What Is Meta Muse Spark 1.3? Benchmarks and Pricing
The Short Answer
Muse Spark 1.3 is Meta’s frontier-class model for long-horizon agentic work, released September 2, 2026. It costs $1.25 per million input tokens and $4.25 per million output tokens, carries a 1 million token context window, and is available in Muse Code and the Meta Model API.
What makes it interesting:
- The best long-context retrieval published — 98.1% on MRCR 512K–1M, where GPT-5.6 Sol manages 73.8%.
- It beats Claude Opus 5 on agentic coding on Meta’s own table, at roughly a quarter of Opus 5’s input price.
- It is the value pick on the independent coding-agent leaderboard, at $1.72 per task.
- Two things are missing: max reasoning mode, and the promised open weights.
Last verified: September 4, 2026.
Specifications
| Muse Spark 1.3 | |
|---|---|
| Lab | Meta |
| Released | September 2, 2026 |
| Input / MTok | $1.25 |
| Cached input / MTok | ~$0.15 |
| Output / MTok | $4.25 |
| Context window | 1,000,000 tokens |
| Input modalities | Text |
| Available in | Muse Code, Meta Model API |
| Open weights | Roadmap, not shipped |
| Max reasoning mode | In safety testing |
| Predecessor | Muse Spark 1.2 (Aug 5, 2026) |
Pricing is listed by model trackers including Benchable, LMMarketCap and Kingy.ai rather than published on a prominent official Meta pricing page, so confirm against your own billing. One tracker shows a “starts at” tier of $0.10/$0.20 that likely reflects a batch or contributor tier — Muse Spark 1.2 carried a contributor tier at $0.10/$0.20 where data is used for training.
The Benchmarks
These are Meta’s published figures from the September 2, 2026 release and evaluation report.
| Benchmark | Muse Spark 1.3 | Claude Opus 5 | GPT-5.6 Sol | Muse Spark 1.2 |
|---|---|---|---|---|
| DeepSWE v1.1 (agentic coding) | 75.4% | 74.0% | 73.0% | — |
| Terminal-Bench 2.1 | 88.8% | — | 88.8% (tie) | — |
| SWEAtlas CodeBase QnA | 59.4% | — | — | — |
| MRCR 256K–512K | 98.5% | not posted | 91.5% | 66.3% |
| MRCR 512K–1M | 98.1% | not posted | 73.8% | 55.5% |
| OSWorld 2.0 (partial) | 66.9% | 75.4% | — | — |
| GDPval-AA v2 | 1754 | 1824 | 1711 | — |
| AutomationBench | ~49% | 26.9% | 19.6% | — |
The MRCR numbers are the story. Long-context retrieval degrades sharply for most models as the window fills — GPT-5.6 Sol drops from 91.5% at 256K–512K to 73.8% at 512K–1M. Muse Spark 1.3 goes 98.5% to 98.1%, essentially flat. Its own predecessor managed 66.3% and 55.5%. That is not a tuning gain; it is a different capability class, and it is the single strongest argument for using this model.
DeepSWE v1.1 is the credibility check. At 75.4% it edges past Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on a long-horizon agentic benchmark covering 113 tasks across 91 repositories and five languages. Beating a $5/$25 frontier model at $1.25/$4.25 is a real result.
OSWorld 2.0 is the honest weak spot. At 66.9% partial accuracy it trails Opus 5’s 75.4% and Claude Fable 5.1’s 77.9%. Computer use is not where this model leads.
What “Long-Horizon” Actually Means Here
Meta’s research blog describes behavioural training rather than raw capability scaling. Muse Spark 1.3 is trained to:
- Juggle multiple workflows in one long thread without losing track of parallel objectives.
- Ask clarifying questions when a prompt is ambiguous, instead of guessing and burning a long run on the wrong task.
- Confirm before consequential actions, which matters when an agent has shell or file access.
- Track what it has already learned, reducing redundant re-derivation.
- Resist prompt injection more effectively than 1.2.
The measured payoff: Meta engineers report roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 for equivalent work. Fewer tool calls is the underrated metric — every tool call is latency, an error surface, and re-read context.
The Independent Verdict
Artificial Analysis titled its coverage “Muse Spark 1.3: Meta reaches the frontier” and scored it around 62.0 on the Intelligence Index composite — ahead of Kimi K3 (59.6), GLM 5.3 (59.4 at max) and Qwen 3.8 Max (57.7), though behind Claude Fable 5.1 (65.6) and Claude Opus 5 (63.0).
On the AA Coding Agent Index it is the value standout:
| Agent configuration | Index | Cost per task |
|---|---|---|
| Codex · GPT-6 Astra (low) | 62.6 | $1.41 |
| Muse Code · Muse Spark 1.3 (xhigh) | 64.2 | $1.72 |
| Opencode · Gemini 3.8 Flash (high) | 61.1 | $2.04 |
| Codex · GPT-6 Astra (xhigh) | 67.0 | $3.27 |
| Claude Code · Fable 5.1 (max) | 70.4 | $9.18 |
It beats Gemini 3.8 Flash on quality and cost per task — the only model on that table to do so from the mid-price tier.
What Is Missing
Max reasoning mode is still in safety testing. The top effort level is unavailable, so today’s numbers are not the ceiling — and any benchmark comparison against a rival’s max-effort result is not apples to apples.
Open weights have not shipped. Meta’s roadmap includes an open-weights release, and that would reset build-versus-buy math for self-hosted AI. It has not happened. Until it does, if self-hosting or data residency drives your decision, the MIT-licensed Chinese frontier models — DeepSeek V4, Kimi K3, GLM 5.3 — are the shipping options, not Muse Spark.
Text input only. No image, audio or video, unlike Gemini 3.8 Flash.
No independent per-benchmark verification at launch. Meta published its own evaluation report; treat vendor tables as directional.
Who Should Use It
Use Muse Spark 1.3 if you run long agent missions over large contexts — codebase-wide refactors, multi-document research sweeps, extended terminal automation. The MRCR results and the tool-call reduction compound exactly there, and the price is a quarter of the flagship tier.
Do not use it if you need computer use (Opus 5 and Fable 5.1 lead), multimodal input (Gemini 3.8 Flash), open weights today (DeepSeek V4 or Kimi K3), or the absolute top of the coding leaderboard (Claude Fable 5.1 at 70.4).
Watch for two events: the max reasoning mode leaving safety testing, and the open-weights release. Either would materially change this recommendation.