Spark 1.3's index jump lands on the three tests that carry half the score
Meta's fourth Muse Spark in five months gains four points on Artificial Analysis' Intelligence Index, almost entirely in agentic tests, while two scores fall and the tier Meta benchmarked against rivals is still a limited preview.
The Engineer · Build desk
What happened
- Meta shipped Muse Spark 1.3 through Muse Code and the Meta Model API, its fourth model in five months after the April launch, version 1.1 in July and 1.2 in August.
- Artificial Analysis puts the available xhigh tier at 61 on its Intelligence Index, four points above 1.2 and level with GPT-5.6 Sol max, Grok 4.6 high and Claude Opus 5 high.
- Meta's own comparison table was run at a new max reasoning level that stays in limited preview pending further safety testing, while xhigh is the highest tier developers can call.
- Token prices held at $1.25 and $4.25 per million, yet the cost of one index task rose to $0.55 from $0.40 on version 1.2, still the cheapest of any model scoring 59 or higher.
- Two Artificial Analysis scores went backwards against 1.2: AA-LCR from 83 to 79 percent, and AA-Omniscience factual accuracy by up to three points as the model declines more questions.
Why it matters
- What it limits constraint Because half the index weight sits in three agentic tests, a 61 tells you how well the model drives tools and says little about the long-context retrieval work where this release regressed.
- What it costs cost The tier Meta benchmarked has no published price and consumes 62 percent more reasoning tokens than xhigh, so nobody outside the preview can put the headline numbers in a budget.
- Who has to decide decision With a new model inside a month, the durable investment is the eval harness and the tool schemas, not the vendor-specific glue that has to be rewritten when the ranking moves again.
- Who's exposed exposure Anyone planning to escape hosted-API pricing by self-hosting the promised open weights is exposed to licence terms Meta has not written down yet.
Next story · Product · 4d ago
Muse Wonder Full sits just below Claude Opus 5.5 on the Artificial Analysis index, Meta saysConfirmed25 publishers
Artificial Analysis weights GDPval-AA v2 at 20 percent of its Intelligence Index, Terminal-Bench 2.1 at 16 percent and tau3-Bench Banking at 14 percent, and Meta's largest gains landed in exactly those three tests [13]. Half the composite comes from those three tests [22]. The 61 is largely a statement about driving tools in a loop. On tau3-Banking, version 1.2 sat at 35 percent; the xhigh tier you can call today reaches 47, tying Claude Fable 5.1 max and GLM-5.3-Flash, and the preview max tier hits 52, which Artificial Analysis calls the only outright lead the model holds [14]. Terminal-Bench 2.1 climbs from 80 to 85 on xhigh [9]. Claude Fable 5.1 is still at 91.4 in its max tier, so the generally available Spark trails the leader by 6.4 points on the terminal coding test [21]. On the highest-weighted test, Meta moves 1,615 to 1,709 at xhigh against Claude Fable's 1,853, on a scale calibrated to human expert performance at 1,000 across 220 professional tasks [15] [23].
The budget arithmetic is cleaner. Input and output prices did not move: $1.25 and $4.25 per million tokens [5]. The cost of one index task did, from $0.40 on 1.2 to $0.55 [5]. Same per-token rate, 37.5 percent more money per task, which means roughly 37.5 percent more tokens burned per task [24]. Reasoning models buy scores with output. Rivals at the same index level run $0.94 to $1.23, so Spark lands between 1.7 and 2.2 times cheaper per task [12], and Artificial Analysis calls it the most cost-efficient model at its measured intelligence, on its Pareto line [20]. That is good engineering and worth saying plainly. Meta buys the max variant's single extra index point with 62 percent more reasoning tokens, which is one way to spend a point [10] [4].
What would have to be true for Meta's own table to describe your workload. The DeepSWE 75.4 percent and the 98.1 on the 512K-1M MRCR split were assembled by Meta, with rival figures drawn from a mix of its own evaluations, leaderboards and vendor-reported results, and its testing of third-party models described as best-effort [7] [8]. The headline comparisons ran at max, and xhigh is the highest level developers can call [3]. Your tasks would need to look like tool sequences against a verifier, and you would need access Meta has not opened. Meanwhile AA-LCR fell from 83 to 79 percent, and factual accuracy on AA-Omniscience slipped by up to three points because the model declines more often when unsure [17]. Long-context retrieval work gets a slightly worse model with a better leaderboard position.
The New Stack headlined its coverage with 1.3 edging out Gemini [19]. On GPQA Diamond, the same Artificial Analysis run has Gemini 3.8 Flash high at 95.3 against Muse Spark's 94 [16]. Both readings are available depending on which test you weight, which is the argument against standardizing a toolchain on any one vendor: 1.2 shipped in August, 1.3 arrived less than a month later, and Zuckerberg has already teased a larger model codenamed Watermelon [1] [2] [18]. The durable move is a vendor-neutral eval harness and tool schemas, with the model bill treated as a monthly line item rather than a fixed bet. The promised open-weight release would loosen the coupling further, but Meta has not published the licence, and its previous open-weight models carried varying restrictions on use [6].
What to watch
- The licence text accompanying the open-weight Spark release, which decides whether self-hosting is legally available at all.
- General availability and a published price for the max tier, without which Meta's headline comparison cannot be reproduced by customers.
- Any firm date for Watermelon, the larger model reportedly still in training in July.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence68
- Adoption25
- Hype gap+30
- Incentives65
- Confidence70
Perspective Coverage
6 publishers- Builder
- Builder 52%
- Operator
- Operator 26%
- Investor
- Investor 22%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Meta released Muse Spark 1.3 through Muse Code and the Meta Model API; it is the company's fourth model in five months, with the series launching in April, 1.1 following in July and 1.2 in August.
