Models#Benchmark#AI safety#Agent memory
Meta ships its first frontier models, Muse Wonder and Muse Wonder Full
Meta's first frontier models landed Oct 12: Muse Wonder at $4/$20 per million tokens, Wonder Full for Muse Code subscribers; independent evals scored 52/56.
Meta’s Superintelligence Labs shipped its first frontier models on October 12: Muse Wonder and Muse Wonder Full. The official blog calls them its most capable models yet, delivering “state-of-the-art performance on personal and professional agentic tasks.” Muse Wonder is already served through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with a 1M-token context window, per Meta’s developer page; Muse Wonder Full skips the API entirely and ships only inside Muse Code subscriptions. Independent evaluator Artificial Analysis scored the two at 52 and 56 on its Intelligence Index (v4.3) — the previous Muse Spark 1.3 sits at 48.
Key points
- The two models take different paths: muse-wonder-1.0 is metered on the Meta Model API and OpenRouter and also comes with Muse Code; muse-wonder-full-1.0 is subscription-only, unlockable from the $50/month Power Usage plan (developer page, checked October 13).
- Artificial Analysis calls the release “a material jump” over previous Meta models: 52 for Wonder and 56 for Full on the AA Index, against 48 for Spark 1.3; Full posts 68 on the Coding Agents Index, tying Claude Code with Sonnet 5.5 and clearing Claude Opus 5.5’s 66.
- The headlined capability is Active Memory, which Meta defines as “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand.” Both models trained at the gigawatt-scale Prometheus data center in Ohio, which came online earlier this year per TechCrunch.
Background
The line comes out of a year of rebuilding. Zuckerberg rebooted Meta’s AI effort as Meta Superintelligence Labs, hired Scale AI founder Alexandr Wang to run it, and spent billions; WIRED calls it an expensive bet that “has sometimes been rocky,” with researchers leaving for OpenAI along the way. The revamped lab’s first model, Muse Spark, arrived in April and was refreshed to Spark 1.3 about six weeks ago, per Artificial Analysis. The commercial plumbing is even newer: the Meta enterprise AI platform packaged Muse API and Muse Code for businesses on September 28 with pricing left unstated — this release fills that gap. On the consumer side, the Muse assistant app passed 3.4 million downloads in under three weeks and topped the US App Store, as we covered in September. One more date worth noting: Zuckerberg said in an X post last month that the release had slipped to work on safety, per WIRED.
The facts
Availability and pricing, from Meta’s developer page (checked October 13):
| Model | Channels | Pricing (per million tokens) | Context |
|---|---|---|---|
| muse-wonder-1.0 | Meta Model API, OpenRouter, Muse Code | $4 input, $0.20 cached input, $20 output | 1M |
| muse-wonder-full-1.0 | Muse Code subscription only (Power Usage $50/mo, Maximum $200/mo) | not sold by the token | 1M |
The capability story has two pillars. Active Memory carries context about a user across conversations and pulls back what a task needs; Active Focus is sustained work on a long goal — trying other approaches, working through blockers, adapting to new information, and checking in before consequential actions (both per the official blog). The benchmarks Meta names are Harvey’s Legal Agent Benchmark (LAB), FocusBench, and SWE-Marathon for long-running software projects; scores appear only as charts and a model card, with no numbers in the post text. TechCrunch, relaying Meta’s own testing, says Wonder outscores Fable 5.1, Opus 5.5, and GPT-6 Astra on MemoryBench, a benchmark for continual learning and memory construction — that one is a vendor claim, and MemoryBench does not appear on the blog page. On safety, Meta shipped the release alongside its Superintelligence Scaling Framework, which guards against CBRN and cyber misuse while preserving access to frontier cyber capabilities for defensive use; WIRED adds that the models were trained to refuse requests like bioweapon recipes or hacking help.
Artificial Analysis’s third-party numbers (v4.3, checked October 13):
| Metric | Wonder (max) | Wonder Full (max) | Reference |
|---|---|---|---|
| AA Intelligence Index | 52 | 56 | Spark 1.3 is 48; WIRED reports Full above Fable 5.1, below Opus 5.5 |
| Coding Agents Index | 63 | 68 | Opus 5.5 is 66; Full ties Claude Code with Sonnet 5.5 |
| AA-Omniscience (knowledge) | 25 | 30 | Opus 5.5 is 46, GPT-6 Astra 43, Gemini 4 Argon 42 |
| Cost per task | $4.52 | not token-priced | Opus 5.5 $5.98, GPT-6 Astra $3.26 |
What others say
Artificial Analysis’s verdict is “a material jump in intelligence over previous Meta models,” counting improvements across all ten Intelligence Index evaluations for Wonder Full over Spark 1.3. The same review flags the weak spot: knowledge recall on AA-Omniscience, where Full’s 30 trails Opus 5.5’s 46. TechCrunch quotes Meta research scientist Kavya Kopparapu: “We put a tremendous amount of effort into making sure your model can remember what’s important to you.” On Artificial Analysis’s per-task math, Wonder costs $4.52 against Opus 5.5’s $5.98. The Verge cautions that Meta’s claim to rival OpenAI and Anthropic flagships rests on internal testing, and that how the models compare with Google’s Gemini 4 Argon is still unclear. WIRED supplies the safety timeline — the delay post last month, the scaling framework this month — and its own caveat: benchmarks are not always a reliable measure of real-world performance.
Our take
The plain fact of this release is that the frontier tier is a four-way race again: Meta, OpenAI, Anthropic, Google. Two signals matter more if you build on these APIs. Pricing: $4/$20 matches Claude Opus 5.5 dollar for dollar, and per Artificial Analysis’s per-task math Wonder runs about 24% cheaper than Opus 5.5 ($4.52 versus $5.98). Distribution: Meta keeps its strongest model locked to its own subscription and meters only Wonder, so using this generation fully means entering the Muse Code system — the same play as September’s enterprise platform. The honest catches: the knowledge-recall gap (30 versus 46) suggests retrieval-heavy work may not favor these models; the blog page itself publishes no scores and Artificial Analysis is so far the only independent number; and Wonder Full, sold by subscription alone, drops out of any cross-vendor price comparison.
How to try it
API route: get a key at dev.meta.ai; the developer page documents an OpenAI SDK-compatible client with the model ID muse-wonder-1.0, and OpenRouter carries it too. Cookbooks cover multi-agent orchestration, computer use, and more. Subscription route: Muse Code’s Power Usage plan ($50/month) unlocks muse-wonder-full-1.0. Meta says Wonder will “soon” power the Muse assistant, starting with subscribers; no date given.