Muse Spark vs Grok 4.7：IndexとArenaが食い違う点

On paper this is the least ambiguous matchup in the current field. M​use Spark 1.3, the current checkpoint of M​eta Superintelligence Labs' M​use Spark line, scores 48 on Artificial Analysis' Intelligence Index at its max reasoning setting and costs $1.60 to complete an index task. Grok 4.7, released September 21, 2026, scores 46 at its xhigh setting and costs $3.74 for the same task. M​use Spark wins the index, wins price per token on both sides of the invoice, wins context window, wins output speed, and emits fewer tokens doing it. Then you open a second independent board and the picture changes: Grok 4.7 sits at rank 89 of LMArena's text leaderboard with 1444 Elo from nearly twelve thousand votes, and M​use Spark 1.3 is not on that leaderboard at all. That divergence — not the two-point index gap — is the thing worth understanding before you pick, because whichever board your users resemble is the board that will decide whether your migration looks like an upgrade.

同じ賭けをする2つのモデル

Grok 4.7 and M​use Spark 1.3 are near-identical in design philosophy. Both are positioned as flagship reasoning models that run on a long-horizon agentic pitch. Both ship their headline benchmark numbers at an extreme effort setting rather than the default. Both are, by the independent measure, unusually talkative. And both changed nothing about pricing when they released — Grok 4.7 was served at the same price and speed as Grok 4.6, and M​use Spark has held $1.25 in and $4.25 out across three consecutive checkpoints.

Grok 4.7's own documentation describes it as x​AI's flagship for "code and everything else: agentic tool calling, minimal hallucinations, configurable reasoning", which is a claim about reliability rather than raw capability. M​eta's line for M​use Spark 1.3 is "trained for agentic workflows and optimized for competitive coding performance". Two vendors, describing the same job, with numbers that mostly favour one of them.

One naming note before the numbers, because it bites people: M​use Spark is a family. Muse Spark 1.1 (July 2026) and Muse Spark 1.2 (August 2026) share the same modality row and context window as 1.3 and differ only in capability. The unversioned "M​use Spark" refers to the original April 2026 model. Comparing Grok 4.7 to anything other than 1.3 is comparing it to a re-pointed older checkpoint.

独立した数値を並べて

以下はすべて、Artificial Analysis 自身がホストして取得した測定であり、各モデルの高コンピュート構成で行われたものです。どちらの行もベンダーによる主張ではありません。

• Intelligence Index — M​use Spark 1.3 (max): 48, rank 24 of 217. Grok 4.7 (xhigh): 46, rank 30 of 217.

• Cost per index task — M​use Spark 1.3: $1.60. Grok 4.7: $3.74. M​use Spark is 2.3x cheaper per completed task.

• Output tokens during the index — M​use Spark 1.3: 170M. Grok 4.7: 240M. Both are flagged as very verbose; G​rok is 41% more so.

• Output speed — M​use Spark 1.3: 121.9 tokens per second, rank 43 of 186. Grok 4.7: 90.8 tokens per second, rank 87 of 186.

• Time to first token — Grok 4.7 averages 37.26 seconds, which the listing places at the high end against a tier median of 3.70s. No comparable figure is published for M​use Spark 1.3, which is a gap rather than a win; we are not going to treat a missing number as a good one.

• List pricing — Grok 4.7: $2.00 in / $6.00 out with cached input at $0.50 for prompts under 200,000 tokens. M​use Spark 1.3: $1.25 in / $4.25 out with cached input at $0.15, no threshold.

• Context — Grok 4.7: 500,000 tokens. M​use Spark 1.3: 1,048,576.

• Modalities — both accept text and image and return text; M​use Spark 1.3 adds video input, which Grok 4.7's documentation does not list.

誰も言わない価格の閾値

Grok 4.7の価格は段階制で、しきい値はプロンプトトークン200,000に設定されている。それ未満では入力$2.00、出力$6.00を支払い、それ以上ではリクエスト全体が入力$4.00、出力$12.00で請求され、キャッシュ済み入力は$0.50から$1.00に上がる。このルールは限界使用量課金ではない——境界を越えるとリクエスト内のすべてが再価格付けされる。500,000トークンのコンテキストウィンドウに対して、これはウィンドウの深さの40%に位置する落とし戸であり、ドキュメントをGrokプロンプトに詰め込むあらゆる検索パイプラインは、出荷された翌月にそれを見つけることになる。

M​use Spark 1.3 publishes no comparable tier structure: one rate across the whole 1M-token window. That is worth more than the two-point index gap to anyone whose prompts are large by design. It also compounds with the cache difference — G​rok's cached input is $0.50 per million against M​use Spark's $0.15 — so in a heavily cached agent loop running near the threshold, the effective cost gap between these two is wider than the headline rates suggest, not narrower.

Grok 4.7 が真に正当な主張を持つ場面

上記の偏った表は議論のすべてではなく、G​rok側には率直に述べておくべき3つの点がある。

First, latency tiers. x​AI ships a Grok 4.7 Fast variant at double the speed and double the price, distributed through Cursor and G​rok Build rather than the public API. If your constraint is interactive responsiveness during peak hours, having a vendor-sanctioned fast tier for the same model is a real operational option. M​eta publishes no equivalent for M​use Spark 1.3.

Second, documentation completeness. Grok 4.7 publishes a knowledge cutoff of May 2026 and a full rate-limit and regional-availability picture; M​use Spark 1.3's published record does not include a knowledge cutoff or a max-output figure we could verify. When two models are close, the one whose limits are written down is the one you can plan against, and a vendor being silent about a ceiling is not the same as the ceiling not existing.

