Meta says Muse Spark 1.3 has frontier functionality — however its highest effects come from a mannequin builders can’t extensively use but

Meta’s latest AI mannequin Muse Spark 1.3, unveiled yesterday, is quicker and extra performant on third-party benchmarks than its predecessor — with a caveat.

"Muse Spark 1.3 is rolling out as of late with frontier functionality nearly too reasonable to meter," Meta co-founder and CEO Mark Zuckerberg wrote on X, calling it Meta’s “greatest soar” but in coding and agentic paintings.

There may be substance in the back of each portions of that declare. Muse Spark 1.3 makes important positive aspects over last month’s 1.2 release, in particular on long-running agent duties. The model builders can get entry to now could also be probably the most most powerful price-performance choices close to the highest of impartial mannequin scores.

Meta’s most powerful Muse Spark 1.3 benchmark effects come from its max reasoning configuration. Meta says that model remains to be finishing further protection checking out and can arrive “in a while”; the third-party benchmarking company Artificial Analysis says it evaluated max in a restricted spouse preview, and these days lists no API supplier at keen on the configuration.

The model extensively rolling out this week thru its Muse Code harness and the Meta Model API makes use of Meta’s up to now to be had reasoning settings, together with xhigh.

That makes the extra related endeavor query no longer whether or not Muse Spark 1.3 can achieve frontier territory, however how shut the mannequin corporations can in fact deploy as of late will get — and at what actual price.

The transport mannequin is superb, however no longer the benchmark chief

Meta does reveal effects for each configurations in its underlying analysis document, so this isn’t a case of the corporate hiding the deployable mannequin. However its release fabrics prominently exhibit the max variant, and probably the most biggest rankings belong to that configuration.

For instance, Meta reviews GDPval-AA v2 rankings of one,754 Elo for optimum as opposed to 1,709 for xhigh, OSWorld 2.0 rankings of 66.9 as opposed to 57.2, and JobBench rankings of 64.9 as opposed to 61.2.

On some checks the glory is negligible or reversed: DeepSearchQA is tied at 89.4, whilst xhigh rankings 89.2 on Terminal-Bench 2.1 as opposed to max at 88.8.

Artificial Analysis rankings Muse Spark 1.3 max at 62 on its Intelligence Index and the transport xhigh model at 61. The latter ties GPT-5.6 Sol max, Grok 4.6 excessive and Claude Opus 5 excessive. However Anthropic nonetheless occupies the highest of the leaderboard: Claude Delusion 5.1 reaches 66 at max and 65 at xhigh, whilst Claude Opus 5 reaches 63 at max and xhigh.

In different phrases, Muse Spark 1.3 xhigh is legitimately within the frontier cluster, however it isn’t the mannequin these days environment the frontier.

This is nonetheless a considerable alternate from Muse Spark 1.2. VentureBeat’s protection of remaining month’s release discovered Meta fielding a reputable coding challenger that however most often trailed Anthropic’s highest mannequin. Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1 as opposed to Opus 5’s 86.7%, and in addition completed in the back of Opus at the different primary coding comparisons Meta introduced.

With 1.3, Meta is now not simply appearing up in that contest. On a number of coding and agentic opinions, it’s buying and selling wins with OpenAI and Anthropic.

Meta says the underlying mannequin has additionally change into more straightforward to perform. Muse Spark 1.3 is skilled to deal with a couple of workflows in an extended thread, acquire context with equipment, come across gaps in its personal plans, ask customers for rationalization when essential and ensure sooner than consequential movements. In Meta engineers’ interior comparisons, it used kind of 20% fewer instrument calls and 25% fewer tokens than 1.2 throughout coding paintings.

For enterprises paying for hundreds or tens of millions of agent loops, the ones behavioral enhancements may topic greater than any other leaderboard level.

‘Nearly too reasonable to meter’ does no longer imply Meta lower its costs

Muse Spark 1.3 didn’t obtain an API payment lower. Meta saved Same old pricing precisely the place it used to be for Muse Spark 1.2: $1.25 in line with million enter tokens, $4.25 in line with million output tokens and $0.15 in line with million cached enter tokens.

That makes Zuckerberg’s “nearly too reasonable to meter” line much less a commentary about decrease token costs than about what Meta believes builders can accomplish with the ones tokens.

Artificial Analysis provides proof for that argument, but in addition a complication. It measures Muse Spark 1.3 xhigh at 235.2 output tokens in line with 2d and estimates a price of $0.55 in line with Intelligence Index process.

At 61 at the Intelligence Index, that provides it the lowest price in line with process of any these days measured mannequin at that intelligence degree.

Muse Spark 1.2 price simplest $0.40 in line with Synthetic Research process, whilst scoring 57.

Regardless of unchanged per-token pricing, the impartial benchmark’s price of finishing a mean process subsequently greater generation-over-generation. Synthetic Research attributes the rise essentially to heavier input-token intake on agentic opinions.

