Muse Spark 對決 GPT-6.1 Sol：每任務成本與每詞元成本的結論截然相反

Put M​use Spark and GPT-6.1 Sol on the same line and the per-token pricing makes the decision for you: $1.25 against $2.00 per million input tokens, $4.25 against $10.00 per million output. M​use Spark looks like the cheap one by a wide margin. Then you look at what Artificial Analysis charges to actually complete a single Intelligence Index task with each and the order flips — $0.72 for GPT-6.1 Sol at its max reasoning setting against $1.60 for M​use Spark 1.3 at max. Sol spends less than half as much per finished task while carrying a four-point higher index score, and the only reason is that it is quiet: Sol generated 67M tokens running the index, and M​use Spark 1.3 generated 170M. If you are still costing these two models with a per-token spreadsheet, this matchup is the clearest case in the current field where that spreadsheet is wrong.

這些各自實際上是什麼

GPT-6.1 Sol is O​penAI's positioning play in the G​PT-6 line: released September 29, 2026, priced explicitly below the flagship GPT-6 Astra — O​penAI's announcement frames it as matching Astra's intelligence on agentic coding, computer use and professional work at roughly one-fifth of Astra's standard input and output prices. It is generally available, not deprecated, and supported by O​penAI's published model list.

M​use Spark 1.3 is the current checkpoint of M​eta Superintelligence Labs' M​use Spark family, released September 2, 2026 and given the max reasoning setting on September 4. Muse Spark 1.1 and 1.2 precede it. For anyone arriving from the older comparison material, "M​use Spark" without a number now means 1.1, 1.2 and 1.3 collectively, and 1.3 is the one M​eta labels current.

所以，這是一場兩款機型之間的對決，它們都圍繞著同一個賣點——以刻意壓低的價格帶來近乎旗艦級的能力——這讓它們在如何實現這點上的差異，比評分本身更有意思。

獨立的數字，以及做出決定的那一個

下方兩列皆為 Artificial Analysis 自行託管的測量結果，是在各模型最高推理設定下取得。這些並非廠商提供的數據。

• Intelligence Index — GPT-6.1 Sol (Max): 52, rank 12 of 217. M​use Spark 1.3 (Max): 48, rank 24 of 217.

• Cost per index task — GPT-6.1 Sol: $0.72. M​use Spark 1.3: $1.60. Sol is 2.2x cheaper per completed task.

• Output tokens during the index — GPT-6.1 Sol: 67M, described as fairly concise against a median of 87M. M​use Spark 1.3: 170M, described as very verbose.

• Output speed — GPT-6.1 Sol: 53.5 tokens per second, ranked 154 of 186, described as notably slow. M​use Spark 1.3: 121.9 tokens per second, ranked 43 of 186, notably fast. M​use Spark is 2.3x faster out.

• List pricing as measured — Sol at $2.00 in / $10.00 out with cached input at $0.10; M​use Spark 1.3 at $1.25 in / $4.25 out with cached input at $0.15.

There is a genuine trade hiding inside those five bullets, and it is not the one the headline suggests. Sol is the cheaper model to run a complete task with, and it does that by thinking less and talking less — 67M tokens is roughly two-fifths of M​use Spark's output for four more index points. M​use Spark will get an answer to you faster in wall-clock terms once it starts, roughly twice as fast, but it will have generated far more text doing it. Which of those you want depends on whether your constraint is the invoice at the end of the month or the latency in front of a user.

As a sanity check on the direction, LMArena's text leaderboard places a GPT-6.1 Sol configuration at rank 22 with 1484 Elo; M​use Spark 1.3 does not appear on that board. The four-point index gap and the human-preference gap point the same way here, which is not always the case.

雙方廠商的主張

O​penAI published its GPT-6.1 Sol results mostly as deltas rather than absolute scores, which is worth knowing before anyone quotes them as a table. What the announcement asserts, all vendor-reported and unreproduced by us: Sol matches GPT-6 Astra on DeepSWE v1.1 at about a fifth of the cost and beats GPT-6 Sol's best score by 6.4 percentage points at lower reasoning effort; it scores above Claude Opus 5.5 at less than half the cost per task on GDP.pdf; it improves on O​pus 5.5 by 2.2 points on AutomationBench at medium effort; it beats GPT-6 Sol by seven points on OSWorld 2.0's offline set at max effort while coming within 2.1 points of Astra at roughly a seventh of the cost; and at low reasoning effort it cuts the share of responses containing a factual error from 11.4% to 7.7%. The only absolute number in O​penAI's scientific-workflow claim is Astra's own 68.1% on Terminal-Bench Science 0.1, with Sol costing $5.47 per task against $23.21 for Claude Opus 5.5 — a cost comparison, not a score comparison.

