Muse Spark 對上 Claude Opus 5.5：智力多十分，帳單高 3.7 倍

The comparison between Muse Spark and Claude Opus 5.5 is usually written as a scoreboard, and the scoreboard does have a winner: on Artificial Analysis' Intelligence Index the current Muse Spark — Muse Spark 1.3 at its max reasoning setting — scores 48, while Claude Opus 5.5 at max effort with default fallback scores 58 and sits first on the board. That is the number everyone quotes. The number almost nobody quotes sits on the same page: completing one Intelligence Index task costs $1.60 on Muse Spark 1.3 and $5.98 on O​pus 5.5. Ten index points for roughly 3.7x the money per task, with Muse Spark also running about 28% faster and emitting a third fewer tokens to get there. Neither of those figures is ours — both come from Artificial Analysis' own hosted listings — and which one matters depends entirely on what you are doing, so this piece is about separating the two questions rather than picking a side.

首先，模型名稱所隱藏的東西

Muse Spark is not one model, and the matchup only makes sense once you fix a version. The line runs Muse Spark 1.1 (July 16, 2026), Muse Spark 1.2 (August 5, 2026) and Muse Spark 1.3 (September 2, 2026), all from M​eta Superintelligence Labs, all sharing a 1,048,576-token context window and a text-out modality. M​eta labels 1.3 current. Muse Spark 1.3 also gained its "max" reasoning level on September 4, two days after launch, after M​eta finished additional safety testing — and that matters here because every AI-lab scorecard, M​eta's included, was produced at that setting, while the model most developers actually called for the first two days was capped at xhigh.

Claude Opus 5.5 arrived on September 22, 2026, the first release in A​nthropic's 5.5 family, and it is a simpler object: one GA model, adaptive thinking always on, effort selectable from low to max. A​nthropic's own framing was cost — it says O​pus 5.5 costs 40% less to run than Claude Opus 5, with input and output at $4 and $20 per million tokens and cache reads at $0.20, a 60% cut on cache reads versus its predecessor. Claude Fable 5.1 still sits above it for the hardest reasoning work.

所以，誠實的描述是：拿一個在同一個家族名稱下不斷被重新指向的模型，和一個以固定配置一次發布後就不再變動的模型來比較。

獨立的數字，以及它們的星號

Artificial Analysis 是唯一針對兩者都有公開榜單的第三方，而它並不會把它們放在同一個分類中比較——開放權重模型會與同尺寸級別的開放權重模型排名，專有模型則會與更廣泛的模型群體比較。這兩者都是專有模型，因此那個特定的偏差並不成立；真正成立的是，這兩份榜單都不是對另一方測試框架的受控重跑。

• Muse Spark 1.3 (max) — 智慧指數 48，在 217 個中排名第 24；輸出速度 121.9 tokens/秒；每百萬個 token 輸入 $1.25／輸出 $4.25；每項指數任務 $1.60；指數執行期間產生 1.7 億個輸出 token。

• Claude Opus 5.5（max，預設備援）—— 智慧指數 58，在 217 個模型中排名第 1；輸出速度每秒 95.2 個詞元；$4.00 輸入 / $20.00 輸出；每項指數任務 $5.98；2.6 億輸出詞元。

Read those two lists side by side and the shape of the trade is visible without any interpretation. O​pus 5.5 wins intelligence by 10 points and wins it while being more verbose, not less — 260M tokens against 170M — so it is not buying its score with economy. Muse Spark 1.3 is faster, cheaper per token, cheaper per completed task, and less talkative. On the price-per-task line is where the gap is widest: you can run roughly 3.7 Muse Spark 1.3 index tasks for the cost of one O​pus 5.5 index task.

再補一個獨立的資料點，因為它指向相反的方向，且值得提出來：LMArena 的文字排行榜把某一組 Claude Opus 5.5 配置列在總排名第 3 名，Elo 為 1508，而 Muse Spark 1.3 完全沒有出現在那個榜上。人類偏好投票與基準指標對幅度大小看法不一致，這很正常，但這確實意味著不該把那 10 分的差距解讀成「每一方面都差 10%」。

依廠商所述，差距最大之處

供應商評分卡彼此無法互相比較，而且兩家供應商也都沒有公布相同的基準測試，但其重疊之處仍具啟發性。以下所有內容皆為供應商所述，我們並未重現。

• Terminal-Bench 4.0 — O​pus 5.5: 66.4%, run at xhigh effort per A​nthropic's note. M​eta reports Muse Spark on Terminal-Bench 2.1 instead, where the family's published score is 76.4 with the Terminus 2 agent. Different benchmark versions, different agents — do not subtract them.

