Meta says Muse Spark 1.3 has frontier efficiency — however its finest outcomes come from a mannequin builders can’t broadly 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 at the moment with frontier efficiency nearly too low-cost to meter," Meta co-founder and CEO Mark Zuckerberg wrote on X, calling it Meta’s “greatest bounce” but in coding and agentic work.
There is substance behind each components of that declare. Muse Spark 1.3 makes vital beneficial properties over last month’s 1.2 release, notably on long-running agent duties. The model builders can entry now can be one of many strongest price-performance choices close to the highest of unbiased mannequin rankings.
Meta’s strongest Muse Spark 1.3 benchmark outcomes come from its max reasoning configuration. Meta says that model remains to be finishing further security testing and can arrive “shortly”; the third-party benchmarking agency Artificial Analysis says it evaluated max in a restricted accomplice preview, and at the moment lists no API supplier in any respect for the configuration.
The model broadly rolling out this week via its Muse Code harness and the Meta Model API makes use of Meta’s beforehand out there reasoning settings, together with xhigh.
That makes the extra related enterprise query not whether or not Muse Spark 1.3 can attain frontier territory, however how shut the mannequin corporations can really deploy at the moment will get — and at what actual value.
The delivery mannequin is superb, however not the benchmark chief
Meta does disclose outcomes for each configurations in its underlying analysis report, so this isn’t a case of the corporate hiding the deployable mannequin. But its launch supplies prominently showcase the max variant, and a number of the largest scores belong to that configuration.
For instance, Meta reviews GDPval-AA v2 scores of 1,754 Elo for max versus 1,709 for xhigh, OSWorld 2.0 scores of 66.9 versus 57.2, and JobBench scores of 64.9 versus 61.2.
On some assessments the excellence is negligible or reversed: DeepSearchQA is tied at 89.4, whereas xhigh scores 89.2 on Terminal-Bench 2.1 versus max at 88.8.
Artificial Analysis scores Muse Spark 1.3 max at 62 on its Intelligence Index and the delivery xhigh model at 61. The latter ties GPT-5.6 Sol max, Grok 4.6 excessive and Claude Opus 5 excessive. But Anthropic nonetheless occupies the highest of the leaderboard: Claude Fable 5.1 reaches 66 at max and 65 at xhigh, whereas 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 at the moment setting the frontier.
That remains to be a considerable change from Muse Spark 1.2. VentureBeat’s protection of final month’s launch discovered Meta fielding a reputable coding challenger that nonetheless usually trailed Anthropic’s finest mannequin. Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1 versus Opus 5’s 86.7%, and in addition completed behind Opus on the opposite major coding comparisons Meta offered.
With 1.3, Meta is now not merely exhibiting up in that contest. On a number of coding and agentic evaluations, it’s buying and selling wins with OpenAI and Anthropic.
Meta says the underlying mannequin has additionally change into simpler to function. Muse Spark 1.3 is educated to keep up a number of workflows in an extended thread, collect context with instruments, detect gaps in its personal plans, ask customers for clarification when mandatory and make sure earlier than consequential actions. In Meta engineers’ inside comparisons, it used roughly 20% fewer device calls and 25% fewer tokens than 1.2 throughout coding work.
For enterprises paying for hundreds or thousands and thousands of agent loops, these behavioral enhancements might matter greater than one other leaderboard level.
‘Almost too low-cost to meter’ doesn’t imply Meta reduce its costs
Muse Spark 1.3 didn’t obtain an API worth reduce. Meta stored Standard pricing precisely the place it was for Muse Spark 1.2: $1.25 per million enter tokens, $4.25 per million output tokens and $0.15 per million cached enter tokens.
