Muse Spark 1.3 review, pricing and limits

Muse Spark 1.3 is Meta's agentic, developer-first model: a million-token context window, no long-context price premium and a max reasoning level now open on the Standard tier. Audio understanding is degraded, so voice work belongs on Muse Spark 1.2 or Muse Voice Transcribe.

Muse Spark 1.3, shipped September 2, 2026, is the newest Muse Spark model from Meta, with a million-token context window for agentic coding and long-document work. Its xhigh tier scores 94% on GPQA Diamond per Artificial Analysis, and a max reasoning level is now available on the Standard API tier.

Provider: Meta AI · Family: Muse

More about Meta AI on Hokai

Context window: 1,048,576 tokens · Max output: 131,072

Input modalities: text, image, video, pdf, tool-calls · Output: text, tool-calls

About Muse Spark 1.3

Meta AI Research shipped Muse Spark 1.3 on September 2, 2026, the fourth entry in the Muse Spark line after the original release (April 8, 2026), Muse Spark 1.1 (July 9, 2026) and Muse Spark 1.2 (August 5, 2026), all built by the same Meta Superintelligence Labs team at Meta. The April launch targeted Meta's own consumer apps first. Version 1.3 is aimed at developers: Meta's documentation positions it for agentic workflows and long-horizon coding, served through the Meta Model API and the Muse Code terminal agent. Parameter count, training data cutoff and full architecture remain undisclosed, as with every earlier release.

The model accepts text, images, video and PDFs and returns text. Meta's API docs put the window at 1,048,576 tokens and state that there is no long-context premium, so a nearly full prompt costs the same per token as a short one. Pricing has two tiers. Standard keeps your data out of training, while Contributor is far cheaper in exchange for permission to train on your prompts and completions; the pricing notes below carry the exact rates. Rate limits are 3,000 requests and 4,000,000 tokens per minute on Standard, and 100 requests and 3,000,000 tokens per minute on Contributor.

Reasoning effort is set per request. The documented levels run from minimal through xhigh, and a max level now sits above xhigh on the Standard-tier model only, billed at the same per-token price but spending more reasoning tokens. At launch Meta had held max back for safety testing, and Artificial Analysis described it as a limited partner preview, so older write-ups that call max invite-only are out of date. Meta's own table for the max configuration shows 75.4 on DeepSWE v1.1, 88.8 on Terminal-Bench 2.1 and 98.5 on MRCR at the 256K-512K span, against 73.0, 88.8 and 91.5 for GPT-5.6 Sol and 74.0, 86.7 and no reported figure for Claude Opus 5. Those are vendor-run numbers, not independent ones.

Independent scoring is harder to compare over time. Artificial Analysis rated the xhigh tier 61 on launch day and has since moved to index v4.3.2, where its model pages show 45 for xhigh and 48 for max, so only compare figures drawn from one index version. On its launch measurements the xhigh tier also used about 57% more input tokens per agentic task than Muse Spark 1.2, which pushed cost per task up even though per-token prices did not change.

Where it fits: teams building coding agents, or reading very long documents and codebases, get the most from it, and the Contributor tier is a cheap way to prototype where data sharing is acceptable. It is the wrong pick when you need audio input, because Meta's docs warn that audio quality in 1.3 may degrade and point to Muse Spark 1.2 or Muse Voice Transcribe instead. It is also the wrong pick for sensitive data on the cheapest tier, or for teams that need Bedrock, Vertex or Azure hosting, since Meta documents only its own API; OpenRouter now lists both 1.3 variants at Meta's own per-token prices. Inside Meta's line, the larger Muse Wonder, released 12 October 2026, is the step up when cost matters less than headroom, and Muse Spark 1.3 stays on sale beside it. Where the alternatives win: Claude Fable 5.1 and Claude Opus 5 posted higher launch-day index scores, Gemini 3.1 Pro and Grok 4.6 come from vendors with broader cloud availability, and Gemini 3.8 Flash and GLM-5.3 sit close in cost per task. Our LLM selection guide widens the shortlist further.

Pricing

Standard tier: $1.25 per 1M input tokens, $4.25 per 1M output and $0.15 per 1M cached input, with no long-context premium and no training on your data. Contributor tier: $0.10 input, $0.20 output and $0.002 cached input per 1M tokens, in return for letting Meta train on prompts and completions; max reasoning is not available there. Max reasoning on the Standard model uses the same per-token rates but more reasoning tokens.

Web search grounding adds $2.50 per 1,000 queries. Prices from Meta's pricing docs, checked 2026-10-19. Compare with GPT-5.6 Sol and Claude Opus 5.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Expanded Context Window: A window of about a million tokens holds a full multi-file codebase in one prompt, with no long-context price premium per Meta's docs.
• Six Reasoning-Effort Levels: Choose minimal, low, medium, high, xhigh or max per request; max is Standard-tier only and spends more reasoning tokens at the same price.
• Contemplating Mode: Runs multiple reasoning agents in parallel on one problem and synthesizes their output into a single answer, carried over unchanged from the original Muse Spark.
• Muse Code Terminal Agent: A terminal-based coding agent co-trained with the model that can run multiple coding sub-agents concurrently on a single task.
• Contributor Pricing Tier: A steeply discounted per-token rate for developers who let Meta train on their prompts and completions, unchanged since its debut alongside Muse Code.

Pros

• Meta's pricing page promises one flat rate whether a prompt is nearly empty or close to the full window, which keeps long-document costs predictable.
• Max reasoning is offered on the Standard tier at the same per-token price as other effort levels, per Meta's reasoning docs.
• Contemplating Mode's parallel multi-agent reasoning remains a genuine architectural differentiator: few widely deployed frontier models run multiple reasoning agents at inference time the same way.

Cons

• Audio understanding is not fully supported in 1.3; Meta points developers to an earlier release or Muse Voice Transcribe.
• The Contributor tier cannot use max reasoning and trains on your prompts and completions.
• Artificial Analysis found input tokens per agentic task rose by more than half versus the previous release, lifting cost per task at flat per-token prices.
• Meta documents only its own Model API and OpenRouter as routes; there is no Bedrock, Vertex AI or Azure listing for teams tied to one of those clouds.

Benchmarks

• GDPval-AA v2: 1,754 vendor-reported · 19 Oct 2026 — Real knowledge-work deliverables judged against professionals, run by Artificial Analysis.
• GPQA Diamond: 94% cited: Artificial Analysis · 19 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
• Terminal-Bench 2.1: 88.8% vendor-reported · 19 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
• Humanity's Last Exam: 47% cited: Artificial Analysis · 19 Oct 2026 — Expert-written questions across many fields, % correct.
• AA Intelligence Index: 45 cited: Artificial Analysis · 19 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $0.78/M cited: Artificial Analysis · 19 Oct 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 327 tok/s cited: Artificial Analysis · 19 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Muse Spark 1.3 Official Page