That does indirectly contradict Meta’s declare of 25% decrease token use: Meta is describing comparisons in its personal coding workflows, whilst Synthetic Research is measuring a broader suite of reasoning and agentic duties.

However it illustrates why “reasonable” turns into slippery as soon as fashions perform as brokers. Token charges, reasoning effort, choice of turns, instrument calls and retries all give a contribution to the true price of completing paintings.

Meta additionally keeps its surprisingly reasonable Contributor tier — $0.10 in line with million enter tokens and $0.20 in line with million output tokens — in change for permission to make use of activates and completions for coaching.

As VentureBeat famous with Muse Spark 1.2, that can be horny for prototyping however creates a materially other data-governance calculation for enterprises operating with proprietary code or delicate interior data.

Wang’s ‘Gemini who?’ lands on an surprisingly shut comparability

Meta leader AI officer Alexandr Wang used to be significantly much less certified in celebrating the discharge.

After Synthetic Research posted its Muse Spark effects, Wang reposted them on X, adding: “i in point of fact hate to mention it, however… gemini who? 😱💨”

The colour used to be in particular pointed as a result of Google released Gemini 3.8 Flash at the identical day, pitching it at nearly precisely the similar elegance of workload: long-horizon instrument engineering, self sufficient brokers and multi-step skilled reasoning. Google calls 3.8 its highest reasoning and coding Flash mannequin but and says it’s the corporate’s 1/3 Flash unencumber in six weeks.

Unbiased numbers give Wang one thing to paintings with, although hardly ever a knockout.

Synthetic Research provides Muse Spark 1.3 xhigh a 61 Intelligence Index rating at $0.55 in line with process, in comparison with 59 and $0.58 for Gemini 3.8 Flash at excessive reasoning. Meta subsequently edges Google on each intelligence and process price at the ones specific settings.

Google wins decisively on throughput. Synthetic Research measures Gemini 3.8 Flash excessive at about 305 output tokens in line with 2d, as opposed to 235 for Muse Spark — kind of 30% sooner. Gemini additionally has the decrease uncooked API decal payment for now: Google is charging an introductory $0.75 in line with million enter tokens and $3.75 in line with million output tokens, in comparison with Meta’s $1.25 and $4.25.

That promotional Google pricing expires December 31, and then it rises to $1.50 in line with million enter tokens and $7.50 in line with million output tokens.

The outcome is an invaluable snapshot of the way tight frontier-model economics have change into. Meta these days wins this impartial comparability by means of two Intelligence Index issues and 3 cents in line with benchmark process; Google provides considerably upper output throughput and less expensive uncooked tokens throughout its release promotion.

Wang’s “gemini who?” is amusing government trash communicate. For an endeavor architect, the solution is nearer to: Gemini is the speedier possibility; Muse is these days the reasonably more potent high-effort agent by means of this impartial measure.

Meta's evolving open weights stance

The extra consequential factor for some builders could have little to do with as of late’s benchmark race.

When Meta introduced Muse Code and Muse Spark 1.2 in August, VentureBeat noted how dramatically the corporate had moved clear of the open-weight technique that made Llama ubiquitous.

Muse Code and Spark 1.2 had been proprietary, API-served merchandise — a putting posture for the corporate that had spent years arguing that open AI used to be the trail ahead.

5 days later, Meta modified route once more.

On August 10, it launched the 30-billion-parameter Muse Glimmer under an Apache 2.0 license. Zuckerberg additionally mentioned: “Within the coming weeks, we also are going to open the weights for Muse Spark 1.2.” Reuters one at a time reported Meta’s plan to unencumber the Spark 1.2 weights.

Now, Meta has as a substitute shipped Muse Spark 1.3 as any other proprietary mannequin.

That doesn’t but quantity to a damaged promise — “coming weeks” can fairly describe a length longer than 3 weeks. However as of late’s announcement makes the roadmap much less transparent slightly than extra.

Meta’s new put up now not says Muse Spark 1.2. It says its roadmap contains “the Muse Spark open weights unencumber”, with out figuring out a model, unencumber date, mannequin dimension or license. Zuckerberg likewise mentioned on X that “Muse Spark open weights releases” are coming quickly.

For groups that standardized on Llama as a result of downloadable weights supposed self-hosting, customization and regulate over inference economics, that ambiguity might topic greater than whether or not Spark received any other level on a composite benchmark.

Muse Spark 1.3 displays that Meta can now iterate proprietary frontier fashions at abnormal velocity. The transport xhigh configuration is speedy, competitively priced and far nearer to the highest of impartial scores than its predecessors. The max preview displays Meta can push the circle of relatives a bit additional when allowed to spend extra reasoning compute.

The following check is other: whether or not Meta can convert that tempo right into a roadmap enterprises can in fact plan round — together with making its highest features extensively deployable and turning in the open-weight Spark mannequin it has already mentioned is coming.