Top-ten AI models fell to 14% of OpenRouter tokens as buyers chose cheaper ones
The ten highest-quality AI models handled 14% of OpenRouter tokens in Q3 2026, down from 55% in early 2025, The Information reported. For Meta, which has just released its Muse Wonder models, cost per task now sits beside benchmark rank as the test.
Context for builders, not their beat.
The ten highest-quality AI models handled 14% of OpenRouter tokens in Q3 2026, against 55% in early 2025, per The Information. Artificial Analysis puts Muse Wonder at $4.52 per benchmark task and Spark 1.3 at $1.60.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Top-ten share, Q3 2026 | 14% of tokens | 1 | |
| Top-ten share, early 2025 | 55% of tokens | 2 | |
| Muse Wonder, cost per task | $4.52 per task | 5 | |
| Spark 1.3, cost per task | $1.6 per task | 5 |
What happened
- The Information reported that cheaper alternatives gained ground on OpenRouter even as the best models grew smarter.
- Meta released Muse Wonder and Muse Wonder Full on Monday; Artificial Analysis scored Full at 56, two points behind Claude Opus 5.5 at 58.
- Artificial Analysis puts standard Muse Wonder at $4.52 per benchmark task, nearly triple the $1.60 for Spark 1.3.
Why it matters
- What it costs cost Moving from Spark 1.3 to standard Muse Wonder buys four index points, 48 to 52, at the higher per-task cost above.
- What it limits constraint With no API price for Full, no one can work out its cost per task, so the model with the best scores cannot be set against cheaper rivals.
- Who has to decide decision Wonder is priced below Anthropic's lineup and above GPT-6 Astra and Gemini 4 Argon, so a buyer picks it only if the quality gap justifies paying more than for those two.
The drop is 41 percentage points [12]. The 14% share is about a quarter of the 55% the same group held in early 2025 [13]. Both figures come from The Information, as summarized by Digg [1][2].
What the figure turns on is how much quality buyers will pay for. Meta's new pricing shows the trade on a small scale. Muse Wonder costs $4 per million input tokens and $20 per million output, with a 1M-token context window [4]. Artificial Analysis scored it 52, four points above Spark 1.3 at 48 [10][14]. Meta's own claim is larger: Mark Zuckerberg wrote that Muse Wonder "brings us to the frontier" [9].
In our view, frontier standing and token volume are separate measures. The tier Meta says it has joined carried 14% of OpenRouter tokens last quarter [1]. A model can score near the top and still sell mostly where it is cheap enough to be the default.
The "stays in-house" reading overstates what the sources show. Standard Muse Wonder is in the Meta Model API [4]. Muse Wonder Full, the model behind the headline scores, has no API pricing yet and is limited to Muse Code subscribers [7]. The sources do not say how The Information ranks quality, which ten models make up the 14%, or when Meta will price Full for API use.
Until it does, Full's 56 can be compared with rivals on quality alone [8].
What to watch
- Whether Meta publishes API pricing for Muse Wonder Full, which would allow a cost-per-task comparison against the cheaper models.
- Whether The Information's next quarterly reading shows the top-ten share falling further below 14%.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The ten highest-quality AI models handled 14% of OpenRouter tokens in Q3 2026.
- [2]
The ten highest-quality models handled 55% of OpenRouter tokens in early 2025.
- [3]
Cheaper alternatives gained ground on OpenRouter even as the best AI models grew smarter.
- [4]
Muse Wonder is live in the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output, with a 1M-token context window.
- [5]
Artificial Analysis pegs Muse Wonder at $4.52 per benchmark task, nearly triple Spark 1.3's $1.60.
- [6]
Muse Wonder's per-task cost is cheaper than Anthropic's lineup and pricier than GPT-6 Astra or Gemini 4 Argon.
- [7]
Muse Wonder Full has no API pricing yet and is limited to Muse Code subscribers.
- [8]
Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index at max reasoning effort, two points behind Claude Opus 5.5 at 58.
- [9]
Mark Zuckerberg wrote in an X thread that Muse Wonder "brings us to the frontier".
- [10]
Artificial Analysis scored standard Muse Wonder at 52 on its Intelligence Index and Muse Spark 1.3 at 48.
- [11]
Meta released Muse Wonder and Muse Wonder Full on Monday, its most capable models to date.
- [12]
The top-ten share fell 41 percentage points between early 2025 and Q3 2026.
- [13]
The Q3 2026 share is roughly a quarter (about 25%) of the early-2025 share.
- [14]
Standard Muse Wonder scores four Intelligence Index points above Spark 1.3.
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- aibreakfast.beehiiv.com4d agoMeta's Frontier Model: Muse Wonder