Muse Spark 1.3 reviewed: benchmarks, pricing, API availability, and what changed from 1.2

Source check: October 13, 2026. Muse Spark 1.3 launched September 2, and max reasoning is now available according to Meta. Standard and Contributor prices are both directly listed for version 1.3. ¹²³

The most consequential change is not a single leaderboard score. Meta has trained 1.3 around the failure modes of long-running agents: maintaining multiple workstreams in one thread, keeping complex constraints intact, asking for clarification when needed, recovering from planning gaps, and using fewer unnecessary tool calls. In comparisons run by Meta engineers, 1.3 used roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. That is a vendor-side comparison, not an independent benchmark. ¹

The benchmark picture needs similar care. Several launch scores came from Meta’s own evaluation runs and had not yet appeared on the corresponding public third-party leaderboards when this research was completed. Meta’s current Muse Spark 1.3 model card now explicitly lists 1,754 GDPval-AA v2 Elo. The September recheck did not show a corresponding independent Artificial Analysis 1.3 entry, so the number remains a Meta-reported result, not independent replication. ⁴³

Muse Spark 1.3 at a glance

• Release: September 2, 2026. ¹
• Available through: Muse Code and Meta Model API. ¹
• Max reasoning: currently available through Muse Code and Meta Model API. ¹³
• Primary focus: coding, long-horizon agentic work, multitasking, and instruction following. ¹
• GDPval-AA v2: 1,754 Elo is now directly listed on Meta’s 1.3 model card, but that is not independent verification. ³
• DeepSWE v1.1: launch reporting cites 75.4%, but Datacurve did not yet show a visible 1.3 entry in the parsed public leaderboard at research time. ⁵⁶
• Standard API: $1.25 input / $0.15 cached input / $4.25 output per million tokens, directly confirmed by Meta. ²
• Contributor API 1.3: $0.10 input / $0.002 cached input / $0.20 output; requests may be used by Meta for training. ²
• Open weights: planned, with no launch date or license disclosed. ¹

What is Muse Spark 1.3?

Muse Spark is Meta’s proprietary multimodal reasoning and agentic model family. The original Muse Spark launched on April 8, 2026; version 1.1 followed on July 9, version 1.2 on August 5, and 1.3 on September 2. ⁷⁸⁹

Muse Spark 1.1 pushed the family toward tool use, computer control, coding, and an actively managed one-million-token context. Version 1.2 was explicitly coding-focused and co-trained with Muse Code for long repository-scale projects. ⁷⁸

Version 1.3 focuses on making those agent workflows behave better over time:

1. Longer-horizon execution. Meta says the model can maintain an open-ended objective, gather its own context with tools, and repair gaps in its plan. ¹
2. More explicit collaboration. It is trained to ask clarifying questions, request help when blocked, and confirm before consequential actions. ¹
3. Better constraint retention. Meta says 1.3 is less likely to lose detailed workflow requirements during multi-step tasks. ¹
4. More efficient coding loops. Meta engineer comparisons report about 20% fewer tool calls and 25% fewer tokens versus 1.2. ¹

The last point is potentially more important in production than a small benchmark gain. Fewer calls can reduce latency, cost, and the number of opportunities for a tool-using agent to go off track. Independent same-task measurements are still needed.

Benchmark snapshot — and the evidence status behind each number

Meta’s methodology document is unusually useful because it explains the harnesses and scoring. It also contains a critical caveat: third-party model runs are best-effort, and prompts, tools, and runtimes may not be optimized for proprietary competitors. Depending on the benchmark, Meta selects comparable results from its own runs, official leaderboards, or provider-reported results. ¹⁰

