Meta put a number on the table this week, and it is not a small one. Muse Wonder Full, the company’s newest large language model, scored 56 on the Artificial Analysis Intelligence Index, matching Anthropic’s Claude Sonnet 5.5 exactly. The lighter Muse Wonder came in at 52, a single point behind OpenAI’s GPT-6 Astra. Meta posted the announcement on October 12, 2026, and updated the page again on October 13 as outlets including Mashable, Reuters, Quartz, Deccan Chronicle, and South Korea’s financial press picked up the story.
The headline framing writing itself across tech coverage this week is that Meta has finally caught up to the two labs that have defined the frontier conversation for the past two years. The benchmark data backs part of that claim and complicates the rest. Muse Wonder Full ties or beats rivals on several measures and trails badly on at least one widely used coding test. This is a story about where Meta actually stands, not where the press release says it stands.
What Meta Actually Announced
Meta shipped two models under the Wonder name: Muse Wonder and Muse Wonder Full. Both take text, images, audio, and video as input and return text, and both carry a context window of roughly one million tokens, enough to hold a full codebase or a lengthy research paper in a single pass. The split between the two models is about access tier, not just capability. Muse Wonder is available through the Meta Model API for any developer, while Muse Wonder Full is gated behind a Muse Code subscription.
That gating matters for anyone trying to put the model through its own benchmark suite. VentureBeat noted that only the weaker of the two models ships through the open API, meaning third-party developers testing Muse Wonder in production right now are not working with the same model that produced Meta’s best scores. Anyone building on top of the API should account for that gap before assuming parity with the headline numbers.
The announcement also leaned on two feature names Meta is using to differentiate the release from a pure benchmark story: Active Memory, which is meant to carry context across separate conversations instead of resetting state each session, and Active Focus, designed to keep the model working through errors or dead ends and ask for permission or clarification rather than stalling out. Neither feature shows up in a leaderboard number, but both speak directly to the agentic workloads Meta is chasing with this release.
How to Try Muse Wonder Right Now
Developers can reach Muse Wonder today through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input priced at $0.20 per million tokens. Meta has set rate limits at 3,000 requests per minute and 4 million tokens per minute, limits aimed squarely at production agent workloads rather than hobbyist testing. Muse Wonder Full, the higher-scoring model, only shows up inside Muse Code for paying subscribers. Shattered.io previously covered the pricing structure behind Muse Code’s subscription tiers, including a top tier that had been testing at $200 a month before this launch.
For teams already running Muse inside Meta’s consumer assistant, the company said Muse Wonder will begin powering that product directly, starting with subscribers before a wider rollout. That is a notable shift from Muse Spark, the smaller, lower-latency model Meta had been running as its default assistant engine. Shattered.io’s earlier look at Muse Spark’s math-solving demo showed a model built for speed over raw reasoning power; Wonder is explicitly the opposite bet.
The Intelligence Index Numbers, Side by Side
Artificial Analysis runs one of the more widely cited composite scores in the industry, blending ten separate evaluations into a single Intelligence Index number. According to the figures Meta and Artificial Analysis published this week, Muse Wonder Full landed at 56, tying Claude Sonnet 5.5 on that specific measure. The smaller Muse Wonder scored 52, a point behind GPT-6 Astra. Artificial Analysis described the jump in its own assessment: “Both models exhibit a material jump in overall Intelligence Index score…Muse Wonder Full (max) places in the top three on six of the ten Intelligence Index evaluations, and is particularly strong in agentic knowledge work and scientific reasoning.”
