On paper 1.3 is a quiet point release: identical context window, identical price, identical modality list to 1.2. The actual story is the second SKU shipped alongside it — a Contributor tier that costs 12.5x less on input and 21.25x less on output for what is, by every published specification, the same model. This guide covers what 1.3 is, what changed, and specifically what you trade away to get that price.

What is Muse Spark 1.3?

Muse Spark 1.3 is a closed, API-only multimodal reasoning model from Meta, aimed at long-horizon agentic and coding work. Meta's own model page describes it as "trained for agentic workflows and optimized for competitive coding performance," with the promise of "higher first-attempt accuracy and reliable tool calling."

It is served exclusively on Meta Model API at https://api.meta.ai/v1. There are no downloadable weights and no Hugging Face repository. That distinction matters within Meta's own line-up: Muse Glimmer 30B is the open-weights Apache 2.0 model you self-host, while Muse Spark stays behind the API.

* Audio is the asterisk that matters. Meta's models documentation carries an explicit warning: "Audio understanding in Muse Spark 1.3 is currently not fully supported, and response quality for requests including audio content may be degraded. For audio, use Muse Spark 1.2, or Muse Voice Transcribe for dedicated speech-to-text."

Meta does not publish a knowledge cutoff for Muse Spark 1.3 anywhere in its documentation or model catalogue, and the field is empty in third-party listings. If your application depends on recency, test it rather than assuming.

What changed from Muse Spark 1.2?

Less than you would expect from a version bump, and one thing moved backwards.

Context window, pricing, modality list, reasoning-effort levels and supported parameters are byte-for-byte identical between 1.2 and 1.3. What Meta claims changed is reasoning depth and coding: the model page positions 1.3 as offering "max reasoning for challenging reasoning and agentic tasks and improved real-world usability," and the docs call it "tuned for agentic workflows (multi-step tool, browser, and long-horizon tasks) with improved coding over 1.2." It is now the recommended default for new work and the model used in every code sample in Meta's docs.

Two things stand out. First, the audio regression is real and documented by Meta itself — if you built a pipeline on 1.2's audio understanding, 1.3 is not a drop-in upgrade. Second, the Contributor tier arrived on day one for 1.3, where 1.2 waited sixteen days for it. Meta clearly considers the cheap tier part of the product now, not an experiment.

What is the Contributor tier, and what do you give up?

This is the part of the release that actually changes decisions.

Meta ships two SKUs of the same model. muse-spark-1.3 is the standard tier. muse-spark-1.3-contributor is the discounted one. Every published specification is identical — same 1,048,576-token context, same 943,718-token max output, same modalities, same reasoning efforts, same tool and structured-output support. The only differences are price, throughput, and what Meta is allowed to do with your data.

The documented terms, verbatim

Meta is unusually direct about this, which is to its credit. Three separate places in its own documentation state the trade:

• The pricing and rate limits page, on the Contributor tier: "Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models."
• The same page, on the Standard tier: "Standard pricing; your prompts and completions are not used to train Meta models."
• The models page: the Contributor variant "trades a lower price for permission to train on your prompts and completions."

Meta's public pricing table puts it in four words per row: the contributor SKU is "used to improve our products," the standard SKU is "not used to improve our products." Third-party catalogues restate it the same way — the OpenRouter listing for the contributor SKU carries the line "Prompts and outputs may be used to improve Meta's products," and the standard listing carries no such line.

What Meta does not document

Being clear about the headline term is not the same as documenting the arrangement. As of 3 September 2026, Meta's public documentation does not state:

• How long contributor prompts and completions are retained.
• Whether humans review them.
• Whether you can revoke training rights after a request has been sent, or request deletion.
• Whether file and image attachments, tool-call arguments, and tool results count as "prompts" for training purposes. In an agentic loop these carry the most sensitive material — file contents, stack traces, database rows.
• Whether a zero-data-retention option exists for Muse Spark. Meta mentions ZDR only for Muse Voice Transcribe, where it notes ZDR "is priced at parity with Standard."

We are not going to guess at any of these. If those answers matter to your compliance posture, they need to come from Meta in writing, not from a docs page.

The structural problem with a config-string decision

The tier is selected by a model ID. There is no separate contract, no signed data-sharing agreement, no admin toggle a security team can lock. A 12-to-21x line-item saving sits behind a one-word change to a string in a config file — a change any engineer can make in thirty seconds while trying to bring an AI bill down.

Training is also the one AI data decision you cannot walk back. Logs get deleted and retention windows expire, but once material is absorbed into model weights there is no extraction procedure. We covered this dynamic in depth when Meta first shipped a contributor SKU for its terminal agent, in Muse Code's Contributor Tier: What 21x Cheaper Actually Costs You.

