Muse Wonder's Two Headline Features Are Measured Only by Meta

Meta’s first frontier model, as TechCrunch described it, is out, and the version you can buy per token scores exactly the same as GPT-6.1 Sol.¹ That comes from Artificial Analysis, the one independent evaluator I could find with results on launch day. Muse Wonder (muse-wonder-1.0) lands at 52 on the Artificial Analysis Intelligence Index, one point under Claude Fable 5.1 and GPT-6 Astra, at $4.52 per index task against Sol’s $0.72.²

That’s a respectable debut for a lab that had avoided state-of-the-art releases since its August 2025 restructuring.¹ The bigger, subscription-only Muse Wonder Full reaches 56, close behind Claude Opus 5.5 at 58.² Read the launch closely, though, and a pattern shows up. Most of Meta’s coding numbers carry an outside evaluator’s name. The two features Meta is selling the model on, Active Memory and Active Focus, are measured only by benchmarks Meta built, ran, and hasn’t published.³⁴ Until someone outside Meta tests them, they’re hypotheses, not results.

What We Know

Two models, one price tag

Meta released two models on October 12.³ Muse Wonder is the one in the Meta Model API: $4.00 per million input tokens, $0.20 cached, $20.00 output, with a 1M-token context window.⁵ Muse Wonder Full is available only through a Muse Code subscription and has no per-token price.⁵ That split matters for every chart in the launch, because most of the strongest numbers belong to Full.

Model	Price per 1M tokens (in / out)	AA Index	Cost per AA task
Muse Wonder	$4 / $20	52	$4.52
Claude Opus 5.5	$4 / $20	58	$5.98
Claude Fable 5.1	$10 / $50	53	$7.63
GPT-6 Astra	$10 / $50	53	$3.26
GPT-6.1 Sol	$2 / $10	52	$0.72

Prices come from each vendor’s pricing page; index scores and per-task costs come from Artificial Analysis, all at max effort.⁵⁶⁷²

TechCrunch described the price as roughly equivalent to GPT-6 Sol or Opus 5.5.¹ The Opus half holds. The Sol half doesn’t: OpenAI lists GPT-6.1 Sol at $2/$10, half of Wonder’s rate.⁷

List price also hides how much the model spends. Artificial Analysis recorded 130M output tokens to run its index, against an 85M median for comparable models, and a 19.27-second time to first token against a 5.52-second median.⁸ The result is a model that costs less per task than Opus 5.5 but more than GPT-6 Astra, which charges 2.5 times as much per token.

View full size

Open asset Close

Artificial Analysis data, October 12, 2026. Muse Wonder and GPT-6.1 Sol share a score; Wonder costs about six times as much per index task.

Meta’s own table, row by row

The model card compares five models on 14 benchmarks, all at max reasoning effort except where noted.⁴ A Muse model leads 7 rows, Opus 5.5 leads 6, and GPT-6 Astra leads one. The API model beats all three rivals on four of them: Harvey LAB-AA (legal agent work), MemoryBench, DeepSWE 1.1, and SWE-Marathon.⁴

The card is more transparent than most launch charts. Most rows name an outside source: Artificial Analysis for Harvey LAB-AA and Terminal-Bench 4.0, Vals AI for finance, Datacurve for DeepSWE, Abundant AI for SWE-Marathon.⁴ Artificial Analysis’s per-benchmark pages didn’t list Muse models yet when I checked, so those rows are attributed, not confirmed.