That “top three on six of ten” detail is worth sitting with. A composite score can tie two models while hiding very different strengths underneath it. Muse Wonder Full’s agentic and scientific-reasoning results are reportedly strong enough to pull its average up to Sonnet-level territory even where other sub-scores lag. That is a different shape of model than a system that is uniformly strong across every category, and it points toward where Meta wants Wonder to actually get used: long-running agent tasks rather than single-shot trivia.
| Model | Maker | Artificial Analysis Intelligence Index | Access |
|---|---|---|---|
| Muse Wonder Full | Meta | 56 | Muse Code subscribers |
| Claude Sonnet 5.5 | Anthropic | 56 (tied) | API, Claude apps |
| GPT-6 Astra | OpenAI | 53 | API, ChatGPT |
| Muse Wonder | Meta | 52 | Meta Model API |
SWE-Marathon and ProgramBench: A Split Decision on Code
Coding performance is where the “Meta catches the frontier” narrative gets messier, and more interesting. On SWE-Marathon, a benchmark built around long-duration software engineering tasks rather than isolated bug fixes, Muse Wonder Full scored 57.5% against Claude Opus 5.5’s 46.3%. That is a double-digit gap in Meta’s favor on a test specifically designed to measure whether a model can stay coherent across an extended engineering session instead of losing the thread after a few tool calls.
Flip to ProgramBench, which leans harder on raw code generation and correctness across discrete programming problems, and the lead reverses hard. Claude Opus 5.5 scored 70.5%, more than eleven points ahead of Muse Wonder Full’s 59%. Put the two benchmarks next to each other and you get a genuinely split result: Meta’s model appears to hold up better over long, multi-step engineering sessions, while Anthropic’s model still produces more reliably correct code on a per-task basis. Neither company gets to claim a clean coding win here, and that split is probably the single most useful data point in this entire release for anyone deciding which model to route coding workloads to.
Shattered.io has tracked this kind of benchmark fragmentation before. Our coverage of Kimi K3’s Elo gains over GPT-6.1 Sol showed the same pattern: a single leaderboard number rarely tells the whole story once you break a model’s performance down task by task.
MemoryBench: Where Wonder Pulls Ahead Cleanly
The one benchmark in this release where Muse Wonder Full posts a clean, unambiguous lead is MemoryBench, which tests how well a model retains and uses information across a long session or across multiple sessions. Muse Wonder Full scored approximately 85.5% against Claude Opus 5.5’s approximately 81.5%, a four-point gap that lines up directly with Meta’s Active Memory feature pitch. If a model is explicitly engineered to carry context forward instead of discarding it, that should show up on a benchmark built to measure exactly that, and it did.
This is also the result that best supports Meta’s framing of Wonder as an agent model first and a chatbot second. Long-running agents, the kind that manage a multi-day research task or keep a coding project moving across dozens of sessions, live or die on memory retention. A four-point MemoryBench edge is a smaller number than the SWE-Marathon gap, but it may matter more for the use case Meta is actually targeting with this launch.
| Benchmark | What It Measures | Muse Wonder Full | Claude Opus 5.5 | Lead |
|---|---|---|---|---|
| SWE-Marathon | Long-duration software engineering | 57.5% | 46.3% | Meta, +11.2 pts |
| ProgramBench | Discrete code generation and correctness | 59% | 70.5% | Anthropic, +11.5 pts |
| MemoryBench | Cross-session context retention | ~85.5% | ~81.5% | Meta, ~4 pts |
Pricing and Access Tiers Compared
At $4 per million input tokens and $20 per million output tokens, Muse Wonder’s API pricing sits in the same general range as the mid-tier offerings from OpenAI and Anthropic, though exact rival pricing shifts often enough that direct comparisons age quickly. The more important detail is the cached-input discount, $0.20 per million tokens, which rewards applications that repeatedly send similar context, a common pattern in agent frameworks that re-send tool definitions and system prompts on every call.
The bigger access question is Muse Wonder Full’s subscription-only status. Developers who want the model that actually ties Claude Sonnet 5.5 on the Intelligence Index have to go through Muse Code rather than the open API, a decision VentureBeat flagged as a meaningful limitation on how quickly third-party developers can validate Meta’s benchmark claims for themselves. Enterprises evaluating a switch will need to weigh subscription costs against committing engineering time to a model they cannot yet fully test through standard API tooling.
