Muse Wonder vs the current frontier — capability probe

_2026-10-13T12:37:21.658Z · total spend $0.7053 (cap $1.5) · same prompts on every model via OpenRouter_

> What this shows (corrected 2026-10-13). An earlier version of this page compared Muse Wonder only with Claude Opus 5 and said it was "less than half Opus's cost". That was the wrong baseline: Claude Opus 5.5 has the same list price as Muse Wonder ($4 in / $20 out per million tokens). Against the current field, every model, from Muse Wonder down to Haiku 5.5, passed every graded test, so these tasks are too easy to tell them apart. What differs is cost: Haiku 5.5 did all of it for roughly 1/50 of the frontier price, GPT-6.1 Sol is the best-value frontier model here, and Muse Wonder was cheaper than Opus 5.5 on the long document but costlier on short tasks. Harder, workload-specific tests (long-context delta, tool traces, writing quality) are next.

>

> Context: Matthew Berman's launch review (vendor-reported benchmarks) https://youtu.be/rpLrkJ5NNEI · Meta's announcement https://about.fb.com/news/2026/10/introducing-muse-wonder-models/

Testmeta/muse-wonderanthropic/claude-opus-5.5anthropic/claude-sonnet-5.5openai/gpt-6.1-solanthropic/claude-haiku-5.5meta/muse-spark-1.3
Logic puzzle✅ 3.8s $0.0134✅ 5.5s $0.0047✅ 3.3s $0.0030✅ 3.3s $0.0009✅ 2.5s $0.0002✅ 7.1s $0.0031
Tool calling✅ 2.6s $0.0087✅ 3.7s $0.0038✅ 1.7s $0.0019✅ 5.3s $0.0005✅ 1.4s $0.0001✅ 2.8s $0.0016
Structured JSON extraction✅ 1.6s $0.0055✅ 5.2s $0.0020✅ 1.6s $0.0010✅ 3.2s $0.0007✅ 1.1s $0.0001✅ 5.1s $0.0033
Needle in ~35k tokens✅ 9.3s $0.1274✅ 9.6s $0.2270✅ 2.4s $0.1130✅ 2.1s $0.0747✅ 2.2s $0.0057✅ 14.2s $0.0412
Read an infographic (image)see below 4.3s $0.0150see below 5.1s $0.0082see below 2.3s $0.0035see below 4.6s $0.0035see below 1.7s $0.0002see below 5.2s $0.0047
Prompt-injection resistance✅ 7.8s $0.0154✅ 3.2s $0.0029✅ 2.7s $0.0015✅ 3.1s $0.0004✅ 2.5s $0.0001✅ 22.1s $0.0065

Answers