Third, and most importantly for the framing of this piece: the human-preference signal points the other way. Grok 4.7 holds rank 89 on LMArena's text leaderboard with 11,591 votes behind a 1444 Elo — a modest placement, but a measured one. M​use Spark 1.3 is not on that board. An index score and an arena score are answering different questions: the index asks whether the model can complete a set of hard tasks, the arena asks whether people prefer its answers. Grok 4.7's index rank of 30 against an arena rank of 89 says something specific — that its capability does not translate into preference at the same rate as its peers'. M​use Spark 1.3's absence says something less specific but not neutral: the evidence for how people react to its output is simply missing.

ベンダーの主張は、別途記載

以下のどちらの数値の組も私たちによって再現されたものではなく、またこの2つのリストは、互いに差し引くのに十分な共通の評価を共有していません。

x​AIの発表によると、Grok 4.7はCursorBench 4.0で46.3%、Terminal-Bench 4.0で37.6%、EEBenchで64.0%、DeepSWE v1.1で71.0%（high effort時）、Harvey Legal Agentで19.6%、HealthBench Professionalで56.7%、GDPval Eloで1695（xhigh時）を記録しています。

M​eta's scorecard for M​use Spark 1.3 reports DeepSWE v1.1 at 75.4, OSWorld 2.0 at 66.9, Terminal-Bench 2.1 at 76.4 with the Terminus 2 agent, MCP-Atlas at 89.2, GDPval-AA v2 at 1,754, SciCode at 58.5, HLE-Diamond without tools at 26.1 and JobBench at 61.8.

DeepSWEで75.4対71.0を見て、それを4ポイント差のMetaの勝利と読んでしまいたくなる。だが、そうしてはいけない。xAIはその数値を高エフォートで公表しており、Metaの最高スコアは最大設定によるもので、両社は異なるハーネスを使っているからだ。この二つのリストが本当に明らかにするのは、不一致の大きさではなく、その形である。

実際には何が意思決定を変えるのか

• プロンプトサイズ — プロンプトが200,000トークンを超える場合、Grok 4.7のしきい値再価格設定はこの比較における最大の単一明細項目であり、トークンあたりの表を誤解させる原因となるのはまさにこれです。

• Retrieval and long context — M​use Spark 1.3's 1M window against Grok 4.7's 500K is a real capability difference for whole-repository or whole-document work, not a spec-sheet flourish.

• Output volume — both models are verbose; Grok 4.7 more so at 240M tokens on the index against 170M. If your bill is dominated by output tokens, you are choosing between two expensive options, and the cheaper one is M​use Spark.

• Reasoning controls — Grok 4.7 exposes low, medium, high and xhigh, defaulting to high. M​use Spark 1.3 runs from minimal to max. Neither can be turned off entirely.

• Streaming and structured output — Grok 4.7 supports function calling and structured outputs, and its Responses API always returns encrypted reasoning content; batch processing is not supported. If you depend on batch pricing, that is a disqualifier on G​rok's side and worth checking against M​use Spark's own API terms before you assume the same.

• Preference risk — if your product's quality bar is set by user judgement rather than by benchmark completion, Grok 4.7 has a measured — if unflattering — arena position and M​use Spark 1.3 has none. Neither number is a reason to migrate; the absence of one is a reason to run your own comparison.

両方を1つのエンドポイントにまとめる

Grok 4.7 が OrcaRouter に登場、併せて当社が提供する2つの Muse Spark チェックポイント、Muse Spark 1.2 と Muse Spark 1.1も、各プロバイダーの定価を、0%のマークアップでそのままパススルーして — 200以上のモデルにまたがる単一のAPI、それらの間の自動フェイルオーバー、ルールごとにトラフィックを分割するルーティングDSL。この組み合わせは、この記事があなたに代わって答えることのできない問いに答えるための、最も安価な方法です。Grok側の200Kしきい値は、ルーティングルールの特に良い候補になります。Grokをしきい値で上限設定し、サイズ超過のリクエストは定額制のモデルに送り、どちらかが劣化した日にはフェイルオーバーに任せましょう。

Two disclosures. M​use Spark 1.3 itself is not in our catalogue, so the M​use Spark side of any test you run here starts at 1.2 — the checkpoint Artificial Analysis measures one point below M​use Spark Open, and one that now carries a deprecation notice pointing at 1.3. And the only numbers in this piece we would defend as measurements are the Artificial Analysis rows; the vendor figures on both sides are claims, and the LMArena placement is preliminary by its own marking.

今日はどっち？

If you are choosing on published, independent evidence and your workload is long-context, video-in, cache-heavy or cost-sensitive, M​use Spark 1.3 is the stronger model on nearly every line and the cheaper one per completed task by a factor of 2.3. If your workload lives under 200,000 tokens, depends on structured output and function calling, benefits from a vendor-run fast tier, or is judged by users rather than by benchmark completion, Grok 4.7 has a coherent case that the index comparison does not capture — and it is the only one of the two with a human-preference number at all. The honest summary is that this matchup is decided by prompt size and by who is grading the output, and by almost nothing else on either scorecard.

スコアボードが測定するものと、導入が実際に満たすものとの間にあるこの隔たりこそが、この記録が重要である理由です。このクラスのモデルをめぐる、文書化されたAIの障害やエージェントのインシデントは、the OrcaRouter AI Incident Archiveに集められており、各エントリには独自の情報源、重大度、信頼度グレード、異議ありフラグが付いています。