Benchmark	Muse Spark 1.3	Evidence and October 13 source check	What it measures
GDPval-AA v2	1,754 Elo	Meta model card; not independently replicated	professional work products
JobBench	64.9	launch-reported	practical tasks across 35 occupations
OSWorld 2.0	66.9	launch-reported	long-horizon desktop workflows
DeepSearchQA	90.3 F1	Meta model card, not independently rerun	multi-step web research
AutomationBench	49.6%	Meta model card, not a Mac benchmark	tool-based business automation
DeepSWE v1.1	75.4%	Meta run; public leaderboard not yet synchronized	long-horizon software engineering
SWE Atlas Codebase QnA	59.4	Meta run; Scale page not yet synchronized	deep codebase comprehension
Terminal-Bench 2.1	88.8%	launch-reported	terminal-based agent tasks
MRCR v2 256K–512K	98.5	launch-reported	long-context retrieval
MRCR v2 512K–1M	98.1	launch-reported	long-context retrieval
Agentic IF Index	57.8	Meta internal	complex instruction adherence

The non-GDPval scores above are drawn from launch reporting and cross-checked against Meta’s methodology rather than treated as independent reproductions. ¹⁰⁵⁴

GDPval-AA v2: strong, but not the overall leader

Meta’s official Muse Spark 1.3 model card now lists 1,754 Elo on GDPval-AA v2 directly. This verifies the manufacturer’s claim, not independent reproduction. At the original September 2 check, no corresponding 1.3 entry was visible on Artificial Analysis. ⁵⁴

GDPval-AA v2 is useful because it is not a multiple-choice quiz. It uses 220 tasks sampled from OpenAI’s GDPval benchmark, which spans 44 occupations and nine major U.S. industries. Agents create real deliverables such as documents, spreadsheets, slide decks, and diagrams in an agentic harness. ⁴¹¹

That makes the score relevant to real agent work, but it is still harness- and judge-dependent, and rankings can change as new models are added.

DeepSWE: why 75.4% is not yet an independently confirmed first place

Launch reporting gives Muse Spark 1.3 a 75.4% DeepSWE v1.1 score. Meta’s methodology says it ran Muse Spark 1.3 max using a mini-swe agent while comparison results for other models were taken from the official Datacurve leaderboard. ¹⁰⁵

When checked on September 2, Datacurve’s parsed public v1.1 leaderboard did not yet show a visible Muse Spark 1.3 entry. It showed Gemini 3.8 Flash high at roughly 74%, Claude Opus 5 max at roughly 74%, and GPT-5.6 Sol max at roughly 73%. ⁶

That does not falsify Meta’s 75.4% run. It means the responsible wording today is “Meta reports 75.4%”, not “Muse Spark 1.3 is independently verified as the DeepSWE leader.”

SWE Atlas: the scaffold is part of the system

The 59.4 launch score on SWE Atlas Codebase QnA also had not appeared on Scale’s visible public leaderboard at research time. ¹² More importantly, Scale’s own analysis shows why scaffold details matter: models in native coding harnesses can explore, search, and execute 1.5–2 times more than the same model in generic scaffolds, with meaningfully higher scores. ¹³

Coding-agent benchmarks often measure model + scaffold + tools + reasoning budget, not a model in isolation. Any “Model A beats Model B” headline that ignores the harness can therefore be misleading.

Long context: excellent MRCR retrieval is not the same as general million-token reasoning

Launch reporting lists 98.5 on MRCR v2 at 256K–512K tokens and 98.1 at 512K–1M. ⁵ Meta’s methodology explains that it uses 100 examples per band, the 8-needle variant, and rule-based grading without agent tools. ¹⁰ OpenAI’s MRCR materials describe the benchmark as long-context retrieval rather than a full test of arbitrary long-context reasoning. ¹⁴

These are strong retrieval results. They should not be translated into the broader claim that every million-token repository, research corpus, or multi-document synthesis task will be nearly error-free.

The 1.3 API specification is now explicit: Meta’s coding-agent quickstart documents 1,048,576 context tokens and 131,072 output tokens for muse-spark-1.3. Long-context retrieval benchmarks still do not guarantee accuracy on arbitrary full-window agent tasks. ¹⁵

Muse Spark 1.3 prices: confirmed Meta API tiers, no 1.2 extrapolation

Meta pricing and rate limits rechecked October 13, 2026. Meta now directly lists both Muse Spark 1.3 model IDs in its first-party price schedule and model catalog. These are U.S. dollar prices per million tokens for the Meta Model API, not an assumed 1.2 price or a quote for every third-party router. ²³