|
Model |
Input ($/1M) |
Output ($/1M) |
Total ($/1M) |
Source |
|
Muse Spark 1.2 / 1.3 Contributor |
$0.10 |
$0.20 |
$0.30 |
|
|
MiMo-V2.5 Flash |
$0.10 |
$0.30 |
$0.40 |
|
|
DeepSeek-V4-Flash — off-peak |
$0.22 |
$0.66 |
$0.88 |
|
|
GPT-5.6 Luna |
$0.20 |
$1.20 |
$1.40 |
|
|
MiniMax-M3 |
$0.30 |
$1.20 |
$1.50 |
|
|
LongCat-2.0 — limited-time promo |
$0.30 |
$1.20 |
$1.50 |
|
|
DeepSeek-V4-Flash — peak hours |
$0.44 |
$1.32 |
$1.76 |
|
|
MiMo-V2.5 |
$0.40 |
$2.00 |
$2.40 |
|
|
DeepSeek-V4-Pro — off-peak |
$0.66 |
$1.98 |
$2.64 |
|
|
LongCat-2.0 — normal |
$0.75 |
$2.95 |
$3.70 |
|
|
MiMo-V2.5 Pro (≤256K) |
$1.00 |
$3.00 |
$4.00 |
|
|
Gemini 3.7 Flash — via Dec. 31, 2026 |
$0.75 |
$3.75 |
$4.50 |
|
|
Gemini 3.8 Flash — via Dec. 31, 2026 |
$0.75 |
$3.75 |
$4.50 |
|
|
DeepSeek-V4-Pro — peak hours |
$1.32 |
$3.96 |
$5.28 |
|
|
Muse Spark 1.1 / 1.2 / 1.3 |
$1.25 |
$4.25 |
$5.50 |
|
|
GLM-5.3 |
$1.40 |
$4.40 |
$5.80 |
|
|
Grok 4.6 — <200K immediate tokens |
$2.00 |
$6.00 |
$8.00 |
|
|
MiMo-V2.5 Pro (>256K) |
$2.00 |
$6.00 |
$8.00 |
|
|
Qwen3.8-Max |
$2.00 |
$6.00 |
$8.00 |
|
|
Gemini 3.7 Flash — beginning Jan. 1, 2027 |
$1.50 |
$7.50 |
$9.00 |
|
|
Gemini 3.8 Flash — beginning Jan. 1, 2027 |
$1.50 |
$7.50 |
$9.00 |
|
|
GPT-5.6 Terra |
$2.00 |
$12.00 |
$14.00 |
|
|
Grok 4.6 — ≥200K immediate tokens |
$4.00 |
$12.00 |
$16.00 |
|
|
GPT-5.4 |
$2.50 |
$15.00 |
$17.50 |
|
|
Kimi K3 |
$3.00 |
$15.00 |
$18.00 |
|
|
Claude Opus 5 |
$5.00 |
$25.00 |
$30.00 |
|
|
Sakana Fugu Ultra (≤272K) |
$5.00 |
$30.00 |
$35.00 |
|
|
GPT-5.6 Sol — Standard mode |
$5.00 |
$30.00 |
$35.00 |
|
|
Claude Fable 5 / Claude Mythos 5 |
$10.00 |
$50.00 |
$60.00 |
|
|
Claude Fable 5.1 / Claude Mythos 5.1 |
$10.00 |
$50.00 |
$60.00 |
|
|
GPT-5.6 Sol — Fast mode |
$10.00 |
$60.00 |
$70.00 |
That makes Zuckerberg’s “nearly too low-cost to meter” line much less a press release about decrease token costs than about what Meta believes builders can accomplish with these tokens.
Artificial Analysis affords proof for that argument, but additionally a complication. It measures Muse Spark 1.3 xhigh at 235.2 output tokens per second and estimates a value of $0.55 per Intelligence Index process.
At 61 on the Intelligence Index, that provides it the lowest value per process of any at the moment measured mannequin at that intelligence degree.
Muse Spark 1.2 value solely $0.40 per Artificial Analysis process, whereas scoring 57.
Despite unchanged per-token pricing, the unbiased benchmark’s value of finishing a median process due to this fact elevated generation-over-generation. Artificial Analysis attributes the rise primarily to heavier input-token consumption on agentic evaluations.
That doesn’t straight contradict Meta’s declare of 25% decrease token use: Meta is describing comparisons in its personal coding workflows, whereas Artificial Analysis is measuring a broader suite of reasoning and agentic duties.
But it illustrates why “low-cost” turns into slippery as soon as fashions function as brokers. Token charges, reasoning effort, variety of turns, device calls and retries all contribute to the precise value of ending work.
Meta additionally retains its unusually low-cost Contributor tier — $0.10 per million enter tokens and $0.20 per million output tokens — in change for permission to make use of prompts and completions for coaching.
