Frontier v4 analysis

Muse Wonder Full against Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra: combined score, time and cost

2026-10-15, updated 2026-11-04 · Morgan Linton · VulcanBench Frontier v4, Low to Max effort · combined score, time and cost

With its code now judged, Muse Wonder Full scores 87.0 on the combined score at Low, 88.3 at Medium, 88.2 at High, 87.4 at Extra-high and 88.4 at Max. It is level with the frontier at Low and second at Medium, but it does not gain from more effort: the other three climb at higher levels, it is fourth at Extra-high, and Max, added on 4 November, brings it back only to Medium’s level. It is also the slowest of the four at every level, at about 27 to 41 minutes per task, and the most expensive at every level but Max, at about $4 to $6.

87.0combined score at Low, out of 100; 88.3 at Medium†, 88.2 at High‡, 87.4 at Extra-high, 88.4 at Max
27 minper task at Low, the slowest of the four; 27 at Medium†, 36 at High, 30 at Extra-high, 41 at Max
~$3.93per task at Low, estimated, the most expensive of the four; $4.09 at Medium†, $4.72 at High, $4.86 at Extra-high, $5.91 at Max
59 vs 76reviewed Code quality at Low from its two judges, Opus 5.5 and GPT-6.1 Sol
Grouped bar charts of combined score, time per task and cost per task at Low, Medium, High, Extra-high and Max for Muse Wonder Full, Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra, with a table of the same numbers
Combined score, time per task and cost per task at all five levels. Medium† ran 21 tasks on Muse Code 1.4.6 and 2 on 1.4.9; High, Extra-high and Max ran on 1.4.9.

The numbers

Means per task over every run, one run per task and level. The combined score is 50% functional score (the share of hidden behavior tests passed), 8.5% quality, 8.5% security and 33% judged Code quality, with its standard error.

LowMedium†HighExtra-highMax
ModelCombinedFunctionalMin$/taskCombinedFunctionalMin$/taskCombinedFunctionalMin$/taskCombinedFunctionalMin$/taskCombinedFunctionalMin$/task
Muse Wonder Full87.0 ±1.295.827.3~3.9388.3 ±0.998.627.3~4.0988.2 ±1.0‡93.236.2~4.7287.4 ±1.097.530.3~4.8688.4 ±0.899.040.9~5.91
Claude Opus 5.586.4 ±1.394.213.71.7090.9 ±0.6100.018.62.8491.1 ±0.6‡95.721.43.2790.7 ±1.197.820.74.1190.2 ±1.097.237.58.75
GPT-6.1 Sol86.2 ±0.897.26.90.3186.4 ±0.898.710.60.4088.2 ±0.4100.010.20.3388.8 ±0.4100.015.40.4388.3 ±0.3100.014.50.46
GPT-6 Astra87.4 ±0.599.74.21.72+87.7 ±0.3100.03.81.48+88.2 ±0.4100.04.91.71+89.2 ±0.3100.08.12.30+89.3 ±0.4100.010.32.57+

Every cell covers 23 runs. † Muse Wonder Full's Medium ran 21 tasks on Muse Code 1.4.6 and the last 2 (cellarcore, paddockcore) on 1.4.9, after the Muse Code quota reset; its High, Extra-high and Max ran entirely on 1.4.9. ‡ Combined over 22 judged runs: Muse Wonder Full's High paddockcore run hit the 3-hour task limit and left no finished code to judge (84.4 with it counted as 0 over all 23), and Claude Opus 5.5's High depotcore run was stopped by a safety classifier. Muse Wonder Full's Extra-high depotcore is its clean rerun of 19 October, which replaces a run that timed out while another job shared the machine. Functional score, minutes and cost count every run. Data: frontier-comparison.csv.