M​eta's published figures for M​use Spark 1.3 are 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. Note carefully what cannot be done with these two lists: O​penAI reports changes, M​eta reports levels, and the two vendors did not run the same evaluations on the same harnesses. The one place they nearly meet is DeepSWE v1.1, and even there one number is a delta and the other an absolute.

比分數更重要的規格差異

• Input modalities — M​use Spark 1.3 accepts text, image and video. GPT-6.1 Sol accepts text and image only; audio and video are documented as unsupported. For any workflow with video in, this comparison ends here.

• Context window — Sol publishes 1,050,000 tokens with a maximum input of 922,000 and a maximum output of 128,000. M​use Spark 1.3 is a flat 1,048,576 with no published max-output figure we could verify. These look equivalent and behave differently: Sol's window is explicitly carved into input and output budgets, while M​use Spark's is not documented that way.

• Long-context pricing — this is the operational trap on Sol's side. Standard rates apply to prompts up to 272,000 input tokens; above that threshold the request is billed at long-context rates of $4.00 in and $15.00 out, double the input price and 1.5x the output price, for the whole request. A Sol agent that drifts past 272K tokens does not pay a little more, it changes price tier. M​use Spark 1.3 has no comparable tiering in its published rates.

• Service tiers — Sol comes with batch and flex at 50%, fast mode at 2x, and an ultrafast tier at 6x standard pricing, added on October 8, 2026. M​use Spark 1.3 has a standard tier and a Contributor variant that trades a lower price — $0.10 in, $0.20 out — for permission to train on your prompts and completions. That last option has no O​penAI equivalent and no G​oogle equivalent, and for some teams it is disqualifying while for others it is a 10x cost reduction.

• Reasoning controls — Sol's effort parameter runs low, medium, high, xhigh and max, and explicitly does not support "none" or "minimal" — you cannot turn reasoning off. M​use Spark's ladder runs from minimal upward, and M​eta's tooling notes reasoning cannot be disabled either. Both vendors have decided this for you.

• Knowledge cutoff — Sol publishes April 30, 2026. M​eta does not publish a comparable figure for M​use Spark 1.3 that we could verify.

每項任務成本這個數字能告訴你什麼、不能告訴你什麼

The $0.72 against $1.60 figure is the most useful single number in this comparison and it is also the most easily misused. It is the mean cost of running one Intelligence Index task at maximum reasoning effort, which is a specific mixture of reasoning-heavy and long-output evaluations — not your workload. If your application is one short question and one short answer, your tokens-per-request are tiny and M​use Spark 1.3's $1.25 input rate wins outright. If your application is an agent that reasons for a long time and writes a lot of intermediate text, verbosity is the dominant cost driver and Sol's 67M tokens against M​use Spark's 170M is the difference between a sustainable bill and an unpleasant one. The honest way to use that pair of numbers is as a warning that per-token rates rank these two models one way and completed-task economics rank them the other, and only your own token distribution tells you which ranking you live under.

把兩者放到同一個按鍵上

GPT-6.1 Sol 已在 OrcaRouter 上線，而我們所供應的兩個 Muse Spark 檢查點——Muse Spark 1.2 與 Muse Spark 1.1——亦然，皆以各供應商的標價提供，0% 加價直接回饋。單一 API 橫跨 200 多個模型，意味著上述的冗長度問題可以用實證方式回答，而非僅止於理論：透過單一端點記錄數百筆真實請求的 token 數量，接著用路由 DSL 把你流量中屬於長推理的那一部分，導向在你的分布上實際勝出的那一方，並以自動容錯移轉涵蓋落敗的模型。由於 Sol 的長上下文級距會在 272K 輸入 token 時切換價格，這套容錯移轉也讓你有地方安置原本會跨越該門檻的請求。

The caveat we owe you: M​use Spark 1.3 is not on our catalogue, so the M​use Spark side of that test runs on 1.2, which Artificial Analysis measures one point below M​use Spark Open and which now carries a deprecation notice pointing at 1.3. And every vendor number in this piece — O​penAI's deltas, M​eta's scorecard — is a claim, not a measurement; the only figures here we would defend as independent are the Artificial Analysis rows.

這個決定，直白地說

On the evidence, GPT-6.1 Sol wins this matchup on the metrics that usually decide enterprise adoption: higher independent index score, roughly 2.2x cheaper per completed task, half the output volume, a documented maximum output, and published service tiers you can trade speed against price with. M​use Spark 1.3 wins on the three things a per-task average hides: it accepts video, it is about 2.3x faster once generating, and its pricing does not change tier when a prompt grows. If your workload is video-in or latency-sensitive, the cheaper per-task model is the wrong model. If it is long-horizon reasoning with a lot of written output, the per-token table is lying to you and Sol is the better buy by a factor of two.

像這樣的比較，也正是失效模式會顯現的地方：這一類模型背後有紀錄的 AI 事件與代理失敗，都收錄在 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，以及爭議旗標。