The card also discloses its own soft spots:

• Where no comparable third-party result exists, Meta uses “the highest of our internally validated run, self-reported score, or official leaderboard score.”⁴
• In-house competitor runs “remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models.”⁴
• SWE-Marathon, home of Full’s biggest coding lead (57.5 against Opus 5.5’s 46.3), is “not a controlled common-harness comparison,” and Full ran there at xhigh effort rather than max.⁴

On ProgramBench, the coding row where every model ran the same mini-swe-agent harness, Opus 5.5 scores 70.5 against Full’s 59.0 and Wonder’s 50.0. Opus also leads Terminal-Bench 4.0, 59.6 to Full’s 57.1 and Wonder’s 52.0.⁴ The strongest Muse coding row is DeepSWE 1.1, where Datacurve’s runs put Full at 79.0 and Wonder at 78.8 against Opus 5.5’s 74.2. The Astra and Fable numbers on that row come from their own vendors’ reports, so the comparison mixes sources, but it’s the gap most worth replicating.⁴

Mark Zuckerberg was candid about how the team optimizes. “As we’ve gone on, we steer the models more towards our internal benchmarks for the use cases that we’re focused on,” he said on the Cheeky Pint podcast.⁹ That’s a defensible product choice. It also means the rows Meta built itself deserve the most scrutiny.

MemoryBench isn’t the MemoryBench you’ll find

Active Memory is the launch’s headline feature, pitched as carrying relevant details across conversations.³ The evidence offered for it is one row: MemoryBench, where Wonder scores 85.6, Full 85.5, Opus 5.5 81.5, Astra 81.4, and Fable 5.1 79.0.⁴

The card defines it in a single paragraph. “MemoryBench is an internal benchmark that tests how accurately a model can recall critical information across long, complex multi-turn conversations of 64K to 100K tokens,” and “an LLM judge scores each answer against weighted, question-specific rubric criteria.”⁴ There’s no dataset, no named judge model, no error bars, and no paper.

Two more gaps. Recall inside one long conversation is a different skill from remembering across separate sessions, which is what “across conversations” promises. And the card reports no public long-term memory benchmark, such as LongMemEval, whose authors found commercial chat assistants drop about 30% in accuracy when memorizing information across sustained interactions.¹¹ Judging adds its own uncertainty: a preprint by Panickssery et al. found LLM evaluators recognize and favor their own generations.¹² Meta doesn’t say which model judges MemoryBench, so a four-point lead with no intervals tells us little.

Active Focus: model or harness?

Active Focus is the second pitch: keep working toward a goal when a task gets hard, recover from errors, and check in before acting on anything important.³ Its benchmark, FocusBench, is “an internal evaluation of personal agentic tasks,” and it reaches the table only inside the Personal Superintelligence Index, another internal composite.⁴ On that index Full leads Opus 5.5, 59.1 to 55.2. On Artificial Analysis’s independent index the order flips, 58 to 56.⁴²

The product layer muddies this further. In Muse Code, a built-in judge reviews eligible approval requests, and “risky actions cause the harness to stop for your decision.”¹³ Muse Code’s memory is Markdown files the tool keeps for the agent and reads into its context.¹⁴ Both are sensible features. But if Wonder checks in before a risky command, or remembers your project’s conventions inside Muse Code, the harness may deserve the credit. Muse Code’s default model is also muse-spark-1.2, so a casual trial may not be testing Wonder at all.¹⁴

Consensus: Muse Wonder Full sits near the frontier without leading it. The API model is a mid-pack frontier model priced like Opus 5.5, and Artificial Analysis lists it as proprietary.²⁸

Active disagreements: Meta’s internal composite ranks Full above Opus 5.5 and the independent index doesn’t. Whether the DeepSWE and SWE-Marathon leads reflect a real long-horizon coding edge is unsettled, and a benchmark-design audit (a preprint) found that flaws in task setup or reward design can misestimate agent performance by up to 100% in relative terms.¹⁵

Open Questions

• Which model judges MemoryBench, and will Meta publish the data, rubric, and confidence intervals?
• Does Wonder’s recall hold past the card’s 64K to 100K range, out to the 1M tokens the API accepts?
• How much of Active Focus inside Muse Code comes from the approval judge and sandbox rather than the model?
• Will Datacurve, Abundant AI, and Artificial Analysis publish the per-benchmark numbers the card attributes to them?
• In a real agent loop, do 130M-token habits and a 19-second first token erase the list-price parity with Opus 5.5?

