--- title: Meta releases Muse Spark 1.2, its third model in four months description: Meta has released Muse Spark 1.2, scoring 54 on the Artificial Analysis Intelligence Index — up from 51 in version 1.1 and 43 at launch in April. The model... canonical: https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Meta releases Muse Spark 1.2, its third model in four months | daily.dev og:description: Meta has released Muse Spark 1.2, scoring 54 on the Artificial Analysis Intelligence Index — up from 51 in version 1.1 and 43 at launch in April. The model... og:url: https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv og:image: https://api.daily.dev/og/posts/W7fg7hcmv.png og:image:alt: Meta releases Muse Spark 1.2, its third model in four months og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Meta releases Muse Spark 1.2, its third model in four months **[Collections](https://daily.dev/sources/collections)** · 5 min read · 0 upvotes · 0 comments ## Summary Meta has released Muse Spark 1.2, scoring 54 on the Artificial Analysis Intelligence Index — up from 51 in version 1.1 and 43 at launch in April. The model shows notable gains in agentic tasks, with GDPval-AA v2 Elo jumping 260 points to 1631, placing it fifth among all benchmarked models. Scientific reasoning benchmarks were mostly flat or slightly down. A key behavioral shift: the model now abstains more when uncertain, cutting hallucination rate from 38% to 28% but also dropping attempt rate from 82% to 67%. Pricing remains unchanged at $1.25/$4.25 per million tokens, though cost per task rose due to higher token usage. Meta is also running a training data flywheel through its internal coding agent MetaCode, where engineer corrections to AI mistakes feed back into training data for future models — over 7,000 weekly active users have submitted more than 800 fixes. ## Content ## What Meta released On August 5, 2026, Meta launched Muse Code, its first AI coding agent, alongside Muse Spark 1.2, the underlying model. Both come out of Meta Superintelligence Labs, led by Chief AI Officer Alexandr Wang. Muse Code is terminal-based, installs with a single command, and targets full software engineering workflows — planning, writing, and validating code — rather than just autocomplete. Zuckerberg also announced that Muse Spark 1.2 weights will be released publicly in the coming weeks, continuing Meta's open-weights strategy after Llama. ## Where the model actually stands Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index — up 3 points from 1.1 (51) and 11 points from 1.0 (43, released in April). That puts it roughly tied with GPT-5.5 and Grok 4.5, and behind the current frontier: Claude Opus 5 (61), Claude Fable 5 (60), GPT-5.6 Sol (59), and Kimi K3 (57). On coding benchmarks specifically, Muse Spark 1.1 scored 53% on DeepSWE, trailing GPT-5.6 Sol at 73% and Claude Opus 5 at 74%. Muse Spark 1.2 reaches 60.3% on WeirdML, roughly comparable to GPT-5 or GPT 5.4 Mini — better, but still not frontier. The clearest improvement in 1.2 is agentic work. GDPval-AA v2 Elo jumped 260 points (1371 to 1631), Terminal-Bench rose 2 points (78% to 80%), and Tau-Bench Banking gained 2 points. Scientific reasoning scores were largely flat, with small regressions on SciCode and Humanity's Last Exam. One pattern worth noting: Muse Spark 1.2 abstains more aggressively than its predecessor. The hallucination rate dropped from 38% to 28%, but the attempt rate also fell from 82% to 67%, and accuracy slipped from 41% to 38%. It's getting more cautious, not necessarily smarter. ## The cost picture Pricing is $1.25 per million input tokens and $4.25 per million output tokens — unchanged from Muse Spark 1.1. At $0.40 per Intelligence Index task, it's among the cheapest models at its capability level. Only Grok 4.5 high ($0.37) and GPT-5.6 Sol medium ($0.39) are cheaper in the same cluster. GPT-5.5 xhigh runs $1.18 per task; Kimi K3 max runs $0.86. On Text Arena's cost-quality comparison, Muse Spark 1.2 sits at 1498 rating versus Claude Fable 5's 1507 — a 9-point gap — while costing roughly 91% less per normalized price. For long-running agents, that gap compounds. Small per-token differences multiply across reasoning steps, tool calls, and retries. ## The contributor tier: where it gets complicated Muse Code has two pricing tiers. The standard tier is pay-as-you-go at the rates above, with zero data retention. The contributor tier is 12–21x cheaper, but Meta trains on your prompts and code completions. The contributor tier is the default. Engineering leaders interviewed about the launch are largely unwilling to accept that trade-off for proprietary work. Source code, developer prompts, and agent interactions are considered too sensitive. The consensus from practitioners is that the contributor tier makes sense for open source projects, personal work, and throwaway prototypes — not production codebases. What makes this more than a standard privacy concern is what coding data actually captures. Every session contains a problem statement, an attempt, and a verdict from real work. It self-grades. That's training signal that public repositories rarely contain — and it's exactly what Meta is collecting through MetaCode, its