First-party API model ID	Input	Cached input	Output	How Meta uses requests
muse-spark-1.3	$1.25	$0.15	$4.25	Meta says not used for model training
muse-spark-1.3-contributor	$0.10	$0.002	$0.20	Inputs and completions may train or improve Meta models

Contributor is a data-use tradeoff, not a normal volume discount. Do not route customer records or proprietary repositories through it without a deliberate data governance review. Standard-tier training exclusions do not, by themselves, establish retention, geographic residency, or DPA terms. ²

Three direct 1.3 cost examples

The following are our calculations using the published 1.3 rate card and no caching or paid tools:

Scenario	Input	Output	1.3 Standard	1.3 Contributor
Light	1M	0.2M	$2.10	$0.14
Medium	10M	2M	$21.00	$1.40
Intensive	100M	20M	$210.00	$14.00

For example: Standard 10 × $1.25 + 2 × $4.25 = $21.00. The Contributor rate is exactly 15× cheaper at this specific 5:1 input/output mix; a different mix, cache hit ratio or external tooling changes the outcome.

Actual cached input: A request involving 10M total input tokens, with 8M verified cache hits, and 2M output costs $12.20 Standard or $0.616 Contributor. Formula: (total input − cached input) × uncached rate + cached input × cached rate + output × output rate. Check cached_tokens in the response: the same tokens must not be charged in both input categories. Meta lists no long-context surcharge within the model window. ²

Rate limits, search fees and billing boundaries

Meta lists 3,000 RPM / 4,000,000 TPM for Standard, compared with 150 RPM / 3,000,000 TPM for Contributor, shared per team rather than per API key. A separate default Standard limit of 600 background submissions/minute applies. Respect remaining-quota response headers and use exponential backoff with jitter for HTTP 429. ²

Integrated web_search grounding costs $2.50 per 1,000 searches on top of token charges. One hundred tool-initiated searches add approximately $0.25. These are first-party U.S. dollar list prices; state/local sales tax and third-party routing charges may be additional. ²

Mac API setup: cloud inference, not an on-device model

macOS setup: Install the OpenAI-compatible Python SDK in a virtual environment, not system Python. Store MODEL_API_KEY only in your local shell or a secret manager. Never paste credentials into a public repository or browser-side JavaScript. The request below may incur token charges.

python3 -m venv .venv-meta-spark
source .venv-meta-spark/bin/activate
python -m pip install --upgrade openai

Meta’s current quickstart documents https://api.meta.ai/v1, MODEL_API_KEY, and model muse-spark-1.3. A minimal reproducible test through the OpenAI Responses SDK is:

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", input="List three ways to test a coding agent.")
print(response.model_dump_json(indent=2))

The Meta agent quickstart specifies a 1,048,576-token context and a 131,072-token output limit. For agentic tool loops, it recommends carrying encrypted reasoning items (for example include: ["reasoning.encrypted_content"]) through a compatible Responses client. OpenAI compatibility is endpoint-dependent; this is remote inference from a Mac, not a release of runnable Mac-local Muse Spark 1.3 weights. ¹⁵

Muse Spark 1.3 vs. 1.2

Muse Spark 1.2 was already trained for coding and long-running repository-scale tasks, with planning, context compaction, and persistent background agents in Muse Code. ⁷ Version 1.3 looks less like a new category and more like a reliability and efficiency upgrade for those workflows.

Area	Muse Spark 1.2	Muse Spark 1.3
Coding focus	strong, co-trained with Muse Code	more long-horizon coding training
Tool use	agentic workflows	Meta reports fewer unnecessary calls
Token efficiency	baseline	Meta reports ~25% fewer tokens
Multitasking	supported	explicitly improved
Clarification/confirmation	agentic	stronger user collaboration and consequential-action checks
Instruction following	improved over 1.1	another explicit focus
Open weights	unavailable	announced, not yet released

For a new agentic coding workload, 1.3 is the obvious version to test if it is available at the same rate. Existing production agents still need regression testing: a model that asks more clarifying questions, changes tool-call frequency, or becomes terser can alter a carefully tuned agent loop.

