Read this before you look for a benchmark table
There isn't one, and nobody honest can write one today. Meta published no numeric scores for Muse Wonder — only the claims that it performs "comparably to" Fable 5.1 and GPT-6 Astra and "higher than" them on MemoryBench. Every comparison below is on something measurable.
What you can actually compare
| |
Meta Muse Wonder |
GPT-6.1 Sol |
| Input / 1M tokens | $4.00 | $2.00 |
| Output / 1M tokens | $20.00 | $10.00 |
| Cached input | Not published | $0.10 |
| Released | 12 October 2026 | 29 September 2026, DevDay |
| Parameter count | Not published | Not published |
| Context window | Not published | Published |
| Numeric benchmarks | None | Published with system card |
| Track record | 1 day | 2 weeks, default in Codex |
The cached input line is the one to notice
GPT-6.1 Sol charges $0.10 per million cached input tokens against $2.00 uncached — a twentyfold reduction on repeated context.
Meta has published no cached input price for Muse Wonder at all. For agent workloads that resend a long system prompt on every turn, that single line can dominate the bill. If you are running anything with persistent context, the headline $4 versus $2 understates the gap, possibly by a lot.
Worth checking before you budget. Meta may well offer caching and simply not have published the price. But you cannot model a cost you have not been quoted, and "we will tell you later" is not a pricing page. Ask before you commit a workload.
What Meta claims, and what that is worth
Alexandr Wang called Wonder and Wonder Full "true frontier models, competitive with the best models in the world." Meta's model card says they score comparably to Fable 5.1 and GPT-6 Astra, and higher than Fable, Opus and Astra on MemoryBench.
Those are rankings without magnitudes. "Comparable" could mean within a point or within ten. There is no way to tell, and no third party has yet tested the model.
OpenAI published a system card for GPT-6 Sol and Luna on 7 October that classifies the models as High capability in cybersecurity and biological domains — a document that makes specific, checkable, and in places unflattering claims. That is a different standard of disclosure, and today it is the main difference between these two options.
Where Wonder might genuinely win
Active Memory is the one capability Meta describes as a lead rather than parity. The model retains information across conversations and retrieves it when relevant, and MemoryBench is the single benchmark where Meta claims to beat rather than match.
If your workload is an assistant that needs to remember a user between sessions, that is a real differentiator worth testing — persistent context is the gap between a chatbot and something that feels like it knows you. Active Focus, which keeps an agent pursuing a goal through obstacles, targets the other common agent failure.
Neither claim is verified. But they are specific, which is more than most launch claims manage.
Decision framework
- Shipping to production this month → GPT-6.1 Sol. Half the price, published benchmarks, two weeks of real-world use, and it is already the default in Codex.
- Running agents that must remember users across sessions → test Muse Wonder. This is the one case with a claimed advantage attached to it.
- Cost-sensitive at volume → GPT-6.1 Sol, and the cached input price widens the gap further the more context you resend.
- Already paying for Muse Code → Wonder Full is included. Nothing to decide, just use it and form your own view.
- You do not need a frontier model at all → neither. Claude Haiku 5.5 is $0.10 input and $0.50 output. Wonder's output price is forty times that.
What would change this
Independent benchmarks, and a cached input price. Both are plausible within weeks.
If third-party testing confirms parity with Sol and Meta publishes competitive caching, the premium becomes a question of preference rather than evidence. If testing shows a gap, a model priced at double the incumbent with no published scores will be a short story.
Until then the sensible position is not scepticism about Meta — it is that a buyer should not pay twice the going rate on a vendor's own say-so, however credible the vendor.
FAQ
Is Muse Wonder better than GPT-6.1 Sol?
Unknown. Meta claims comparable performance to GPT-6 Astra and Fable 5.1 but has published no numeric scores, and no independent testing exists yet.
How much more expensive is Muse Wonder?
Exactly double on both input and output: $4 against $2, and $20 against $10 per million tokens. The gap is wider in practice because Sol publishes a $0.10 cached input price and Meta has published none.
What is MemoryBench?
A benchmark for continual learning and memory construction. It is the one test where Meta claims Wonder beats rather than matches its rivals, though no score is given.
Can I use Muse Wonder Full through an API?
No. Wonder Full is restricted to Muse Code subscribers. Only Muse Wonder is available through the Meta Model API.
Which should I pick today?
GPT-6.1 Sol for anything going to production, on price and evidence. Muse Wonder is worth a trial specifically if cross-session memory is central to what you are building.