BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Muse Wonder's 60% discount depends on which rival you price it against
Meta launched Muse Wonder at $4 per million input tokens and $20 per million output tokens, a dev.to post says. The same post puts it roughly level with Claude Opus 5.5 on input, so the saving depends on the model your traffic runs on today.
The Engineer · Build desk
Per a dev.to post, Muse Wonder lists at $4 per million input tokens and $20 per million output tokens. If Muse Wonder is 60% below a rival on both price lines, that rival lists at $10 per million input tokens and $50 per million output tokens.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Muse Wonder input price | $4 per M tokens | 1 | |
| Implied rival input price | $10 per M tokens | if Muse Wonder is 60% below it | 18 |
| Muse Wonder output price | $20 per M tokens | 1 | |
| Implied rival output price | $50 per M tokens | if Muse Wonder is 60% below it | 18 |
What happened
- The year-end contract is tied: Google 42%, Anthropic 42% and OpenAI 9%.
- Tom Tunguz found, according to the post, that 84% of tokens on OpenRouter go to models that are not state of the art.
- The post also reports token usage up roughly 50% since July while prices fell 41%.
Why it matters
- constraint If the lead rotates in weeks, the switching cost sits in prompts and tool formats written for one model. Keeping those behind one interface is what makes a price move cheap.
- decision Routing on price forces the team to build an evaluation set first. The cheaper model has to pass your own tasks before any traffic moves.
- contradiction The post calls the model layer a commodity, yet its agent-market figures put Anthropic at 93%. Price-only routing optimizes the layer where the post says advantage is thin.
The post says Muse Wonder is 60% cheaper than Fable 5.1 and GPT-6 Astra [16]. For that to hold on both price lines, the rival would list at $10 per million input tokens and $50 per million output tokens [18]. Those are the figures to check on each vendor's own sheet.
Output costs five times input on Muse Wonder [14], so a workload that writes more than it reads is priced mostly at the $20 line. A Hacker News commenter quoted a cached input rate of $0.20 [2], which is 5% of list input [15]. If your traffic repeats a long prefix, the cached rate dominates the bill, and a discount quoted on list input does not tell you much about it.
The leaderboard half of the argument has firmer support. The contract asks which company will have the best model at the end of each month [13] and carries $1.6 million in liquidity [5]. Google went from the low 30s to 71% in early October while Anthropic fell from the 80s to 26% [6]. The post does not name the ranking that settles the contract, so the swing is evidence about that ranking only.
The price evidence is thinner. According to the post, Tunguz found that a supermajority of OpenRouter usage goes to six models priced at 2.5% of what Claude Fable 5 costs, which reach roughly 77% of frontier performance [9]. That is a measurement of other people's workloads. It transfers to yours only if your tasks fail in the same places as whatever benchmark produced the 77%.
Marc Andreessen told a16z's LP meeting that "technological moats that used to last a decade now last a quarter." [11] Russell Kaplan, president of Cognition, argues in the post's account that the durable part is the agent harness, the evaluation loop and the recovery paths [12].
A router needs an evaluation set to decide when the cheaper model is good enough. Building that set is the adoption cost. In our view the first half of the case holds, since a leaderboard position is a weak basis for a contract that outlasts the month. The second half needs one more condition: your own evals must show the cheaper model holds on your tasks.
What to watch
- Published list prices for Fable 5.1, GPT-6 Astra and Claude Opus 5.5, checked against the implied $10 input and $50 output.
- Independent evals of Muse Wonder on agent tasks, since the agent market still prices one lab near 93%.
- Whether the year-end Polymarket contract leaves its 42-42 tie between Google and Anthropic.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption
- Insufficient
- Hype gap+40
- Incentives35
- Confidence35
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Meta launched Muse Wonder, which the post calls its first true frontier model, at $4 per million input tokens and $20 per million output tokens.
- [2]
A Hacker News commenter listed the pricing as Input $4.00, Cached input $0.20, Output $20.00 and noted it matches the frontier competition.
- [3]
On Polymarket's best-model market, Google leads Anthropic 71-26 for October.
- [4]
The year-end best-model contract on Polymarket is a tie: Google 42%, Anthropic 42%, OpenAI 9%.
- [5]
The Polymarket best-model market has $1.6 million in liquidity.
- [6]
Google surged from the low 30s to 71% in early October on the Polymarket contract, while Anthropic dropped from the 80s to 26%.
- [7]
On the Polymarket agent market, Anthropic is at 93%, OpenAI at 4% and Google at 3%.
- [8]
Tom Tunguz found that 84% of tokens on OpenRouter go to non-state-of-the-art models.
- [9]
Six models that deliver roughly 77% of frontier performance at 2.5% of Claude Fable 5's price carry the supermajority of usage.
- [10]
Token usage is up roughly 50% since July while prices fell 41%.
- [11]
Marc Andreessen told a16z's LP meeting that "technological moats that used to last a decade now last a quarter."
- [12]
Russell Kaplan, President of Cognition, argues that what matters is what you build around the model: the agent harness, the evaluation loop and the recovery paths.
- [13]
The Polymarket contract asks which company will have the best AI model at the end of each month.
- [14]
Muse Wonder's output price is five times its input price.
- [15]
The $0.20 cached input rate is 5% of the $4.00 list input price.
- [16]
The post says the Muse Wonder price is 60% cheaper than Fable 5.1 and GPT-6 Astra.
- [17]
Later in the post, Muse Wonder is described as roughly matched with Claude Opus 5.5 on input costs and 60% below GPT-6 Astra.
- [18]
If Muse Wonder is 60% below a rival on both price lines, that rival lists at $10 per million input tokens and $50 per million output tokens.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toThe 'Best Model' Race Is Over. Here's What Actually Matters.
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Agent harness and AI moatsFollow
- Prediction MarketsFollow
- Model pricing and inference economicsFollow