Combined score

Time and cost

How the code was judged

Code quality is a third of the combined score, and two calibrated model judges score it. Each judge first passes an exam on held-out code samples under the same frozen protocol. The other three models were judged by Muse Spark 1.3 and Grok 4.6. Muse Spark is a Meta model, so it is not neutral for a Meta model, and Muse Wonder Full was judged by Claude Opus 5.5 and GPT-6.1 Sol instead. Both passed the exam with no failing gate.

Method and caveats

Runs. Muse Wonder Full (muse-wonder-full-1.0, Contributor mode) with its skill reminder off, at Low (23 tasks, 13 to 14 October 2026, Muse Code 1.4.6) and Medium (23 tasks: 21 on Muse Code 1.4.6, 14 to 15 October, and the last 2 on Muse Code 1.4.9, 16 October) High (23 tasks on Muse Code 1.4.9, 16 to 17 October), Extra-high (23 tasks on Muse Code 1.4.9, 17 October, with depotcore rerun on 19 October) and Max (23 tasks on Muse Code 1.4.9, 3 to 4 November). Meta documents Max as Standard-only for Muse Spark and says nothing about it for Muse Wonder Full in Contributor mode, so it was first left out; a probe on 3 November showed Contributor mode honors it (more than twice Extra-high's reasoning on the same puzzle), and it was run. The Muse Code weekly limit stopped Medium at 21 tasks; after Meta reset the quota and released a more token-efficient Muse Code, the last two and every later level ran on 1.4.9. One High attempt (snapcore) lost its connection to Meta's API partway through; it was unscored and the task was rerun. Muse Code 1.4.9 runs its end-of-run goal and verify reminder observers on Muse Spark 1.3, not on the main model as 1.4.6 did; those calls are recorded apart, and every other call was Muse Wonder Full. The other models are their published Frontier v4 runs: Claude Opus 5.5 in Claude Code 2.1.280, GPT-6.1 Sol in Codex 0.159.0 and GPT-6 Astra in Codex.

Judging. Muse Wonder Full's Code quality is protocol v3.26 (Claude Opus 5.5 at medium effort) with GPT-6.1 Sol's v3.25 scores (medium effort) on byte-identical inputs, the v3.26.1 top-up for the two 1.4.9 Medium runs, v3.26.2 for High, v3.26.3 for Extra-high, v3.26.4 for the Extra-high depotcore rerun and v3.26.5 for Max, judged by the same two judges with no new exam (each is another level of the same sweep). Opus 5.5 is v3.15, GPT-6.1 Sol v3.18 and GPT-6 Astra v3.4.

Cost. List prices were checked on 15 October 2026, and Muse Spark 1.3's on 16 October. The Spark reminder tokens of the 1.4.9 runs are added to Muse Wonder Full's cost at Spark's list price ($1.25 input, $0.15 cached, $4.25 output per million). Opus 5.5 is Claude Code's own total, which reprices exactly from its traces. GPT-6.1 Sol is its Codex token receipts at $2.00 / $0.10 / $10.00 per million. GPT-6 Astra ($10 / $1 / $50) is its report's central estimate; it and GPT-6.1 Sol charge more on requests over 272K input tokens, and request sizes were not recorded, so + marks Astra's upper bound ($3.28 at Low, $2.79 at Medium, $3.22 at High, $4.25 at Extra-high, $4.69 at Max). GPT-6.1 Sol's calls average about 35K input tokens, so its premium is likely small.

Not controlled. The harnesses differ, and minutes are wall-clock time in each model's own harness. Nine of Muse Wonder Full's 23 Medium tasks, 14 of its 23 High tasks and 4 of its 23 Extra-high tasks overlapped another local job for part of their run, which can lengthen minutes; Max ran alone. Extra-high depotcore first hit the 3-hour limit with 145 of its 180 minutes overlapped (it finished clean at High in 111 minutes), so it was rerun alone on 19 October (109 minutes, every hidden test passed), and the judged rerun replaces it; one GPT-6.1 Sol call in that round stopped on a Codex reconnect and was rerun once with identical inputs on 4 November. One run per task and level, except that depotcore rerun.

← All benchmarks