Mercury 3 vs Muse Wonder: Thirty-Two Seconds to First Token

There is a measured number in this comparison that ought to end the argument before it starts. Artificial Analysis records Muse Wonder's time to first token at 32.08 seconds. Mercury 3 claims 1,068 output tokens per second with no independent measurement at all. One model takes half a minute to begin answering; the other says it writes a thousand words in the time it takes to read this sentence. On latency they are not in the same category of product, and no amount of benchmark superiority closes a thirty-fold gap in when the first character appears. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

But that is the easy version of the story, and it is misleading in one important respect. Muse Wonder is not trying to be fast. Meta released Muse Wonder on October 12, 2026 as a frontier reasoning model priced at $4.00 per million input tokens and $20.00 per million output tokens, and at that price the thing being bought is depth — a model that thinks for thirty seconds and gets the hard problem right. Mercury 3, released by Inception on October 19, 2026, is priced twenty to fifty times lower because it is selling the opposite trade. The question a builder actually faces is not which is better. It is which trade your product makes, and whether you have been honest with yourself about which one that is.

The two models, in one paragraph each

Mercury 3 is a diffusion language model — Inception's term is dLLM — that generates tokens in parallel rather than one at a time. It carries a 1M-token context, a 131,072-token maximum output, text-only input, tool calling and structured outputs on an OpenAI-compatible endpoint, and four reasoning-effort settings (instant, low, medium, high) where the default is low. List price is $0.20 per million input and $0.70 per million output, currently halved to $0.10 and $0.35 "at launch" with no published end date. Inception's quality claim is comparative and unscored: it says the model is on par with or better than "frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5", and reports one task result from its own agent harness — 50 per cent completion on 50 live Mind2Web tasks.

Muse Wonder is Meta's frontier reasoning model, shipped as two checkpoints: muse-wonder-1.0, callable on the Meta Model API and inside Muse Code, and muse-wonder-full-1.0, a higher-scoring configuration Meta says is available only to Muse Code subscribers. Both carry a 1M-token context and the same $4.00/$20.00 rate card, with a 95 per cent cache discount. Reasoning is mandatory — the effort ladder runs from minimal to max, defaulting to medium, and reasoning tokens bill as output. Inputs are text, image, speech and video; output is text.

What each one actually costs at the request level

Per-token prices are the least useful number in a comparison like this, because the two models consume wildly different token volumes to do the same job. A reasoning model that defaults to medium effort and spent 130 million output tokens completing an evaluation run is not priced the same as a model optimised for short answers, even before the rate card is applied. So price the same job both ways.

Take a document-review assistant: 20,000 tokens of contract text in, a 500-token structured summary and risk list out, 2,000 reviews a day.

• Mercury 3 at list — 40 million input tokens at $0.20 is $8.00; 1 million output tokens at $0.70 is $0.70. About $8.70 a day.

• Mercury 3 at the launch rate — half of that, about $4.35 a day.

• Muse Wonder — 40 million input at $4.00 is $160; 1 million output at $20.00 is $20. Plus reasoning tokens, which bill as output and which Meta does not let you switch off. At a conservative 1,000 reasoning tokens per call the output line becomes 3 million tokens, or $60, for a daily total of $220.

That is a twenty-five-fold gap on identical traffic, and it is the mildly favourable version for Muse Wonder, because reasoning models routinely spend far more than a thousand thinking tokens on a task that requires care. The effort setting is the only lever, and it is a lever on both cost and quality at once — you cannot turn the thinking down without turning the answer down.

Here is the part that makes the comparison more than arithmetic. Mercury 3 does not publish an output-token count for any evaluation, so nobody knows whether it is terse or verbose against a baseline task. If it turns out to be a talkative model, the daily cost moves by multiples, not percentages. The twenty-five-fold gap could be fifteen-fold or forty-fold in production and there is no published evidence either way.

The throughput claim, and the number that undercuts it

Inception's headline is 1,068 tokens per second on NVIDIA GPUs, described as 4 to 8 times faster than the models it compares itself to. The company does not name the hardware, the batch size, the effort setting or whether the figure is single-stream or aggregate, and those four choices move the number by multiples. There is no independent Mercury 3 measurement as of October 20, 2026 — Artificial Analysis still lists Mercury 2.5 and Mercury 2 and has no Mercury 3 page.

What independent measurement does exist for the previous generation is instructive. Artificial Analysis puts Mercury 2.5 at roughly 706 tokens per second under its own harness. That is a real, reproducible third-party figure, and it is meaningfully below the vendor numbers published at that launch. That gap is the normal vendor-to-independent discount in this category rather than evidence of bad faith — but it is the discount to apply mentally to 1,068 until somebody reproduces it.

Muse Wonder, meanwhile, has a measured 161 tokens per second of output and a measured 32.08-second time to first token. Both figures are Meta's model operating as designed: the first token arrives after a long thinking pass, and the rest of the answer streams at a middling rate. For a batch job that is irrelevant. For anything with a human waiting, it is the whole experience.

The measured quality gap runs the other way. Muse Wonder scores 52 on the Artificial Analysis Intelligence Index, ranked ninth of 217 models, at $4.52 of compute per index task. Mercury 3 has no index score. Muse Wonder Full scores 56 and ranks fourth of 696, with no published price because Meta does not sell it per token. So the honest framing is a frontier model with a measured ceiling and a measured latency problem, against a budget model with a claimed latency advantage and no measured ceiling at all.

Reasoning is a product decision, not a setting

The deepest difference between these models is not price or speed. It is whether the model is allowed to think briefly.

Mercury 3 defaults to low effort when the reasoning parameter is omitted, and the four settings run from instant to high. That default is the detail most likely to be missed and most likely to distort any comparison: a model answering at low effort will look faster and score lower than the same model at high, and Inception's own comparative claims do not state which setting produced them. If you benchmark Mercury 3 yourself, pin the effort level explicitly or your numbers will not be comparable to anyone else's — including Inception's.

