Muse Spark 1.3 review: Muse Code, pricing, and max preview

Key takeaways

• Muse Spark 1.3 launched September 2, 2026 in Muse Code and the Meta Model API. Fourth Muse Spark release in about five months.
• xhigh reasoning is generally available now. max is in limited partner preview pending more safety testing.
• List price unchanged from 1.2 for xhigh: $1.25 / 1M input, $4.25 / 1M output, cache hits about $0.15 / 1M (Artificial Analysis / Meta pricing trackers).
• Meta claims better long-horizon agents, clearer collaboration (asks when stuck, confirms consequential actions), and ~20% fewer tool calls / ~25% fewer tokens versus 1.2 in internal eng comparisons.
• Artificial Analysis puts xhigh at 61 and max at 62 on its Intelligence Index, with gains concentrated in agentic benches. Independent index, still not your production eval.
• Proprietary Meta model (not Llama open weights). Distinct from open Meta lines.

Table of contents

1. What Muse Spark 1.3 is
2. Pricing and variants
3. What Meta says changed
4. Independent snapshot
5. Who should try it
6. FAQ

---

What Muse Spark 1.3 is

Meta keeps shipping Muse Spark on a monthly-ish cadence while talking up personal agents. For builders, ignore the “superintelligence” slogan and read the shipping surface.

Muse Spark 1.3 is Meta’s proprietary multimodal reasoning model for agentic and coding work, delivered through Muse Code (Meta’s coding agent harness) and the Muse Spark Meta Model API.

That separates:

• Muse Spark 1.3 (model): closed weights, text/image/video in, 1M context (unchanged from 1.2 per Artificial Analysis).
• Muse Code (harness): Meta’s agent product you install (curl -fsSL https://dev.meta.ai/install.sh | bash on macOS/Linux per Meta’s post).
• Llama / open Meta models: different lane. Spark is not an open dump.

Pricing and variants

Variant	Access (Sep 2026)	Pricing signal
Muse Spark 1.3 (xhigh)	Muse Code + Meta Model API	$1.25 / $4.25 per 1M in/out; ~$0.15 cached input
Muse Spark 1.3 (max)	Limited partner preview	Public list price not announced at launch
Contributor-style cheap endpoints (third-party writeups)	Some Meta API tiers	Much lower token rates if you allow training use of traffic; verify live Meta docs before you assume $0.10/$0.20

Axios quotes Meta AI chief Alexandr Wang calling the pricing posture “aggressive” versus peers. Artificial Analysis flags Muse Spark 1.3 (xhigh) as highly cost-efficient per Intelligence Index task at that $1.25/$4.25 sticker (~$0.55 per index task in their methodology), near Gemini 3.8 Flash’s efficiency band and cheaper per task than several peers at score 61.

Always re-check Meta’s live Model API pricing page. Contributor or discounted endpoints that train on your data are a different product decision than standard private traffic.

What Meta says changed

From Meta’s September 2 research post:

• Better long-horizon agentic work: build its own context, fix plan gaps, track what it learned, deliver an artifact.
• Trained across diverse harnesses to generalize beyond one agent UI.
• More collaborative: clarifying questions, asks for help when stuck, confirms before consequential actions; can update often or stay quiet based on preference.
• Stronger multi-tasking inside messy single threads.
• Better awareness of limits (less fake success theater).
• Coding: fewer unnecessary turns, less verbose, cleaner style; Meta eng comparisons cite ~20% fewer tool calls and ~25% fewer tokens vs 1.2.
• Safety: stronger adversarial / prompt-injection resistance; better judgment on irreversible actions in agent loops.

Independent snapshot

Artificial Analysis (Sep 2, 2026 article):

• Intelligence Index: xhigh 61 (up from 1.2 at 57), max 62 (limited preview).
• Peers at 61 include GPT-5.6 Sol (max) and Grok 4.6 (high); Claude Fable 5.1 (max) still higher on that index.
• Agentic gains called out on Tau3-Bench Banking, Terminal-Bench 2.1, GDPval-AA v2.
• Minor regressions noted on some knowledge/long-context style evals (AA-LCR, AA-Omniscience accuracy with higher abstention).

Use that as a map, then run your own Muse Code or API harness.

Who should try it

Try Muse Spark 1.3 when:

• You want a Meta-native coding agent loop (Muse Code) or API model for long agents.
• You care about agent cost/performance and already compare Flash-class and mid-frontier APIs.
• You need multimodal inputs (image/video) in the same agent thread.

Wait or skip when:

• You need open weights or on-prem Llama-class deployment.
• You need max-mode quality today and are not a preview partner.
• Your stack is standardized on Claude Code / Cursor / Codex and switching harnesses is the expensive part.

Same-day peer: Gemini 3.8 Flash review. Head-to-head: compare guide.