Next Steps

I haven’t run Wonder yet. There’s no Meta Model API key in this lab today. When there is, the plan is four small arms on a pinned muse-wonder-1.0, the same harness for every model, and at least three repeats each:

1. Persistence: repos with planted blockers, such as a broken dependency or a misleading test, to see whether the model recovers or quits.
2. Check-ins: risky actions with every harness gate switched off, comparing the raw API against Muse Code with its judge on.
3. Recall: public long-context tasks with deterministic scoring at 64K, 200K, and 900K tokens, with facts updated mid-conversation.
4. Cross-session memory: Muse Code sessions with the memory files diffed, so file-based memory and model behavior stay separate.

Two API details will trip up existing eval scaffolds. tool_choice accepts only "auto", and "required", "none", or a named tool returns HTTP 400.¹⁶ Reasoning effort "none" also returns 400 on Wonder.¹⁷ Rate limits are 3,000 requests and 4M tokens per minute, shared across a whole team rather than per key.¹⁸

Angles I might be missing: if you’ve run Wonder against a public memory benchmark or a fixed coding harness, I’d like to compare notes, especially results that contradict this one.

Sources

Footnotes

1. TechCrunch. “Meta releases its first frontier model, Muse Wonder.” October 12, 2026. techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder. ↩ ↩² ↩³

2. Artificial Analysis. “LLM Leaderboard,” Intelligence Index v4.3.2. Accessed October 12, 2026. artificialanalysis.ai/leaderboards/models. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Meta. “Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents.” October 12, 2026. about.fb.com/news/2026/10/introducing-muse-wonder-models. ↩ ↩² ↩³ ↩⁴

4. Meta. “Muse Wonder Model Card.” October 12, 2026. research.meta.ai/static/muse-wonder-model-card. Benchmark table, methodology, and per-benchmark notes. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³

5. Meta. “Muse Wonder” model page. dev.meta.ai/models/muse-wonder. ↩ ↩² ↩³

6. Anthropic. “Pricing.” claude.com/pricing. ↩

7. OpenAI. “API pricing.” developers.openai.com/api/docs/pricing. ↩ ↩²

8. Artificial Analysis. “Muse Wonder.” Accessed October 12, 2026. artificialanalysis.ai/models/muse-wonder. ↩ ↩²

9. Cheeky Pint. “Muse Wonder: Zuckerberg and Wang’s $100B AI.” Transcript. October 12, 2026. cheekypint.substack.com/p/muse-wonder-zuckerberg-and-wangs. ↩

10. Ai, Q., Tang, Y., Wang, C., Long, J., Su, W., and Liu, Y. “MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems.” ICML 2026, PMLR 306. proceedings.mlr.press/v306/ai26e.html. ↩

11. Wu, D., et al. “LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory.” ICLR 2025. arxiv.org/abs/2410.10813. ↩

12. Panickssery, A., Bowman, S. R., and Feng, S. “LLM Evaluators Recognize and Favor Their Own Generations.” Preprint. arxiv.org/abs/2404.13076. ↩

13. Meta. “Muse Code: Permissions.” dev.meta.ai/docs/muse-code/permissions. ↩

14. Meta. “Muse Code: Configuration.” dev.meta.ai/docs/muse-code/configuration. ↩ ↩²

15. Zhu, Y., et al. “Establishing Best Practices for Building Rigorous Agentic Benchmarks.” Preprint. arxiv.org/abs/2507.02825. ↩

16. Meta. “Tool calling.” dev.meta.ai/docs/tool-calling. ↩

17. Meta. “Reasoning.” dev.meta.ai/docs/reasoning. ↩

18. Meta. “Pricing and rate limits.” dev.meta.ai/docs/pricing-rate-limits. ↩