When the Contributor tier is fine, and when it isn't

Reasonable uses: benchmarking and model evaluation, prompt engineering against public or synthetic data, throwaway prototypes, personal projects, load-testing an integration, coursework. Meta's own framing — "prototyping, testing integrations, and scaling experiments where training on your data is acceptable" — accurately describes the intended market.

Do not use it for: client code under an NDA, proprietary source you would not publish, regulated data of any kind (health, financial, personal data under GDPR), anything covered by a customer DPA that restricts sub-processor training, or unreleased product material. Trade-secret protection depends on the information having been kept secret; code used to train a third party's commercial model is hard to characterise that way afterwards. If you use the tier at all, pin it explicitly and keep the model ID out of any config a production service can read.

Why is the max output 943,718 tokens?

943,718 is not an arbitrary number. It is exactly 90% of the 1,048,576-token context window. Meta reserves the remaining 10% — 104,858 tokens — and lets the model spend everything else on generation.

That is far beyond what the rest of the market allows:

Is it real in practice? Partly. The ceiling is genuine — the API will accept a max_completion_tokens in that range, and Meta charges no long-context premium, so a near-maximum generation costs the same per token as a short one. Artificial Analysis notes that 1.3 emitted 120M output tokens across its evaluation suite against a peer median of 72M, so this is a model that genuinely writes long when allowed to.

But three constraints bite before the ceiling does. Output and input share one budget: a 943,718-token completion leaves roughly 105,000 tokens for your entire prompt, which rules out the long-context-in, long-context-out case people imagine. Cost scales linearly — one maximum-length completion is about $4.01 on the standard tier, or $0.19 on contributor. And Meta's own recommended agent configuration, in the quickstart guide, sets output: 131072 rather than anything near the API maximum — a fair signal of where Meta expects real workloads to sit.

The practical value is not one enormous document. It is that you effectively never have to think about output truncation in an agent loop: long refactors, full test-suite generation, and exhaustive migration diffs complete in a single response instead of needing continuation logic.

How good is Muse Spark 1.3?

Separate the vendor claims from the independent numbers, because the vendor side is unusually thin here.

Vendor. Meta publishes no numeric benchmark table for Muse Spark 1.3. Its model page carries a benchmark chart as an image with no machine-readable values behind it, and the accompanying prose claims only that the model "performs competitively with frontier models across several coding evaluations." There is no scorecard against SWE-bench, Terminal-Bench, GPQA or anything else. For a coding-positioned model that is a conspicuous gap, and it is a reason to weight independent evaluation more heavily than usual.

Independent. Artificial Analysis scores Muse Spark 1.3 (max) at 62 on its Intelligence Index v4.1.1, ranking it #6 of 636 models tracked. Muse Spark 1.2 (xhigh) scores 57. That five-point gain is the single most credible piece of evidence that 1.3 is a real improvement rather than a relabel. The index aggregates nine evaluations — GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR — though individual per-benchmark scores are not yet public for this model.

On LMArena, Muse Spark 1.3 has no rating yet. Muse Spark 1.2 (xHigh) sits at rank 5 with 1499 ±10 from 3,240 votes, behind four Anthropic models — claude-fable-5 at 1507, claude-opus-4-6-high at 1505, claude-fable-5.1-max at 1504 and claude-opus-4-7-high at 1502. Human-preference data for 1.3 will take a few weeks to accumulate. Our benchmark breakdown of Muse Spark 1.2 against Claude Opus 5 holds up as the closest available proxy until then, and we go deeper on the 1.3 matchup in Muse Spark 1.3 vs Claude Opus 5.

How does Muse Spark 1.3 compare on price?

The standard tier is priced as a mid-market frontier model. The contributor tier is priced as something else entirely.

That is the framing that matters. On a 3:1 input-to-output blend, the contributor tier lands at $0.125 per million tokens — within a rounding error of GLM-5.3-Flash and DeepSeek V4-Flash, the two Chinese models that have been setting the floor on inference pricing all year. Meta has not built a cheaper model to compete there; it has priced the same frontier-class model into that bracket and covered the gap with training data.

A worked example on a typical agent turn of 20,000 input and 2,000 output tokens: $0.0335 on the standard tier, $0.0024 on contributor. Roughly 14x. For a background job running ten thousand of those a day, that is $335 versus $24.