As VentureBeat famous with Muse Spark 1.2, that could be engaging for prototyping however creates a materially totally different data-governance calculation for enterprises working with proprietary code or delicate inside data.
Wang’s ‘Gemini who?’ lands on an unusually shut comparability
Meta chief AI officer Alexandr Wang was significantly much less certified in celebrating the discharge.
After Artificial Analysis posted its Muse Spark outcomes, Wang reposted them on X, adding: “i actually hate to say it, however… gemini who? 😱💨”
The shade was notably pointed as a result of Google released Gemini 3.8 Flash on the identical day, pitching it at nearly precisely the identical class of workload: long-horizon software program engineering, autonomous brokers and multi-step skilled reasoning. Google calls 3.8 its finest reasoning and coding Flash mannequin but and says it’s the firm’s third Flash launch in six weeks.
Independent numbers give Wang one thing to work with, although hardly a knockout.
Artificial Analysis offers Muse Spark 1.3 xhigh a 61 Intelligence Index rating at $0.55 per process, in contrast with 59 and $0.58 for Gemini 3.8 Flash at excessive reasoning. Meta due to this fact edges Google on each intelligence and process value at these specific settings.
Google wins decisively on throughput. Artificial Analysis measures Gemini 3.8 Flash excessive at about 305 output tokens per second, versus 235 for Muse Spark — roughly 30% sooner. Gemini additionally has the decrease uncooked API sticker worth for now: Google is charging an introductory $0.75 per million enter tokens and $3.75 per million output tokens, in contrast with Meta’s $1.25 and $4.25.
That promotional Google pricing expires December 31, after which it rises to $1.50 per million enter tokens and $7.50 per million output tokens.
The result’s a helpful snapshot of how tight frontier-model economics have change into. Meta at the moment wins this unbiased comparability by two Intelligence Index factors and three cents per benchmark process; Google affords considerably greater output throughput and cheaper uncooked tokens throughout its launch promotion.
Wang’s “gemini who?” is enjoyable govt trash speak. For an enterprise architect, the reply is nearer to: Gemini is the sooner choice; Muse is at the moment the marginally stronger high-effort agent by this unbiased measure.
Meta's evolving open weights stance
The extra consequential challenge for some builders might have little to do with at the moment’s benchmark race.
When Meta launched Muse Code and Muse Spark 1.2 in August, VentureBeat noted how dramatically the corporate had moved away from the open-weight technique that made Llama ubiquitous.
Muse Code and Spark 1.2 have been proprietary, API-served merchandise — a placing posture for the corporate that had spent years arguing that open AI was the trail ahead.
Five days later, Meta modified course once more.
On August 10, it launched the 30-billion-parameter Muse Glimmer under an Apache 2.0 license. Zuckerberg additionally stated: “In the approaching weeks, we’re additionally going to open the weights for Muse Spark 1.2.” Reuters individually reported Meta’s plan to launch the Spark 1.2 weights.
Now, Meta has as an alternative shipped Muse Spark 1.3 as one other proprietary mannequin.
That doesn’t but quantity to a damaged promise — “coming weeks” can fairly describe a interval longer than three weeks. But at the moment’s announcement makes the roadmap much less clear slightly than extra.
Meta’s new publish now not says Muse Spark 1.2. It says its roadmap consists of “the Muse Spark open weights launch”, with out figuring out a model, launch date, mannequin measurement or license. Zuckerberg likewise stated on X that “Muse Spark open weights releases” are coming quickly.
For groups that standardized on Llama as a result of downloadable weights meant self-hosting, customization and management over inference economics, that ambiguity might matter greater than whether or not Spark gained one other level on a composite benchmark.
Muse Spark 1.3 reveals that Meta can now iterate proprietary frontier fashions at extraordinary velocity. The delivery xhigh configuration is quick, competitively priced and far nearer to the highest of unbiased rankings than its predecessors. The max preview reveals Meta can push the household somewhat additional when allowed to spend extra reasoning compute.
The subsequent take a look at is totally different: whether or not Meta can convert that tempo right into a roadmap enterprises can really plan round — together with making its finest capabilities broadly deployable and delivering the open-weight Spark mannequin it has already stated is coming.