Muse Wonder is the opposite by construction. Reasoning is mandatory, the default is medium, and there is no switch to turn it off. That is a deliberate design choice for a model aimed at professional and agentic work: Meta's framing is that the thinking is the product. It also means the model has no cheap path. Every call pays the reasoning tax, and at $20 per million output tokens a long agentic loop that fills the 943,718-token output budget is a serious invoice.

For a builder the practical consequence is that these two models cannot be swapped behind a single prompt template and compared fairly. Doing the same task well on Mercury 3 means choosing an effort level; doing it on Muse Wonder means accepting the thinking time. The comparison only becomes meaningful once both are pinned to the same effective behaviour, and for Mercury 3 that behaviour is undocumented.

On Meta's own scoreboard, and why it is not the last word

Meta published a benchmark table at launch covering both checkpoints alongside Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 at maximum reasoning. Every number is vendor-reported, produced in Meta's harness, and none has been independently reproduced. Two rows are worth carrying into this comparison anyway, because they show the model doing what it is sold to do.

• JobBench, long-horizon professional tasks — Muse Wonder Full 75.3, Muse Wonder 30.8, Claude Fable 5.1 27.6, GPT-6 Astra 34.2, Claude Opus 5.5 30.6

• Vals Finance Agent V2 — Muse Wonder Full 59.1, Muse Wonder 10.0, Claude Fable 5.1 6.4, GPT-6 Astra 8.6, Claude Opus 5.5 4.2

• OSWorld V2.1, computer use — Muse Wonder Full 78.4, Muse Wonder 72.6, Claude Fable 5.1 73.1, GPT-6 Astra 66.1, Claude Opus 5.5 75.4

Where Meta and the independent evaluator disagree, the independent number is the one to plan against. Meta's table lists Muse Wonder Full at 59.0 on the Artificial Analysis Index; Artificial Analysis itself publishes 56. Three points on the same index is not a scandal, but it establishes which figure has a methodology behind it.

Mercury 3 has no equivalent table to argue with. Its only structured quality result is the Mind2Web agent test: 50 live navigation tasks in Inception's own production harness, 50 per cent completed, "significantly more than GPT-6.1-Sol", using the same token volume and the same number of agent steps and therefore a fraction of the per-token cost. Fifty tasks is a small sample, the harness is Inception's, and the comparison model is named without a score. It is a real signal that the model can drive a browser at a fraction of the cost of a frontier model doing the same thing. It is not a benchmark.

Which of these two you should actually be evaluating

The decision usually resolves on one question: does a human wait for this output, or does a queue consume it?

If a human waits, Muse Wonder is close to disqualified by its own measurements. Thirty-two seconds to first token is not a latency you can put in front of a customer, and nobody chooses it for a chat surface. Mercury 3 is built for exactly this case, and its price makes the experiment cheap. Even if the throughput lands at half the claim under a neutral harness, it is still several times faster than anything in the frontier tier.

If a queue consumes it — overnight document processing, agentic runs that deliver a report, back-office workflows where nobody is watching — then the 32-second first token stops mattering and the measured quality gap starts to. Muse Wonder at 52 on the index, with a full independent record, is a known quantity. Mercury 3 at an unknown index score, on a workload where a wrong answer is expensive and unmonitored, is a bet.

Three cases sit outside both:

• Video, images or audio in the prompt — Muse Wonder takes all three, Mercury 3 takes none. There is no trade to make.

• Anything requiring the highest-scoring checkpoint — Muse Wonder Full is not an API product. It is a $50-to-$200 monthly subscription to Muse Code, and Meta has not said it will open it.

• Anything where you need to plan against a stable price — Mercury 3's launch discount has no end date, and its predecessors are being retired for new users. Muse Wonder's $4/$20 has no announced change.

What OrcaRouter does and does not do here

Neither Muse Wonder nor Mercury 3 is on our catalogue. A check of our model API returns nothing for meta/muse-wonder and nothing for any Inception model, and we will not describe a route we do not have. What we carry from Meta's lab is the Muse Spark line — Muse Spark 1.2 and Muse Spark 1.1, both with a 1M-token context at $1.25 in and $4.25 out per million tokens.

That matters more for this specific matchup than it sounds. The reason to run Muse Wonder is depth, and the reason to run Mercury 3 is speed. Both of those can be tested against a model you can already call: Muse Spark 1.2 gives you a Meta reasoning family with a published rate card on our side, and the 200-plus models reachable through one OpenAI-compatible key give you the latency comparison without a second contract. If Muse Wonder's $4/$20 is going to be justified for your workload, it has to beat something — and the something should be a model you can measure this afternoon.

The mechanical advantage worth naming is price pass-through. Provider list prices reach your bill unchanged, so a vendor cut — like Inception's 50 per cent launch discount or a hypothetical Meta reduction — is live on our side the day it is announced, not at the next renewal. That is a small thing until the day it is not.

What would settle this

Two measurements would do most of the work. The first is an independent Mercury 3 throughput figure under a stated harness, effort level and batch size; the second is any independent Mercury 3 quality score at all. Until both exist, the fastest model in this comparison is also the one you know least about, and the most capable is the one you can only call with a subscription or at a price that is hard to justify without a use case that tolerates half a minute of silence.

What makes this an easy decision rather than a hard one is that the two models do not compete for the same slot. The only real risk is picking the wrong slot — deploying a thirty-second model into a chat window because its benchmark is better, or deploying an unmeasured model into an unmonitored pipeline because its price is better. Both of those are decisions about your product, not about the models, and neither one is settled by a number on a leaderboard.