| Access Path | Model Available | Pricing | Notes |
|---|---|---|---|
| Meta Model API | Muse Wonder | $4/M input, $20/M output, $0.20/M cached | 3,000 RPM / 4M TPM limits |
| Muse Code subscription | Muse Wonder Full | Subscription-based, reported tiers up to $200/mo | Not available via open API |
| Meta’s Muse assistant | Muse Wonder (rolling out) | Bundled with assistant access | Subscribers first, wider rollout to follow |
From Llama to Wonder: Meta’s AI Model Strategy
Meta’s path to a frontier-tier model has not been a straight line. The company built its early AI reputation on Llama, an open-weight series that traded top-tier benchmark scores for wide developer adoption and free commercial licensing. That strategy won Meta a lot of goodwill among researchers and startups that could not afford frontier-lab API bills, but it also meant Meta rarely showed up at the very top of leaderboards next to GPT or Claude releases.
Muse marked a shift toward a more consumer-facing, assistant-first approach, and Muse Spark pushed that further with a lightweight, low-latency model built to run Meta’s assistant products cheaply at scale rather than chase benchmark supremacy. TechCrunch described Wonder as Meta’s first genuine frontier model, explicitly contrasting it with Spark’s lightweight design philosophy. That framing lines up with what the numbers show: Wonder is bigger, slower to gate behind a subscription, and built to compete on the same leaderboards as GPT-6 Astra and Claude Sonnet 5.5, rather than to simply run cheaply across a billion phones.
Meta CEO Mark Zuckerberg has framed this shift in public terms as well, telling reporters the company’s models have reached the AI “frontier” and pushing back on calls from some in the industry to slow development pace, according to Yahoo Finance. Whether Wonder’s benchmark mix fully supports “frontier” as a label is a separate question from whether Meta believes it internally, and the split coding results above suggest the honest answer sits somewhere in between.
Competitive Landscape: OpenAI and Anthropic’s Responses
Neither OpenAI nor Anthropic had issued a direct public response to the Wonder benchmarks at the time of this report. Both companies have spent 2026 shipping their own rapid-fire model updates, and the competitive pressure clearly runs in multiple directions rather than Meta simply catching up to a fixed target. OpenAI’s own release cadence this year included shelving one model in favor of a cheaper successor, a move Shattered.io covered in detail when OpenAI shelved GPT-6.1 Astra in favor of GPT-6 Sol at a fraction of the price, underscoring how quickly the baseline for comparison shifts in this market.
Anthropic has been similarly active, rolling out its Claude 5.5 family across two models in under a week, a pace Shattered.io detailed in its report on the Claude 5.5 launch sequence. Against that backdrop, Muse Wonder Full tying Sonnet 5.5 on one composite index is a real achievement for Meta, but it lands in a market where the goalposts move every few weeks regardless of who is ahead on any given Tuesday.
Meta’s own agent product has also drawn direct competitive attention recently. Shattered.io’s coverage of how OpenAI’s dots agent is targeting Meta’s Muse user base showed both companies converging on the same consumer-agent battleground that Wonder’s Active Memory and Active Focus features are built for. This is no longer strictly a model-quality race; it is also a race for which company’s agent stack actually keeps users coming back.
Market Impact and What Remains Unconfirmed
No verified stock-price reaction to the Wonder launch had surfaced in reporting as of publication, and no major analyst note specifically pegged to these benchmark numbers had been independently confirmed. That absence is itself worth noting: a genuinely frontier-redefining model announcement from a company the size of Meta would typically generate immediate market commentary, and the relatively muted initial reaction suggests investors and analysts are treating this as an incremental, if solid, step rather than a watershed moment.
It is also worth being precise about what this launch does not prove. A tied score on one composite index, alongside a benchmark where Meta trails by more than eleven points, is not the same as Meta having definitively “caught” or surpassed OpenAI and Anthropic across the board. The honest read is narrower and more useful: Meta has closed a meaningful gap on agentic and memory-heavy tasks while still trailing on pure code-generation accuracy. That is a real result. It is not a sweep.
Data Training and Retention Questions for Enterprises
VentureBeat’s review of the launch flagged data training defaults and retention policy as an area that needs careful review before enterprise teams commit workloads to Muse Wonder, a caution that applies broadly across the industry whenever a consumer-facing company extends a chat-first model into developer and enterprise-facing API products. Meta has not published a fully detailed breakdown of how enterprise API traffic differs from consumer assistant traffic in terms of training-data use, and that gap is likely to come up in procurement conversations well before the benchmark scores do.