The rate limits change the shape of the saving, though, and this is the detail most coverage misses. Contributor's throughput cut is asymmetric: requests per minute drop 30x, from 3,000 to 100, but tokens per minute only drop 25%, from 4M to 3M. Contributor is therefore well-suited to a small number of very large requests — document analysis, long refactors, batch summarisation — and badly suited to high-frequency small calls like autocomplete or a chat frontend. To saturate both limits your average request needs to be around 30,000 tokens. Limits apply per team, not per API key, so extra keys buy you nothing.

How do you use Muse Spark 1.3?

Meta Model API is drop-in compatible with the OpenAI SDK, the Anthropic Messages format, and most OpenAI-compatible agent harnesses. Three values get you connected: base URL https://api.meta.ai/v1, your MODEL_API_KEY, and the model ID.

curl -X POST "https://api.meta.ai/v1/responses" \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "muse-spark-1.3",
    "input": "Summarise this changelog and list breaking changes."
  }'

The same call in Python, using the OpenAI SDK against Meta's base URL:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-spark-1.3",           # standard tier: not used for training
    # model="muse-spark-1.3-contributor",  # 12-21x cheaper: Meta trains on this
    input="Summarise this changelog and list breaking changes.",
)

print(response.output_text)

Two practical notes. Prompt caching is automatic prefix caching with no cache key to manage — put stable content (system prompt, examples, history) first and volatile content last, then read cached_tokens in the response's usage object to confirm hits. And if you are wiring 1.3 into an agent harness that replays reasoning across turns, include reasoning.encrypted_content; without it the model reasons from scratch every turn and loses the thread in long tool loops.

Web search grounding is billed separately at $2.50 per 1,000 search queries on top of token cost, on both tiers.

Which tier should you actually pick?

A simple decision rule. If the prompt contains anything you would not paste into a public GitHub issue, use muse-spark-1.3 and pay the standard rate. If it does not, the contributor tier is an honest trade and a genuinely good deal — you are getting frontier-class reasoning at open-weights prices, and Meta is telling you exactly what it wants in return.

The failure mode is not choosing wrong; it is choosing by accident. The tier is a string in a config file, the saving is real enough to be tempting, and the consequence is irreversible. Decide it deliberately, write the decision down, and make sure the person who owns your IP is the one who made it.

For where Muse Spark 1.3 sits in the wider model landscape, see our Muse Spark family guide and the open-source LLM landscape for 2026.

FAQ

What is Muse Spark 1.3?

Muse Spark 1.3 is Meta's closed, API-only multimodal reasoning model, listed on 2 September 2026. It accepts text, images, video, PDFs and audio and returns text, with a 1,048,576-token context window and up to 943,718 output tokens. Meta positions it for long-horizon agentic workflows and coding, and it is the recommended default for new work.

How much does Muse Spark 1.3 cost?

The standard tier costs $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million cached input tokens — identical to Muse Spark 1.2 and 1.1. The Contributor tier costs $0.10 input, $0.20 output and $0.002 cached. Web search grounding adds $2.50 per 1,000 queries on both tiers. There is no long-context premium.

What is the Muse Spark Contributor tier?

It is a second SKU, muse-spark-1.3-contributor, with identical specifications to the standard model but heavily discounted pricing: 12.5x cheaper on input, 21.25x on output, 75x on cached input. In exchange, Meta may use your prompts and completions to train future models. It is also rate-limited to 100 requests per minute against standard's 3,000.

Does Meta train on your data with the Contributor tier?

Yes, and Meta says so explicitly. Its pricing documentation describes the tier as "heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models," and the public pricing table marks it "used to improve our products." The standard tier is marked "not used to improve our products." Retention periods, human review and deletion rights are not documented.

Is Muse Spark 1.3 open source?

No. Muse Spark 1.3 is closed and API-only — there are no downloadable weights and no Hugging Face repository, and independent trackers list it as proprietary. Meta's open-weights model is Muse Glimmer 30B, released under Apache 2.0 and designed to be self-hosted on vLLM, SGLang, llama.cpp or ExecuTorch.

What's new vs Muse Spark 1.2?

Context window, pricing, modalities and supported parameters are unchanged. Meta claims deeper reasoning for hard agentic tasks and improved coding, and independent scoring supports that: Artificial Analysis rates 1.3 at 62 on its Intelligence Index versus 57 for 1.2. One regression: audio understanding is degraded in 1.3, and Meta recommends staying on 1.2 for audio work.

How does Muse Spark 1.3 compare to Claude Opus 5?

Muse Spark 1.3 is far cheaper — $1.25/$4.25 against Opus 5's $5.00/$25.00, roughly 5x lower blended — and allows 943,718 output tokens against Opus 5's 128,000. Claude leads on human preference: Anthropic models hold the top four LMArena text positions, and Muse Spark 1.3 has no Arena rating yet. Judge on your own evals until it does.