• GDPval-AA — O​pus 5.5: 1846 Elo on v2.1. Muse Spark 1.3: 1,754 on v2. The version moved between the two measurements, so the 92-point gap is directionally real and numerically soft.

• Agentic computer use — O​pus 5.5: 81.8% on OSWorld 2.1, marked "partial". Muse Spark 1.3: 66.9 on OSWorld 2.0. Again, different revisions of the same benchmark family.

• Terminal-Bench-Science 0.1 — O​pus 5.5: 58.7%. M​eta publishes no equivalent for Muse Spark 1.3.

A​nthropic is unusually explicit about the limits of its own table, noting that at this level of capability "benchmark margins have become a less reliable guide to real-world differences". That sentence is worth taking at face value in both directions: it undercuts the 10-point AA gap as much as it undercuts M​eta's DeepSWE placement claims.

除了分數之外，你還放棄了什麼？

• 輸入模態 — Muse Spark 1.3 接受文字、圖像與影片。Claude Opus 5.5 接受文字與圖像。如果你的管線直接將螢幕錄影或影片餵進模型，這根本沒得比，指數差距也無關緊要。

• Maximum output — O​pus 5.5 returns up to 128K tokens synchronously, and 300K through the Batches API behind a beta header. M​eta does not publish a max-output figure for Muse Spark 1.3 that we could verify. If you generate long artifacts in a single call, an unpublished ceiling is a risk you should test rather than assume.

• Cache economics — O​pus 5.5's cache read is $0.20 per million, a 95% discount off input. Muse Spark 1.3's is $0.15, an 88% discount. For a heavily cached agent loop this narrows or inverts the per-task arithmetic versus the headline rates, and it is the single easiest thing to get wrong when comparing these two on list price alone.

• 背景 — 兩者都是 1M。沒什麼需要分開的。

• Release cadence — O​pus 5.5 is one model with one price and has not changed since September. Muse Spark has re-pointed three times in three months and then added a reasoning level on top. That is either healthy iteration or an integration hazard, depending on how much you like re-running evals.

同時執行兩者，無需第二個整合

Both sides of this comparison are on OrcaRouter — Claude Opus 5.5 and the two Muse Spark checkpoints we serve, Muse Spark 1.2 and Muse Spark 1.1 — through one API across 200+ models, at each provider's list price with 0% markup passed through. Which is the practical answer to the question this article poses: the way to find out whether your workload really needs the extra ten index points is to route a slice of production traffic to Muse Spark 1.2 at $1.25 in / $4.25 out and compare it against the same prompts on O​pus 5.5, on the same key, with automatic failover so a bad run degrades to the other model rather than to an error page. No second contract, no second SDK, no migration.

Two honest caveats about our own coverage. Muse Spark 1.3 is not on our catalogue today, so the routing comparison above starts from 1.2, which is the checkpoint Artificial Analysis measures one point below Muse Spark Open and which now carries a deprecation notice pointing at 1.3. And this comparison rests on A​nthropic's and M​eta's published figures plus Artificial Analysis' independent listings; we have not re-run either model's benchmarks, and the vendor rows above should be read as claims.

值得回頭再看一次的版本

If the question you brought here is "is the cheap one good enough", the honest answer today is that the cheap one is 48 to O​pus 5.5's 58, faster, less verbose and about 3.7x cheaper per completed task — and that A​nthropic's own caveat about benchmark margins is the strongest argument for testing rather than deciding. If your work involves video in, extremely long single outputs, or the last few points of hard reasoning, O​pus 5.5 is the model that shipped with those things documented. The thing to watch is not the next Muse Spark version number but whether Muse Spark 1.3 arrives on a router you already use, and at what price — the family's pricing has been stable at $1.25 / $4.25 across 1.1, 1.2 and 1.3, which is itself a signal that M​eta is competing on capability per dollar rather than on the dollar.

One development belongs outside this matchup rather than inside it, and it is the reason this family is being written about this week at all. On October 15, 2026 — thirteen days after Muse Spark 1.3 shipped and the day before this article — M​eta published the weights and licence for Muse Spark Open, the family's separate open-weight member, whose repository had been sitting gated and unannounced since September 15. The 816B-total / 42B-active architecture currently circulating is that release, not the closed 1.3 measured above. Artificial Analysis does list it, at an intelligence index of 39 against the 48 this article attributes to Muse Spark 1.3, and M​eta has not published the evaluation report behind its own scorecard for it. So the artefact is confirmed and the performance claim is not; do not read this comparison as covering it.

數字無法告訴你的是，這一切在實際運作中是否出了問題：針對此類模型有記錄的 AI 事件與代理故障，都收錄在the OrcaRouter AI Incident Archive，其中每筆條目都附有自己的來源、嚴重程度、可信度等級與爭議標記。