This is a recurring pattern across the industry rather than something unique to Meta. Security and compliance teams evaluating any new model family, regardless of vendor, need an explicit answer on data retention before green-lighting production use, and the absence of that detail in Wonder’s launch materials is a gap worth tracking as more enterprise reviews are published over the coming weeks.
Five Predictions for What Comes Next
- Expect pressure on Meta to widen Muse Wonder Full’s access beyond Muse Code subscribers if enterprise demand materializes faster than the subscription funnel can absorb it.
- OpenAI and Anthropic are likely to publish their own benchmark rebuttals or refreshed scores within weeks, given how quickly both labs have been shipping updates through 2026.
- ProgramBench’s double-digit gap in Anthropic’s favor means pure code-generation workloads probably stay on Claude Opus 5.5 for now, even as agentic and memory-heavy workloads shift toward testing Wonder.
- Data retention and training-default disclosures will likely become a bigger differentiator for enterprise buyers than the next point or two of Intelligence Index score.
- Active Memory and Active Focus, not the composite benchmark tie, are the more likely long-term differentiators if Meta can show them holding up in real multi-day agent deployments rather than controlled benchmark runs.
What Developers Should Actually Test Before Switching
Teams evaluating a move to Muse Wonder should run their own workload-specific tests rather than relying on composite scores alone, for one simple reason: the API-accessible Muse Wonder is not the same model that produced Meta’s best headline numbers. Muse Wonder Full, the model behind the 56 Intelligence Index score and the MemoryBench lead, is locked behind Muse Code. Anyone benchmarking through the open API is working with the smaller model that scored 52, a point below GPT-6 Astra.
For coding-heavy workloads specifically, the SWE-Marathon and ProgramBench split means the right choice depends heavily on task shape. Long-running agentic engineering sessions favor Meta’s approach; discrete, single-shot code generation still favors Anthropic’s. Running both benchmarks against your own representative tasks, rather than trusting either company’s headline framing, remains the only reliable way to pick a model for a specific production use case.
Frequently Asked Questions
What is Muse Wonder?
Muse Wonder is a new AI model family from Meta, announced October 12, 2026, built to handle text, image, audio, and video input and designed for agentic, long-running tasks rather than single-turn chat.
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder is available to any developer through the Meta Model API and scored 52 on the Artificial Analysis Intelligence Index. Muse Wonder Full is the larger, higher-scoring model at 56, but it is only available to Muse Code subscribers, not through the open API.
Does Muse Wonder beat Claude or GPT-6?
It is mixed. Muse Wonder Full ties Claude Sonnet 5.5 at 56 on the Intelligence Index and beats Claude Opus 5.5 on SWE-Marathon and MemoryBench. It trails Claude Opus 5.5 by over 11 points on ProgramBench and scores a point below GPT-6 Astra at the smaller Muse Wonder tier. There is no single benchmark where Meta wins across the board.
How much does Muse Wonder cost through the API?
Meta prices Muse Wonder at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million tokens, subject to rate limits of 3,000 requests per minute and 4 million tokens per minute.
How do I try Muse Wonder?
Developers can access Muse Wonder through the Meta Model API today. Muse Wonder Full requires a Muse Code subscription. Meta also said Muse Wonder will begin powering its Muse consumer assistant, starting with subscribers before a broader rollout.
What is Active Memory?
Active Memory is Meta’s name for a feature meant to carry context across separate conversations instead of resetting state every session, aimed at supporting longer-running agent workflows.
What is Active Focus?
Active Focus is designed to let the model keep working through errors or roadblocks during a task and ask for clarification or permission when it hits a genuine decision point, rather than stalling out entirely.
Has the market reacted to the Muse Wonder launch?
No independently verified stock-price move or analyst note tied specifically to these benchmark numbers had surfaced as of this report. The muted initial reaction suggests the market is reading this as an incremental step rather than a dramatic shift in competitive position.
