Meta shipped Muse Wonder on October 12, 2026, and within a day the Hacker News thread about it had cleared 650 points and 360 comments — the most-argued AI launch of the week. Almost none of that argument was about benchmarks. Meta Muse Wonder pricing starts at $4 per million input tokens and $20 per million output tokens, which is exactly twice OpenAI's standard rate for GPT-6.1 Sol and exactly twice the previously reported rate for Google's Gemini 4 Argon. Meta Superintelligence Labs, run by former Scale AI chief Alexandr Wang, reportedly spent around fourteen months and a data center in Ohio getting Meta back to the frontier, and it has two genuinely new ideas to show for it. The developers who paid to try them spent the first twenty-four hours comparing the output to Claude Sonnet and Haiku.
The Two Ideas Meta Is Actually Selling
Strip away the launch framing and Muse Wonder is pitched on two capabilities, both named, both aimed at the same complaint: that models forget you and give up on you. Meta's research post describes the model's strengths in domain expertise — law, finance, health — alongside coding, conversation continuity, and what it calls professional and personal agent tasks. The two named features are where the actual bet lives.
Active Memory: Context You Don't Have to Re-Type
Meta defines Active Memory as carrying important context about you across conversations and actively retrieving the details relevant to the task at hand. The second half of that sentence is the part that distinguishes it from a stored-profile feature: the claim is not just that the model keeps notes, but that it recognizes when something you mentioned earlier has become relevant again. Research scientist Kavya Kopparapu put the ambition in plain terms, saying the team put tremendous effort into making sure the model can remember what's important to you over days, weeks, months, and beyond. That is a meaningfully different product promise from a long context window, and Muse Wonder ships with a one-million-token context window anyway — so Meta is explicitly arguing that stuffing more text into a single request was never the answer.
Active Focus: Not Quitting
Active Focus is described as the model's ability to sustain effort toward difficult goals by trying alternative approaches, working through blockers, and adapting to new information. Meta adds two behavioral specifics that are easy to check in practice: the model asks for clarification when it needs it, and verifies actions before proceeding. Both are testable in an afternoon, which matters, because the first developers to test Active Focus found something closer to its failure mode than its promise — more on that below.
The Price Is Double, Line by Line
Here is every number the launch gave developers, set against what the sources disclose about the obvious alternatives. The gaps are real gaps: the published material simply does not put a token price on Anthropic's or OpenAI's flagship tiers, and does not score Sol or Argon on the same index.
| Model | Artificial Analysis Intelligence Index | Input $/M | Cached input $/M | Output $/M | Rentable per-token? |
|---|---|---|---|---|---|
| Muse Wonder | 52 | $4.00 | $0.20 | $20.00 | Yes — Meta Model API |
| Muse Wonder Full | 56 | — | — | — | No — Muse Code subscription only |
| Claude Opus 5.5 | 58 | not disclosed in sources | — | not disclosed in sources | — |
| Claude Sonnet 5.5 | 56 | not disclosed in sources | — | not disclosed in sources | — |
| GPT-6 Astra | 53 | not disclosed in sources | — | not disclosed in sources | — |
| GPT-6.1 Sol (standard) | not scored in sources | $2.00 | $0.10 | $10.00 | Yes |
| Gemini 4 Argon | not scored in sources | $2.00 | — | $10.00 | Yes |
What Doubling Looks Like on One Real Task
Token prices feel abstract until you attach them to a job. Take a single agentic coding pass that reads 100,000 tokens of repository context and writes 20,000 tokens of diffs, explanation, and retries. On Muse Wonder that is $0.40 of input plus $0.40 of output — $0.80. On GPT-6.1 Sol at standard rates it is $0.20 plus $0.20 — $0.40. Run that loop fifty times in a working day and the difference is $40 versus $20. Nothing about that is catastrophic for one developer. It is a serious number for a team running agents in CI, and it is the reason the pricing, not the index score, is what the thread kept circling back to. Hacker News user hirako2000 compressed the whole objection into one line about the output rate: "$20 per M token make a pelican hard digest."
The Cache Discount Doubles Too
The detail that gets lost is that the 2× multiplier holds all the way down. Muse Wonder's cached input is $0.20 per million; Sol's is $0.10. Cached reads are where long-running agents actually spend their money — the same system prompt and the same repository context, re-read on every turn — so a workload optimized for caching does not escape the premium. It inherits it. There is one wrinkle worth knowing in the other direction: OpenAI applies long-context surcharges above 272,000 input tokens, pushing Sol to $4 input and $15 output per million for the whole request. At genuinely huge context sizes, Sol's input price converges on Meta's and Meta's output price is still a third higher.
The Model That Scores Isn't the Model You Can Rent
This is the structural oddity of the launch, and it is easy to miss when the two variants share a name.