internal tool, where over 7,000 weekly active users have already submitted more than 800 corrections. Those fixes — original task, AI response, engineer fix, and review — are being used to post-train upcoming models. The 20x price gap will drive adoption by default for many teams, which makes the data policy a governance decision that most developers won't consciously make. ## How it performs against Claude Code in practice One head-to-head comparison ran Muse Code and Claude Code (running Fable 5) on three identical tasks using the dayjs JavaScript library: a bug hunt, a multi-file refactor, and a timezone reasoning test. On raw correctness, both agents tied — same bugs found, same gates passed. But Claude's work was more thorough: it reduced a file from 467 to 353 lines versus Muse's 467 to 465, produced a clean test suite, and left no dead code. Muse's output looked correct on the surface but required expert review to confirm it was actually finished. Cost: Muse on the contributor tier ran $0.06 total. Claude ran $9.41. Standard-tier Muse would have been roughly $2. The practical implication: token savings on Muse get partially paid back in review time. If you need to carefully audit the output anyway, the economics shift. ## The open weights announcement The decision to release Muse Spark 1.2 weights publicly is getting significant attention. Open weights let companies run the model on their own infrastructure, fine-tune it for specific domains, and avoid sending data to any third party — including Meta. For teams that want the cost efficiency without the data trade-off, self-hosting the open weights is the obvious path once they're available. It also continues to push down the cost of capable AI more broadly, which benefits the ecosystem regardless of which closed models are at the frontier. ## What Meta is betting on Meta's strategy has two parts. First, compete on price at a capability level that's close enough to frontier for most real work. Second, use production mistakes — not synthetic benchmarks — as training signal for the next generation of models. The risk is that frontier competitors are cutting prices too. Claude and GPT-5.6 Sol are both getting cheaper. If the price gap narrows, Meta's cost advantage shrinks, and the capability gap becomes more relevant. Muse Spark 1.2 is Meta's third model release in four months. The pace is fast. Whether the training data flywheel from MetaCode and the contributor tier produces a meaningful capability jump in the next release is the real question. ## Questions this post answers ### What does Muse Spark 1.2 score on the Artificial Analysis Intelligence Index and how does it compare to GPT-5.5 and Claude Opus 5? Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index, up from 51 in version 1.1 and 43 in the original 1.0 release. That puts it roughly tied with GPT-5.5 (55) and Grok 4.5 (54). The current top cluster sits higher: Claude Opus 5 at 61, Claude Fable 5 at 60, GPT-5.6 Sol at 59, and Kimi K3 at 57. _Developers tracking the AI coding tool landscape follow benchmark shifts like these on daily.dev._ ### What is the pricing for Muse Spark 1.2 per million tokens? Muse Spark 1.2 is priced at $1.25 per million input tokens and $4.25 per million output tokens, with cache hits at $0.15 per million. The context window is 1 million tokens. The cost per Intelligence Index task rose from $0.29 to $0.40 — not from a price increase, but because the model uses roughly 53% more input tokens and 36% more output tokens per task than version 1.1. _Teams budgeting for AI model usage track cost-per-task changes like these on daily.dev._ ### How is Meta using internal engineering workflows to generate training data for its AI models? Meta feeds real engineer corrections from its internal coding tool MetaCode back into model training. When engineers fix MetaCode's mistakes, the original task, the AI's response, the fix, and the review are all captured as training signal for upcoming models like Watermelon. Over 7,000 weekly active users have submitted more than 800 fixes. The bet is that production correction loops are a stronger signal than synthetic benchmarks or public code repositories. _Engineers building or evaluating AI coding agents find coverage of training data strategies like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 71 comments across x (as of 2026-08-11). **TL;DR:** Developers are broadly enthusiastic about Muse Spark 1.2 going open-weight, seeing it as a meaningful comeback for Meta after a disappointing Llama 4. The main concerns are around hardware accessibility and some quirks in the Muse Code agent's context awareness. **Sentiment:** 70% positive · 20% mixed · 10% skeptical **The case for** - Open-weighting Spark 1.2 is widely seen as a major win for indie developers, students, and self-hosters who can't afford frontier API costs. - Early hands-on reports suggest Muse Code handles real codebases better than expected. - The open-weight narrative challenges the idea that closed models will always outpace open ones. **The pushback** - Muse Code appears stateless and lacks environment-awareness that competing agents have, requiring more expert review. - Running frontier-scale open