Is Muse Spark 1.3 better than GPT-5.6 Sol or Claude Opus 5?

There is no defensible one-word answer yet.

On GDPval-AA v2, Muse Spark 1.3’s 1,754 Elo trails Claude Opus 5 max at 1,824. ⁵⁴ Launch reporting gives Muse Spark 1.3 90.3 on DeepSearchQA versus 93.1 for GPT-5.6 and 57.8 on Meta’s Agentic IF Index versus 60.5 for GPT-5.6. ⁵

The model looks much stronger on some coding measurements, including Meta’s DeepSWE and Terminal-Bench runs. But because harnesses differ and several launch results have not yet synchronized to public leaderboards, “better” is currently workload-specific.

For an engineering team, the more useful comparison is:

• success rate on your own repository tasks;
• human interventions per task;
• tool calls and tokens to completion;
• incomplete-edit and regression rate;
• wall-clock time;
• effective cost after cache reuse;
• data-governance compatibility.

What is still unverified in October 2026

Meta’s current Muse Spark 1.3 card now directly documents the 1M-token context, both 1.3 pricing tiers, launch benchmark scores and availability of max reasoning. Those are no longer hypothetical 1.2 carryovers. However, the following remain unresolved or inadequately disclosed:

• Exact model parameters and detailed architectural specification.
• Actual open-weight release date and license for the announced future Muse Spark weights.
• Independent reproduction of all vendor-evaluated agent benchmarks on a matching harness.
• Task-level reliability and operational cost for a specific Mac-driven coding-agent workflow.

Evidence note: Meta now puts 1,754 GDPval-AA v2 Elo on its own current model page. This is directly sourced to Meta, not independently replicated proof of that score on the current Artificial Analysis board. The 75.4% DeepSWE figure also remains a Meta-evaluated result. ³

Where Muse Spark 1.3 fits today

Versus the newer Muse Wonder: Launched October 12, Wonder is a different hosted model. Wonder Full is gated behind Muse Code Power/Maximum, rather than a public pay-as-you-go API. For 10M uncached input and 2M output tokens, Spark 1.3 Standard costs $21 and Wonder Standard $80 at Meta’s published rates. These are list-price calculations before cache, tools and tax, not evidence that one model is better. Spark 1.3 Contributor is cheaper in exchange for permission to use content in training; Wonder has no Contributor Model API tier. Meta rate card · Muse Wonder.

Muse Spark 1.3 remains principally a hosted agent/coding model, not an on-device Mac model. Meta’s current first-party card now explicitly confirms its 1.3 prices, max reasoning, and vendor-evaluated GDPval-AA v2, DeepSWE, long-context retrieval and Terminal-Bench results. That is substantially better evidence than the launch-day price inference, but does not establish independent replication. ³

The 1,754 GDPval-AA-v2 Elo and 75.4% DeepSWE figures are Meta-reported results on a specified evaluation harness. The very cheap 1.3 Contributor tier is directly available, but its request data can be used to improve Meta models. It is not a safe default for sensitive proprietary code simply because the 5:1 example bill is 15× lower. ²³

Compare 1.3 against 1.2, Claude or GPT using identical prompts, harness, tool permissions, context budget and evaluation rules. Measure uncached tokens, actual cached tokens, output, latency, failures and any paid web-search calls separately.

Further reading: Muse Spark 1.1 and the first-party Meta API quickstart.

Source notes

Footnotes

1. Meta: introduction ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

2. Meta: pricing ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

3. Meta: model card ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

4. Artificial Analysis ↩ ↩² ↩³ ↩⁴ ↩⁵

5. Meta: benchmarks ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

6. DeepSWE ↩ ↩²

7. Meta: Muse 1.2 ↩ ↩² ↩³

8. Meta: API launch ↩ ↩²

9. Meta: background ↩

10. Meta: methodology ↩ ↩² ↩³ ↩⁴

11. OpenAI: GDPval ↩

12. Scale: SWE Atlas ↩

13. Scale: SWE Atlas ↩

14. MRCR dataset ↩

15. Meta: API quickstart ↩ ↩²