The Index Numbers
On the Artificial Analysis Intelligence Index, Muse Wonder lands at 52 and Muse Wonder Full at 56. Claude Opus 5.5 sits at 58, Claude Sonnet 5.5 at 56, and GPT-6 Astra at 53. Read that ordering carefully. Muse Wonder Full ties Sonnet 5.5. Plain Muse Wonder — the one with a published token price — comes in below GPT-6 Astra while charging double Sol's rate. Meta also reports Muse Wonder Full at 79.0 on the DeepSWE 1.1 coding benchmark, and attributes Full's better generalization to unfamiliar tasks to its larger, stronger base model compared with Muse Spark 1.3.
A Walled Garden With a Price Tag
Muse Wonder Full is not available through the API at all. The only route to it is a Muse Code subscription: $50 a month for Power Usage, or $200 a month for Maximum Usage, which Meta describes as roughly 80× the allowance of the base Everyday Usage plan. One analysis of the launch put it bluntly — Wonder Full is positioned as the flagship and Meta is keeping it inside a walled garden of paid plans for now. The practical consequence is that the 56 on the index has no per-token price to compare against anything. You cannot benchmark Wonder Full's cost per task against Sol's, because Meta has not sold it that way.
| Plan | Price | Gets you |
|---|---|---|
| Meta Model API | $4 / $20 per M tokens | Muse Wonder only |
| Muse Code — Power Usage | $50/month | Muse Wonder Full access |
| Muse Code — Maximum Usage | $200/month | ~80× the base Everyday Usage allowance |
| Muse app — Free | $0 | Base access |
| Muse app — Standard | $20/month | Advanced model access |
| Muse app — Premium | $100/month | Premium agent capabilities |
What the First Testers Actually Found
The gap between the index numbers and the hands-on reports is the most useful thing in the Hacker News thread, because the testers were specific about which tasks broke.
"Closer to Sonnet or Haiku"
User jjcm ran image-to-HTML conversion — hand a model a screenshot, get markup back — and reported that Muse Wonder was much worse than Opus 5.5, adding: "IMO it's closer to Sonnet or Haiku." The diagnosis was narrower than a blanket dismissal: the vision model isn't mature yet, even where the text benchmarks look strong. That distinction matters for anyone picking a model for a specific job. A 52 on a general intelligence index tells you nothing about whether the thing can read a design mock.
Active Focus, Put Under Load
The sharpest test came from ekojs, who used Muse Wonder for real work — UI components, backend review, Rust server changes — and landed on a formulation worth quoting: "intelligence is there I think, but 'taste' and 'judgment' is not there yet." The concrete failure is the one that should give Meta pause, because it sits directly on top of the Active Focus pitch: the model became fixated on review comments and ground through seven iterations without reconsidering its approach. Active Focus promises persistence plus trying alternative approaches. Seven passes at the same idea is the first half delivered without the second.
The Pelican Problem
Simon Willison ran his familiar pelican-on-a-bicycle SVG test across Muse Spark 1.2, Spark 1.3, and Wonder, and the generations show incremental progression — Spark 1.3 adds animation. The thread's reaction to its own ritual was more interesting than the drawings. User wxce argued the test has outlived its usefulness, calling it a classic case of Goodhart's law, where a measure stops being informative once everyone optimizes for it. User bogzz asked whether pelicans would face left if prompted in a right-to-left language. Both are the sound of a community noticing that its improvised benchmarks have become marketing surface. Meanwhile user 217 offered the generous read on Wonder's position: "losing only to opus 5.5 is really really not bad."
What's Still Unknown
Three things the launch material does not settle, and they are the three that decide whether the premium is defensible:
- Whether Active Memory survives contact with months. Kopparapu's framing is explicitly about days, weeks, and months. Nobody has had the model for more than a few days. The feature cannot be evaluated yet by anyone, including Meta.
- Whether the alignment claims hold up externally. Meta calls Muse Wonder its most aligned model to date, with safeguards against CBRN and cyber misuse while preserving defensive cybersecurity capability. That is a self-assessment until third parties publish on it.
- Whether the open-weights promise lands. User kennywinker noted that Meta committed to an open-weight Muse Spark release on September 2 and developers are still waiting. Meta's reputation for open weights is doing real work in how this launch is received, and that IOU is now over a month old.
The honest summary is that Meta built something frontier-adjacent and priced it like something frontier-leading. The index numbers say Muse Wonder Full is competitive; the token prices say the variant you can actually rent costs double a rival that scores higher than it does. If the next few weeks of real usage show that Active Memory genuinely removes the re-briefing tax — that you stop re-explaining your codebase every morning — then $4 and $20 will look like a bargain, because the thing it saves is not tokens but human time. If it does not, Meta has a second-place model behind a first-place paywall, and the thread that greeted it will have been right on day one. The useful test is not the next benchmark release. It is whether anyone who tried Muse Wonder on October 12 is still paying for it in December.
-EditorZ
Photo by Vitaly Gariev on Unsplash
Post a Comment