weights demands high-VRAM hardware (24GB+), making self-hosting inaccessible for many hobbyists. - The escalating cost of consumer GPU hardware is a growing barrier to actually benefiting from open weights. **By community** - x (positive): Replies are largely celebratory about the open-weight release and Meta's resurgence, with minor concerns about hardware costs and agent context-awareness. **Hottest debate:** Whether Muse Code's statelessness and lack of environment context is a fundamental limitation compared to rivals like Claude Code. **Open questions** - What is the actual architecture of Spark 1.2 beyond the open-weight announcement? - Will Meta release a smaller, more consumer-hardware-friendly variant that hobbyists can realistically run? **Highlights** > @theo Great Video first of all. Secondly, I totally agree on Spark 1.2 going open weight. That is the adult part of this release for me. Closed demos are fun. Weights you can run, poke, and post-train are how more people get real reps. Builders and students who cannot rent frontier > — [GargeyaS on x · 1 points](https://x.com/GargeyaS/status/2087094164781670807) > @theo muse not knowing the harness it is working in is fkin hilarious lmfao. web search results lol. > — [sougata\_x on x · 2 points, 1 comments](https://x.com/sougata_x/status/2087089538686664879) > @theo my fear is the escalating cost of consumer hardware making it less accessible to try these. I'd need a min 24gb VRAM card for my aging box at ~£350-£500 not inc any updates to the underlying hardware. Frustratingly expensive for a hobbyist. > — [cupcake\_trader on x · 1 points](https://x.com/cupcake_trader/status/2087096672186679771) > @theo The “open weights can’t keep up” narrative is getting harder to defend. > — [udohjeremiah\_ on x · 1 points](https://x.com/udohjeremiah_/status/2087098662874689781) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2087134617199661378) · 0 points · 0 comments - [x](https://x.com/theo/status/2087067918077182384) · 0 points · 71 comments ## Similar posts on daily.dev - [Meta enters the crowded AI coding battle with Muse Spark 1.1](https://daily.dev/posts/meta-enters-the-crowded-ai-coding-battle-with-muse-spark-1-1-wtfjrleyt) · TechCrunch · 1 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#agentic-ai](https://daily.dev/tags/agentic-ai) [View this post on daily.dev](https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Meta releases Muse Spark 1.2, its third model in four months","url":"https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv"},"datePublished":"2026-08-06T10:47:20.688Z","dateModified":"2026-08-11T18:42:32.702Z","description":"Meta has released Muse Spark 1.2, scoring 54 on the Artificial Analysis Intelligence Index — up from 51 in version 1.1 and 43 at launch in April. The model...","image":"https://pbs.twimg.com/media/HO_L-w7bMAAHHDi.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HO_L-w7bMAAHHDi.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,machine-learning,llm,agentic-ai","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Meta releases Muse Spark 1.2, its third model in four months"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/meta-releases-muse-spark-1-2-its-third-model-in-four-months-w7fg7hcmv#faq","mainEntity":[{"@type":"Question","name":"What does Muse Spark 1.2 score on the Artificial Analysis Intelligence Index and how does it compare to GPT-5.5 and Claude Opus 5?","acceptedAnswer":{"@type":"Answer","text":"Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index, up from 51 in version 1.1 and 43 in the original 1.0 release. That puts it roughly tied with GPT-5.5 (55) and Grok 4.5 (54). The current top cluster sits higher: Claude Opus 5 at 61, Claude Fable 5 at 60, GPT-5.6 Sol at 59, and Kimi K3 at 57. Developers tracking the AI coding tool landscape follow benchmark shifts like these on daily.dev."}},{"@type":"Question","name":"What is the pricing for Muse Spark 1.2 per million tokens?","acceptedAnswer":{"@type":"Answer","text":"Muse Spark 1.2 is priced at $1.25 per million input tokens and $4.25 per million output tokens, with cache hits at $0.15 per million. The context window is 1 million tokens. The cost per Intelligence Index task rose from $0.29 to $0.40 — not from a price increase, but because the model uses roughly 53% more input tokens and 36% more output tokens per task than version 1.1. Teams budgeting for AI model usage track cost-per-task changes like these on daily.dev."}},{"@type":"Question","name":"How is Meta using internal engineering workflows to generate training data for its AI models?","acceptedAnswer":{"@type":"Answer","text":"Meta feeds real engineer corrections from its internal coding tool MetaCode back into model training. When engineers fix MetaCode's mistakes, the original task, the AI's response, the fix, and the review are all captured as training signal for upcoming models like Watermelon. Over 7,000 weekly active users have submitted more than 800 fixes. The bet is that production correction loops are a stronger signal than synthetic benchmarks or public code repositories. Engineers building or evaluating AI coding agents find coverage of training data strategies like this on daily.dev."}}]} ```