- [2]
Muse Spark 1.3 arrives less than a month after the release of Muse Spark 1.2.
- [3]
Meta ran Spark 1.3 at its new "max" reasoning level for the headline comparisons, while the highest reasoning level generally available to developers is "xhigh"; max remains in limited preview while Meta completes additional safety testing.
- [4]
Artificial Analysis scores the publicly available Muse Spark 1.3 xhigh at 61 on its Intelligence Index, four points ahead of Spark 1.2 and level with GPT-5.6 Sol max, Grok 4.6 high and Claude Opus 5 high; the limited-preview max version scored 62, behind only Claude Fable 5.1 and Claude Opus 5 in its comparison at launch.
- [5]
Muse Spark 1.3 pricing is unchanged at $1.25 and $4.25 per million input and output tokens, and one Intelligence Index task costs $0.55; no model scoring 59 or higher is cheaper, rivals at the same index level run between $0.94 and $1.23, and version 1.2 ran $0.40 per task.
- [6]
Zuckerberg confirmed open-weight releases of Muse Spark are "coming soon", but Meta has yet to publish the licence terms, and its previous open-weight models have carried varying restrictions on use.
- [7]
Meta's benchmark table pitted Muse Spark 1.3 against OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5; among the figures, Spark 1.3 scored 75.4% on the DeepSWE coding benchmark and 98.1% on the 512K-1M version of the MRCR long-context test.
- [8]
The comparison results were assembled by Meta: Spark 1.3 scores came via the Meta Model API while rival figures were drawn from a mix of Meta's own evaluations, official leaderboards and results reported by rival providers, with Meta describing its testing of third-party models as "best-effort" and the table not a single independent head-to-head under identical conditions.
- [9]
On Terminal-Bench 2.1, which tests coding in the terminal, Spark climbs from 80 to 85 percent on xhigh and 86 on max, while Claude Fable 5.1 holds the top spot at 91.4 percent in max, 91.0 at xhigh and 89.9 at high.
- [10]
Meta buys the max variant's edge with compute, burning 62 percent more reasoning tokens than xhigh, and neither Meta nor Artificial Analysis has named a price for the max variant.
- [11]
Mark Zuckerberg described Muse Spark 1.3 as delivering "frontier performance almost too cheap to meter" and wrote on X that it is "the biggest jump we've made so far on coding and agentic work".
- [12]
Rivals at the same index level cost between 1.7 and 2.2 times more per index task than Spark 1.3 xhigh.
- [13]
In the Intelligence Index, GDPval-AA v2 counts for 20 percent, Terminal-Bench 2.1 for 16 percent and tau3-Bench Banking for 14 percent, and Meta's biggest gains land in exactly those three tests.
- [14]
On tau3-Banking, where agents operate tools in a simulated banking scenario, Spark 1.3 max hits 52 percent, which Artificial Analysis says is number one and the model's only outright lead; the available xhigh tier reaches 47 percent, tying Claude Fable 5.1 (max) and GLM-5.3-Flash, while version 1.2 sat at 35 percent.
- [15]
On GDPval-AA v2, the index's highest-weighted test, Meta improves from 1,615 to 1,709 (xhigh) and 1,754 (max) on a scale calibrated to human expert performance at 1,000 across 220 real-world professional tasks; Claude Fable 5.1 (max) sits at 1,853.
- [16]
On GPQA Diamond, Muse Spark rises from 90 to 94 percent, still below Gemini 3.8 Flash (high) at 95.3 and Grok 4.6 (high) at 94.9.
- [17]
Two scores drop against version 1.2: AA-LCR falls from 83 to 79 percent, and factual accuracy in AA-Omniscience slips by up to three points because the model more often declines to answer when unsure.
- [18]
Zuckerberg teased Meta's next model, codenamed Watermelon, understood to be much larger than Spark and reportedly still in training in July according to a Business Insider report, with no firm release date.
- [19]
The New Stack headlined its coverage "Google was ahead only a few hours": Muse Spark 1.3 edges out Gemini as Meta claims its biggest coding leap yet.
- [20]
Artificial Analysis describes Spark 1.3 xhigh as the "most cost-efficient model" at its level of measured intelligence, placing it on the company's Pareto line.
- [21]
On Terminal-Bench 2.1 the generally available Spark 1.3 xhigh trails Claude Fable 5.1 max by 6.4 points.
- [22]
The three tests where Meta's gains concentrated account for 50 percent of the Intelligence Index weighting.
- [23]
On GDPval-AA v2 the xhigh tier trails Claude Fable 5.1 max by 144 points.
- [24]
Per-task cost rose 37.5 percent from version 1.2 to 1.3 while per-token prices stayed the same, implying roughly 37.5 percent more tokens consumed per index task.
Sources
6 independent publishers whose own reporting we read for this story.
- blog.vercel.comMuse Spark 1.3 now available on AI Gateway
1 article · September 1, 2026
- infoworld.comMeta upgrades Muse Spark for long-horizon coding without raising API prices | InfoWorld
1 article · September 3, 2026
- latent.space[AINews] Muse Spark 1.3 matches GPT-5.6-Sol, confirming Meta Superintelligence as the newest Frontier Lab, >90% discount for training
1 article · September 2, 2026
- runtimewire.comAlexandr Wang's Meta team ships Muse Spark 1.3 for coding agents
2 articles · September 4, 2026
- the-decoder.comMeta closes in on the top with Muse Spark 1.3, and undercuts rivals on price
2 articles · September 3, 2026
- thenewstack.io“Google was ahead only a few hours”: Muse Spark 1.3 edges out Gemini as Meta claims its biggest coding leap yet
2 articles · September 3, 2026